Digital Signal and Image

Processing
(CSC701)

Semester VII - Computer Engineering

(Mumbai University)

ge
io led
ic ow

Strictly as per the New Revised Syllabus (Rev- 2016) of

Mumbai University w.e.f. academic year 2019-2020
n
bl kn
at
Pu ch
Te

Dhananjay K. Theckedath
M. Tech. - IIT (Mumbai) - Biomedical Engineering,
B.Sc. Tech. - Electronic Instrumentation
B.Sc. - Physics,
Assistant Professor,
Thadomal Shahani Engineering College,
Mumbai University, Maharashtra, India.

(Book Code : MO80A)

Digital Signal & Image Processing
Dhananjay K. Theckedath
(Semester VII - Computer Engineering) (MU))

Copyright © by Author. All rights reserved. No part of this publication may be reproduced, copied, or stored in a retrieval
system, distributed or transmitted in any form or by any means, including photocopy, recording, or other electronic or
mechanical methods, without the prior written permission of the publisher.

This book is sold subject to the condition that it shall not, by the way of trade or otherwise, be lent, resold, hired out, or
otherwise circulated without the publisher’s prior written consent in any form of binding or cover other than which it is
published and without a similar condition including this condition being imposed on the subsequent purchaser and without

ge
limiting the rights under copyright reserved above.
io led
First Printed in India : August 2008 (For UPTU)
First Edition : July 2019(TechKnowledge Publications)
ic ow

This edition is for sale in India, Bangladesh, Bhutan, Maldives, Nepal, Pakistan, Sri Lanka and designated countries in
n
South-East Asia. Sale and purchase of this book outside of these countries is unauthorized by the publisher.
bl kn
at
Pu ch

P
Te

ISBN : 978-93-89424-30-0

Published by
TechKnowledge Publications

Head Office : B/5, First floor, Maniratna Complex,

Taware Colony, Aranyeshwar Corner,
Pune - 411 009. Maharashtra State, India
Ph : 91-20-24221234, 91-20-24225678.
[CSC701] (FID : MO80) (Book Code : MO80A)

(Book Code : MO80A)

 We dedicate this Publication soulfully and wholeheartedly,
in loving memory of our beloved founder director,
Late. Shri. Pradeepsheth Lalchandji Lunawat, who will always be an
inspiration, a positive force and strong support behind us.

ge
io led
ic ow
n
bl kn
at
Pu ch
Te

Lt. Shri. Pradeepji L. Lunawat

Soulful Tribute and Gratitude for all Your

Sacrifices, Hardwork and 40 years of Strong Vision…….

(Book Code : MO80A)

 Te
Pu ch
bl kn
ic ow
at
io led
n ge

Preface

Dear Students,

I am extremely happy to present this book “Digital Signal & Image Processing” for you. I have

divided the subject into small chapters so that the topics can be arranged and understood properly. The

topics within the chapters have been arranged in a proper sequence to ensure smooth flow and

ge
understanding of the subject.
io led
A number of solved examples have been included. So, I am sure that this book will cater all

your needs for this subject.

ic ow
n
I present this book in the loving memory of Late. Shri. Pradeepji Lunawat, our source of
bl kn

inspiration and a strong foundation of “TechKnowledge Publications”. He will always be remembered

at
Pu ch

in our hearts and motivate us to achieve our new milestone.

Te

I am thankful to Prof. Arunoday Kumar, Shri. Shital Bhandari & Shri. Chandroday Kumar for the

encouragement and support that they have extended. I am also thankful to the staff members of

TechKnowledge Publications and others for their efforts to make this book as good as it is. I have made

every possible efforts to eliminate all the errors in this book. However if you find any, please let me

know, because that will help me to improve the book quality further.

I am also thankful to my family members and friends for their patience and encouragement.

- Author



(Book Code : MO80A)

 Syllabus

Digital Signal & Image Processing

Course Code Course Name Credits

CSC701 Digital Signal & Image Processing 4

Course Objectives :

1. To understand the fundamental concepts of digital signal processing and Image processing.
2. To explore DFT for 1-D and 2-D signal and FFT for 1-Dsignal
3. To apply processing techniques on 1-D and Image signals.

ge
4. To apply digital image processing techniques for edge detection.
Course outcomes :
io led
On successful completion of the course learner will be able to :
1. Apply the concept of DT Signal and DT Systems.
ic ow

2. Classify and analyze discrete time signals and systems

n
3. Implement Digital Signal Transform techniques DFT and FFT.
bl kn

4. Use the enhancement techniques for digital Image Processing

at

5. Differentiate between the advantages and disadvantages of different edge detection techniques
Pu ch

6. Develop small projects of 1-D and 2-D Digital Signal Processing.

Prerequisite: Applied Mathematics
Te

Module Unit
Topic details Hrs.
No. No.

Discrete-Time Signal and Discrete-Time System

Introduction to Digital Signal Processing, Sampling and Reconstruction, Standard DT Signals,

1.1 Concept of Digital Frequency, Representation of DT signal using Standard DT Signals, Signal
Manipulations(shifting, reversal, scaling, addition, multiplication).

1.2 Classification of Discrete-Time Signals, Classification of Discrete- Systems

1.0 Linear Convolution formulation for 1-D and 2-D signal (without mathematical proof), 14
Circular Convolution (without mathematical proof), Linear convolution using Circular
Convolution. Auto and Cross Correlation formula evaluation, LTI system, Concept of Impulse

1.3 Response and Step Response, Output of DT system using Time Domain Linear Convolution.

(Refer Chapters 1, 2 and 3)

(Book Code : MO80A)

 Discrete Fourier Transform

2.1 Introduction to DTFT, DFT, Relation between DFT and DTFT, IDFT

Properties of DFT without mathematical proof (Scaling and Linearity, Periodicity, Time Shift
2.2 and Frequency Shift, Time Reversal, Convolution Property and Parsevals‘ Energy Theorem).
2.0 08
DFT computation using DFT properties.

Transfer function of DT System in frequency domain using DFT. Linear and Circular
2.3 Convolution using DFT, Convolution of long sequences, Introduction to 2-D DFT
(Refer Chapter 3)

Fast Fourier Transform

06
3.1 Need of FFT, Radix-2 DIT-FFT algorithm,
3.0

ge
3.2 DIT-FFT Flow graph for N=4 and 8, Inverse FFT algorithm.

3.3 Spectral Analysis using FFT

io led (Refer Chapters 3 and 4)

Digital Image Fundamentals

4.1 Introduction to Digital Image, Digital Image Processing System, Sampling and Quantization
4.0 08
ic ow

4.2 Representation of Digital Image, Connectivity

n
4.3 Image File Formats: BMP, TIFF and JPEG. (Refer Chapters 5 and 6)
bl kn

Image Enhancement in Spatial domain

at

5.1 Gray Level Transformations, Zero Memory Point Operations,

Pu ch

5.0 5.2 Histogram Processing, Histogram equalization. 10

Te

Neighborhood Processing, Spatial Filtering, Smoothing and Sharpening Filters, Median Filter.
5.3
(Refer Chapters 7 and 8)

Image Segmentation

6.1 Segmentation based on Discontinuities (point, Line, Edge),

06
6.0 Image Edge detection using Robert, Sobel, Previtt masks, Image Edge detection using
6.2
Laplacian Mask. (Refer Chapter 9)

Total 52



(Book Code : MO80A)

 DSIP (MU-Sem.7-COMP) 1 Table of Contents

2.1 Introduction ................................................................. 2-1

Chapter 1 : Introduction to Digital Signal Processing
1-1 to 1-17 2.1.1 Representation of Discrete Time Signals..................... 2-1

2.1.2 Elementary Discrete Time Signals............................... 2-2

Syllabus :
2.2 Classification of Discrete Time Signals........................ 2-4
Introduction to Digital Signal Processing, Sampling and 2.2.1 Even and Odd Signals ................................................. 2-5
Reconstruction, Standard DT Signals, Concept of Digital
2.2.2 Periodic and Aperiodic Signals .................................... 2-7
Frequency, Representation of DT signal using Standard
2.2.3 Energy and Power Signals .......................................... 2-8
DT Signals, Signal Manipulations(shifting, reversal,
scaling, addition, multiplication). 2.3 Operation on Signals ................................................. 2-10

1.1 Introduction ................................................................. 1-1 2.3.1 Time Shifting ............................................................. 2-10

1.1.1 Digital Signal Processing System ................................ 1-2 2.3.2 Time Reversal ........................................................... 2-12

1.1.2 Advantages of Digital signal Processing 2.3.3 Time Scaling ............................................................. 2-12
over Analog Signal Processing .................................... 1-3

ge
2.3.4 Scalar Multiplication .................................................. 2-13
1.1.3 Difference between DSP and ASP .............................. 1-3 2.3.5 Signal Addition and Multiplication .............................. 2-14
1.1.4 Limitations of Digital signal Processing ....................... 1-4 2.3.6 Solved Examples on Signals ..................................... 2-14
io led
1.2 Signals ........................................................................ 1-4 2.4 Discrete Time Systems ............................................. 2-17
1.2.1 Continuous and Discrete Time Signals ........................ 1-4 2.4.1 Representation of a Signal in Terms of
ic ow

1.2.2 Continuous Valued and Discrete Weighted Sum of Shifted Discrete Impulse ............... 2-20
Valued Signals ............................................................ 1-6
2.5 Impulse Response and Convolution .......................... 2-21
n
1.2.3 Deterministic and Random Signals .............................. 1-6
bl kn

2.5.1 Computation of Linear Convolution............................ 2-22

1.3 Concept of Frequency in the Continuous
2.5.2 Linear Convolution using Graphical Method .............. 2-22
and Discrete Time Signals........................................... 1-6
at
Pu ch

2.5.2(A) Solved Examples on Graphical Method ..................... 2-22

1.3.1 Concept of Frequency in Continuous
Time Signals ............................................................... 1-6 2.5.3 Linear Convolution using Tabular Method.................. 2-26
Te

1.3.2 Concept of Frequency in Discrete 2.5.3(A) Solved Examples on Tabular Method ........................ 2-27

Time Signals ............................................................... 1-7 2.6 Correlation................................................................. 2-27

1.3.3 Converting a Analog Signal to a 2.6.1 Solved Examples on Correlation................................ 2-28

Discrete Time Signal ................................................... 1-8
Chapter 3 : Discrete Fourier Transform (DFT)
1.4 Aliasing and Nyquist Rate ........................................... 1-9
3-1 to 3-68
1.4.1 Anti Aliasing Filter...................................................... 1-13

1.5 Basic Elements of a DSP System ............................. 1-13 Syllabus :

1.5.1 Solved Examples ....................................................... 1-14 Introduction to DTFT, DFT, Relation between DFT and
1.6 Applications of DSP................................................... 1-17 DTFT, IDFT, Properties of DFT without mathematical
Chapter 2 : Discrete Time Signal and System proof (Scaling and Linearity, Periodicity, Time Shift and
Frequency Shift, Time, Reversal, Convolution Property
2-1 to 2-29
and Parsevals‘ Energy Theorem). DFT computation using
Syllabus : DFT properties. Transfer function of DT System in
Classification of Discrete-Time Signals, Classification of frequency domain using DFT. Linear and Circular
Discrete- Systems Linear Convolution formulation for 1-D Convolution using DFT, Convolution of long sequences,
and 2-D signal (without mathematical proof), Auto and Introduction to 2-D DFT Circular Convolution (without
Cross Correlation formula evaluation, LTI system,
mathematical proof), Linear convolution using Circular
Concept of Impulse Response and Step Response,
Output of DT system using Time Domain Linear Convolution. Spectral Analysis using FFT
Convolution 3.1 The Fourier Transform................................................. 3-1
 DSIP (MU-Sem.7-COMP) 2 Table of Contents

3.2 Discrete Time Fourier Transform ................................. 3-2 4.1 Introduction ................................................................. 4-1

3.2.1 Solved Examples on DTFT of Standard Signals .......... 3-2 4.2 Computational Complexity of DFT ............................... 4-1

3.3 Relationship between DTFT and DFT.......................... 3-4 4.3 Decimation In Time FFT (DIT-FFT) ............................. 4-2

3.4 Discrete Fourier Transform (DFT)................................ 3-4 4.4 Bit Reversal Format..................................................... 4-6

3.4.1 Solved Examples on DFT of Simple Signals................ 3-5 4.4.1 Solved Examples on Bit Reversal Format ................... 4-6

3.5 The Fourier Spectrum................................................ 3-13 4.5 Decimation in Frequency FFT (DIT-FFT) ................. 4-17

3.6 Inverse Discrete Fourier Transform (IDFT) ................ 3-14 4.5.1 Solved Examples on DIT-FFT ................................... 4-20

3.7 Computing DFT by Matrix Method ............................. 3-15 4.6 Computational Complexity Comparison
between DFT and FFT .............................................. 4-30
3.7.1 Solved Examples of DFT ........................................... 3-16
4.6.1 Solved Examples on Computation of Inverse
3.8 Discrete Fourier Transform (DFT) Properties ............ 3-29
DFT using FFT Algorithms ........................................ 4-31
3.8.1 Linearity..................................................................... 3-29
4.6.2 Filtering using FFT Algorithms ................................... 4-37

ge
3.8.2 Periodicity.................................................................. 3-30
3.8.3 Circular Time Shift ..................................................... 3-31 Chapter 5 : Introduction to Image Processing

3.8.4 Circular Frequency Shift ............................................ 3-38

5-1 to 5-14
io led
3.8.5 Time Reversal ........................................................... 3-39
Syllabus :
3.8.6 Symmetry Property.................................................... 3-40
Introduction to Digital Image, Digital Image Processing
ic ow

3.8.7 Complex Conjugate Property ..................................... 3-44

System, Image File Formats : BMP, TIFF and JPEG.
3.8.8 Parseval's Theorem ................................................... 3-45
n
5.1 Introduction ................................................................. 5-1
3.8.9 Multiplication of Two Sequences ............................... 3-46
bl kn

5.2 What Do We Mean by Image Processing ? ................. 5-1

3.8.10 Circular Convolution Property .................................... 3-47
5.3 Images ........................................................................ 5-2
at

3.9 Circular Convolution .................................................. 3-47

Pu ch

5.4 The Electromagnetic Spectrum ................................... 5-3

3.9.1 Concentric Circle Method .......................................... 3-47
5.5 Units of Intensity.......................................................... 5-3
3.9.1(A) Solved Example on Concentric Circle Method ........... 3-47
Te

5.6 Basic Elements/Steps of an Image

3.9.2 Matrix Method ............................................................ 3-51
Processing System ..................................................... 5-4
3.9.2(A) Solved Example on Matrix Method ............................ 3-51
5.7 Image Types ............................................................... 5-9
3.9.3 Getting Linear Convolution
5.8 Image File Formats .................................................. 5-12
from Circular Convolution .......................................... 3-54
5.9 Image Processing, Computer Graphics
3.9.4 Solved Examples ....................................................... 3-54
3.10 Filtering of Long Duration Signals .............................. 3-60 and Computer Vision ................................................. 5-13

3.10.1 Overlap - Save Method .............................................. 3-60 Chapter 6 : Sampling and Quantization 6-1 to 6-10
3.10.2 Overlap - Add Method ............................................... 3-62
3.10.3 Solved Examples on Method ..................................... 3-64
Syllabus :
Sampling and Quantization. Representation of Digital
3.11 Summary of DFT properties ...................................... 3-66
Image
3.12 2 D- Discrete Fourier Transform (2-DFT) .................. 3-66
6.1 Introduction ................................................................. 6-1
3.12.1 The Separability Property ......................................... 3-66
6.2 Sampling ..................................................................... 6-1
3.12.2 Solved Examples on 2- DFT ...................................... 3-67
6.3 Quantization ................................................................ 6-5
Chapter 4 : Fast Fourier Transform 4-1 to 4-41
6.4 Isopreference Curves .................................................. 6-8

Syllabus : 6.5 Non-uniform Sampling ................................................. 6-9

Need of FFT, Radix-2 DIT-FFT algorithm, DIT-FFT Flow 6.6 Physical Resolution ..................................................... 6-9

graph for N=4 and 8, Inverse FFT algorithm. 6.6.1 Solved Examples on Physical Resolution .................... 6-9
 DSIP (MU-Sem.7-COMP) 3 Table of Contents

8.3.1 Solved Examples on Histogram Equalization .............. 8-7

Chapter 7 : Image Enhancement in Spatial
Domain 7-1 to 7-35 8.4 Histogram Specification............................................. 8-12

8.4.1 Solved Example Histogram ....................................... 8-12

Syllabus :
8.5 Solved Examples on Histogram Modelling................ 8-13
Gray Level Transformations, Zero Memory Point
Chapter 9 : Segmentation 9-1 to 9-23
Operations, Neighborhood Processing, Spatial Filtering,
Smoothing and Sharpening Filters, Median Filter. Syllabus :
7.1 Introduction ................................................................. 7-1 Segmentation based on Discontinuities (point, Line,
7.2 Spatial Domain Methods ............................................. 7-1 Edge), Image Edge detection using Robert, Sobel, Previtt
7.3 Point Processing / Zero Memory Operation ................. 7-2 masks, Image Edge detection using Laplacian Mask.
Connectivity
7.4 Neighbourhood Processing ......................................... 7-9
9.1 Introduction ................................................................. 9-1
7.4.1 Low Pass Filtering (Smoothing) ................................. 7-10

ge
9.2 Image Segmentation based on Discontinuities ............ 9-2
7.4.2 Noise ......................................................................... 7-10
9.2.1 Point Detection ............................................................ 9-2
7.4.3 Low Pass Averaging Filter ......................................... 7-11
io led 9.2.2 Line Detection ............................................................. 9-2
7.4.4 Low Pass Median Filtering......................................... 7-14
9.2.3 Edge Detection ............................................................ 9-3
7.5 Highpass Filtering...................................................... 7-16
9.3 Detection of Edges ...................................................... 9-5
ic ow

7.6 High-Boost Filtering ................................................... 7-18

9.3.1 Computing the Gradient .............................................. 9-5
7.7 Zooming .................................................................... 7-19
n
9.4 Finding Gradients using Masks ................................... 9-6
7.7.1 Replication ................................................................ 7-20
bl kn

9.4.1 Roberts Mask .............................................................. 9-7

7.7.2 Linear Interpolation .................................................... 7-21
9.4.2 Prewitts and Sobel Operators ...................................... 9-8
7.8 Solved Examples ....................................................... 7-22
at
Pu ch

9.4.3 Compass Operators .................................................. 9-11

7.9 Comparison of Contrast Stretching
9.5 Image Segmentation using the Second
and Thresholding ....................................................... 7-34
Te

Derivative - the Laplacian .......................................... 9-13

Chapter 8 : Histogram Modelling 8-1 to 8-20 9.5.1 Laplacian of Gaussian Operator ................................ 9-14

Syllabus : 9.5.2 Canny Edge Detector ................................................ 9-16

9.6 Connectivity............................................................... 9-16

Histogram Processing, Histogram equalization.
9.6.1 Solved Examples on Connectivity.............................. 9-18
8.1 Introduction ................................................................. 8-1
9.7 Distance Transform ................................................... 9-18
8.1.1 Mean and Standard Deviation of Histogram ............... 8-3
9.7.1 Solved Examples on Distance Transform ................. 9-19
8.2 Linear Stretching ......................................................... 8-3
9.8 Additional Solved Examples ..................................... 9-19
8.2.1 Solved Examples on Linear Stretching ........................ 8-3

8.3 Histogram Equalization................................................ 8-6


 Introduction to Digital
1 Signal Processing

Syllabus :
Introduction to Digital Signal Processing, Sampling and Reconstruction, Standard DT Signals, Concept of Digital
Frequency, Representation of DT signal using Standard DT Signals, Signal Manipulations(shifting, reversal, scaling,
addition, multiplication).

1.1 Introduction

ge
 A signal is defined as any physical quantity that varies
with time, space or any other independent variable.
io led
Anything that carries some information can be called a
signal.
 Examples of signals that we encounter frequently in
ic ow

our daily life are speech, music, image, video etc.

Some other examples of signals are
Electrocardiograms (ECG) which provides us
n
information about the health of a person’s heart,
bl kn

Seismic signals which are used to measure the

intensity of an earthquake, AC power supply signals
at
Pu ch

that we have at our homes and offices etc.

 Mathematically, we define a signal as a function of one
Te

or more independent variables. Consider a voltage Fig. 1.1.1

signal and a pressure signal,
 Both these signals in the given example depend on an
v(t) = 15t independent variable t i.e. both voltage and pressure
–2t change with respect to time. Such signals are known as
p(t) = 10e …(1.1.1)
1-Dimensional signals. In the Fig. 1.1.1, the x-axis is the
MATLAB program for displaying 1-dimensional signals
time axis (independent variable) while the
clc y-axis is the voltage (dependent variable).
clear all
 Another example of a signal is an image signal that we
for t=1:1:10 see on the computer monitor. A pixel on the computer
v(t)=5*t; could be represented as,
p(t)=10*exp(–0.2*t);
f (x, y) = 10x + 12y …(1.1.2)
end
Here, f depends on two independent variables x and y
subplot (2, 1, 1)
and hence are known as 2-Dimensional signals. This
plot (v)
subject deals with only 1-Dimensional signals.
xlabel (‘Time’)
 The Equation (1.1.2) represent a class of signals that
ylabel (‘voltage v(t)’)
are precisely defined. However there are many such
grid on
signals that cannot be directly represented by a
subplot (2, 1, 2)
mathematical formula. For example a speech signal
plot (p)
shown in Fig. 1.1.1 cannot be represented by a
xlabel (‘Time’) mathematical equation of the kind shown in
ylabel (‘Pressure p(t) ‘) Equations (1.1.1) and (1.1.2).
grid on
 DSIP (MU-Sem. 7-COMP) 1-2 Introduction to Digital Signal Processing

ge
io led
ic ow

Fig. 1.1.2 : Speech signal

 Associated with all the above mentioned signals are c) Signal processing : Production of a desired output
n
the means by which these signals are generated. signal from an input signal is known as signal
bl kn

For example speech signals are generated by the vocal processing.

chords; the ECG signal is generated by the polarization This chapter provides an overview of signals, systems
at
Pu ch

and depolarization of the ventricles of the heat. and signal processing methods.
 Hence signals can be said to be generated by a system.
1.1.1 Digital Signal Processing System
Te

 A system is a device that produces a signal and also

performs an operation on the signal to give the  Most of the signals that occur in nature are continuous
desired output. (analog) i.e., they are a function of a continuous
variable such as time (e.g. speech and ECG)
 Suppose we want to amplify a signal. In this case the
amplifier is a system. Similarly, if we want to filter out  In many cases, this continuous input signal can be
some noise, we pass it through a filter. In this case, the directly processed in the continuous form itself.
filter is also a system. Hence a system could be an Let us take an example,
interconnection of components that performs an  Suppose we want to amplify our voice. The set up
operation on an input signal and produces an output would be as shown in Fig. 1.1.3.
signal.
 This production of a desired output signal from an
input signal is known as signal processing.
Let us define the three important terms discussed. Fig. 1.1.3
a) Signal : A signal is defined as any physical quantity  The amplifier which is made up resistors, capacitors,
that varies with time, space or any other transistors etc., takes the input from the microphone
independent variable. Anything that carries some which is a continuous signal and amplifies it to gives us
information can be called a signal. a continuous output signal.
b) System : A system is defined as a device or a  Such an operation is called Analog Signal Processing. A
combination of devices which produces signals general analog processing system is shown in
and also performs operation on signals to get the Fig. 1.1.4. In this, the signal is directly processed in its
desired output. continuous (analog) form.
 DSIP (MU-Sem. 7-COMP) 1-3 Introduction to Digital Signal Processing

Fig. 1.1.4

 Digital signal processing, which this chapter is all about, provides an alternate method of processing the same signal.
In this method we convert the continuous signal to a discrete time signal, also known as a digital signal, using an
Analog to Digital Converter (ADC).
 The output of the ADC is fed to a digital processor (could also be a computer) which processes the discrete time signal.
The output of the digital processor is again in digital form. This needs to be converted back to the continuous signal
for any output device to understand it. This is done by using a Digital to Analog Converter (DAC). Let us take the same
example of trying to amplify a voice signal. If we do it using a digital signal processor, we would have the following
block diagram as shown in Fig. 1.1.5.

ge
io led
Fig. 1.1.5

In general we have,
ic ow
n
bl kn

Fig. 1.1.6

Here, x(t)  Continuous time input signal. x(n)  Discrete time input signal.
at
Pu ch

y(n)  Discrete time output signal. y(t)  Continuous time output signal.
5. Complex mathematical algorithms can be easily
1.1.2 Advantages of Digital Signal
Te

implemented using DSP systems while it is extremely

Processing over Analog Signal difficult to build these algorithms using circuits.

Processing 6. Because of the use of software, DSP systems can be

easily upgraded compared to Analog Signal Processing
Though a DSP system has more blocks than an ASP systems.
system, there are many advantages of working in the digital 7. In most cases, DSP systems are cheaper compared to
domain. We will list down the advantages here. ASP systems.
1. Digital Signal Processing (DSP) systems are flexible i.e., Based on the above discussion we shall now form a
they can be reconfigured simply by changing the table for a quick reference.
program. Hence these systems are extremely versatile.
2. DSP systems can be easily replicated (duplicated) as 1.1.3 Difference between DSP and ASP
they do not depend on component tolerances.
3. Accuracy also plays a pivotal role in determining the Table 1.1.1
form of the signal processor. To design an analog
Sr. DSP ASP
system, we need components like resistors, capacitors
and inductors. The tolerances of these components No.
reduce accuracy of the analog system. On the other
1. More flexible Less flexible
hand, DSP systems provide much better control of the
accuracy. 2. Better accuracy Less accuracy
4. Digital signal can be stored without deterioration. This
is not true with analog signals. 3. Digital signals can be It is difficult to store
easily stored. analog signals.
 DSIP (MU-Sem. 7-COMP) 1-4 Introduction to Digital Signal Processing

Sr. DSP ASP 1.2.1 Continuous and Discrete Time Signals

No.  A signal that exists for all time in a given interval is said
to be a continuous signal (also called analog signal).
4. Digital systems can be Tedious to replicate
Most of the naturally occurring signals are continuous
easily replicated. analog systems. in nature. Some of the continuous signals are shown in
Fig. 1.2.1.
5. Complex mathematical It is extremely difficult to
algorithms can be easily implement complex
implemented. mathematical algorithms.

6. Easily upgradable Difficult to upgrade

1.1.4 Limitations of Digital Signal

Processing

ge
In spite of the obvious advantages of DSP systems over (a) (b)
the ASP systems, it would be unfair if we do not highlight
Fig. 1.2.1
io led
the disadvantages of the DSP systems.
These signals can be represented as,
1. A DSP system makes use of an ADC’s and DAC’s. These
devices slow down the process. Hence a DSP system is a) x(t) = A sin (t + )
ic ow

–t/RC
not as fast as an ASP system. b) x(t) = e
2. A typical DSP chip contains thousands of transistors  As can be seen in Fig.1.2.1, continuous time signals are
n
and hence the power consumption is more compared defined for every instance of time.
bl kn

to ASP systems.  Some of the examples of continuous time signals are

3. The most important limitation of DSP systems is that speech signals, vibrations, ECG, seismic waves etc.
at
Pu ch

they are not suitable for signals which have a huge  A discrete time signal is one that is defined only at
bandwidth. The ADC needs to sample the input signals specific values of time. Two discrete time signals are
Te

at twice the bandwidth of the input signal to avoid shown in Fig. 1.2.2.
aliasing. If the frequency content in the input signal is
very larger, the ADC will need to sample it at least
twice that frequency. There is a practical limitation on
the speed of the ADC. Hence it is advisable to use ASP
systems when the bandwidth of the input signal is very
large.
With this background, we now move to discuss signals
and systems.
We shall first study the various types of signals and
(a)
then study the different systems.

1.2 Signals

Signals can be broadly classified as,

1. Continuous and discrete time signals.
2. Continuous valued and discrete valued signals.
3. Deterministic and random signals.

(b)
Fig. 1.2.2
 DSIP (MU-Sem. 7-COMP) 1-5 Introduction to Digital Signal Processing

 We do not know what is the value between 0 and 1 or between 1 and 2. Hence a discrete time signal exists only at
specific values of time.
 A discrete time signal is obtained from a continuous time signal using the sampling operation.

ge
io led
ic ow
n
bl kn

Fig. 1.2.3 : Sampling of a Continuous time signal to obtain a Discrete time signal

 Hence a discrete time signal is an approximation of the continuous time signal. To get a good approximation we need
at
Pu ch

to take more samples in an interval i.e., reduce the sampling period (Ts) (increase sampling frequency). A typical
discrete time signal can be represented by the equation.
Te

x(n) = A cos n

We will discuss more about sampling in the next section.

 There are some signals that are inherently discrete in time and do not require the sampling of a continuous time
signal. One such example is the stock market data. In this the value of a stock is studied at every hour or every minute.
Nobody studies the stock value at every millisecond. Hence the data obtained would be discrete in time. Fig.1.2.4
represents a stock at different instances of time.

Fig. 1.2.4
 DSIP (MU-Sem. 7-COMP) 1-6 Introduction to Digital Signal Processing

1.2.2 Continuous Valued and Discrete Valued Signals

 The values of a continuous time or discrete time signal can be continuous or discrete. If the signal (whether
continuous or discrete) can take all possible values, it is said to be continuous valued.
 On the other hand if the signal can take only a finite set of values, it is called a discrete valued signal. To put it simply,
if the y-axis can be divided into infinite number of levels, it is called a continuous valued signal. If the y-axis can be
divided into only a finite number of levels, it is called as a discrete valued signal. One such discrete valued signal is
shown in Fig. 1.2.5. A discrete time signal having a set of discrete values is called a digital signal.

ge
io led
ic ow
n
Fig. 1.2.5 : Discrete valued signal which is also discrete in nature
bl kn

1.2.3 Deterministic and Random Signals

at
Pu ch

A signal which can be uniquely represented by a

mathematical formula is called a deterministic signal. Since
there is a mathematical formula, all past, present and
Te

future values are known and hence the name deterministic.

A signal which cannot be represented by a
mathematical formula is called a Random signal. Such
signals evolve over time and are unpredictable. Seismic
waves, voice signals etc. are examples of random signals. Fig. 1.3.1

This analog signal has three very nice properties

1.3 Concept of Frequency in the
associated with it.
Continuous and Discrete Time
1) For every fixed value of F, x(t) is periodic i.e.,
Signals
x(t + T) = x(t) (10 Hz, 20 Hz, 100 Hz … etc. are all
periodic signals.
1.3.1 Concept of Frequency in Continuous
2) All continuous time signals with distinct frequencies
Time Signals are distinct i.e., 10 Hz looks different from 11 Hz which
looks different from 100 Hz. Every frequency looks
A continuous time domain signal is represented as,
different.
x(t) = A cos (2 Ft + ) …(1.3.1) 3) Range of frequency, F is –  F  + .
Where F is the frequency in Hz and has a range Though frequency is a positive quantity, we consider
–   F  + . negative frequency as well from Euler’s formula.
One such continuous time signal is shown in Fig.1.3.1. Let us understand the concept of Negative
frequencies
 A continuous time signal is represented as,
x(t) = A cos (2 Ft + )
 DSIP (MU-Sem. 7-COMP) 1-7 Introduction to Digital Signal Processing
j –j
e +e
We know cos  =
2
j (2Ft + ) –j (2Ft + )
x(t) = A  
e +e

 2 
A –j (2Ft + ) A – j (2Ft + )
 x(t) = e + e Fig. 1.3.3
2 2
 Hence x(t) can be considered to be made up of two 1. A discrete time signal is periodic only if its frequency
phases moving in opposite directions. is a rational number.
Let us prove it
For periodicity x (n + N) = x(n)
Hence, for our x(n) to be periodic, we should have
A cos (2f(n + N) + ) = A cos (2fn + )
 A cos (2fn + 2fN) + ) = A cos (2fn + )

ge
This relation will be true only if
2fN = 2k
io led
k
 f =
N
k
ic ow

Fig. 1.3.2 Hence only if f = N (Rational number)

 As time progresses one phasor moves in the counter A cos (2f (n + N) + ) = A cos (2 fn + )
n
clockwise direction while the other moves in the
bl kn

Therefore x(t), which was always periodic, when

clockwise direction.
passed through a ADC remains periodic only if ‘ f ’ is a
 A positive frequency corresponds to a counter
at

rational number.
Pu ch

clockwise uniform angular motion and negative

2. Discrete time signals whose frequencies are
frequency corresponds to negative clockwise uniform
separated by 2k are identical.
Te

angular motion. Hence the range of Analog frequency

Consider x1(n) = A cos (2 f0n + )
is –   F  + .
We know 0 = 2f0.
 Remember, like negative numbers, negative
frequencies do not exist and are used for  x1(n) = A cos (0n + )
mathematical convenience. This signal has a frequency 0
 When we pass this x(t), having the three properties Consider another discrete time signal x2(n) which as a
just discussed, through an Analog to Digital Converter frequency (0 + 2)
(ADC) we obtain x(n). x2 (n) = A cos ((0n + 2)n + )
 We will now see if these 4 properties of x(t) are  x2 (n) = A cos (0n + 2n + )
retained in x(n).
= A cos (0n + )
1.3.2 Concept of Frequency in Discrete = x1 (n)

Time Signals Hence x2(n) looks identical to x1 (n) even though x1(n)
has a frequency 0 and x2 (n) has a frequency (0+ 2).
A discrete time domain signal in represented as,
This can be generalized to state that all time signals
`x (n) = A cos (2fn + ) –   n  +  …( 1.3.2)
xk(n) = A cos (k n + ) k = 0, 1, 2, 3
One such discrete time signal is shown in Fig. 1.3.3. where, k = 0 + 2k –   0  + 
are identical
Therefore while all x(t)’s were distinct, their discrete time
domain counterparts, x(n)’s are not
 DSIP (MU-Sem. 7-COMP) 1-8 Introduction to Digital Signal Processing

1 1 How did a large range –   F  +  get reduced to

3. The range of discrete frequency is – 2  f  + 2
1 1
– 2  f  + 2 ? What does this mean ? How can discrete
From point number (2) we write,
time signals have such low frequencies ?
All discrete time signals
These questions will be answered if we know what ‘f’
xk(n) = A cos (k n + ); k = 0, 1, 2, 3…
is.
where, k = 0 + 2K ; –   0  
This is what the next section is all about.
are identical.
This means only those frequencies which are in the 1.3.3 Converting a Analog Signal to a
range –   0  +  are distinct.
Discrete Time Signal
– +
 The process of converting an analog signal (continuous
∵ = 2f, we can write
signal) to a discrete time signal is known as sampling.
–   2f  + 
 Sampling converts x(t) to x(n) and is performed by an

ge
1 1
 – f+ Analog to Digital Converter (ADC).
2 2

Hence the distinct range of discrete time signals is

io led
1 1
only – 2  f  + 2

Therefore while x(t)’s had a range of –   F  + , their

ic ow

Fig. 1.3.4
discrete time domain counterparts, x(n)’s, have a

n
1 1 The sampling process can be visualized as a simple
range of f +
2 2
bl kn

switch which opens and closes at fixed intervals of

time, Ts.
Hence when x(t), which has a range –  F  + , is
at

Consider a switch which closes every Ts sec.

Pu ch

passed through a ADC, it becomes x(n) and has a frequency

1 1 Let, x(t) be a pure sinusoidal function given by the
range of – 2  f  + 2
formula.
Te

Note : The range of digital frequency is, x (t) = A cos (2  F t + ) … (1.3.3)

1 1 Here, F is the frequency in Hz.
–     +  OR – 2  f  + 2

Fig. 1.3.5

 Since the switch closes every T seconds, we get the output only at fixed intervals of time.
1
The sampling rate Fs = T
s

 From Fig. 1.3.5, it is clear that the x-axis which represents time gets replaced by nTs in the discrete domain.
 t = nTs in the discrete domain
 We can have thus convert an analog signal to a discrete time signal by replacing t=nT, where T is the Sampling period.
 DSIP (MU-Sem. 7-COMP) 1-9 Introduction to Digital Signal Processing

Fig.1.3.6

 Hence we obtain a digital signal that is represented as, Only if Fs  2F, will f remain in the distinct
x(nT) = x(n) = A cos (2  F n Ts + ) …(1.3.4) 1 1
range of – 2  f  + 2
1
 We know that Ts is the sampling period i.e, Ts = F , This is known as the Nyquist criteria.
s
where, Fs is the sampling frequency. If Fs  2F, f will no longer remain in the distinct range
1 and will give rise to aliasing.
 Substituting Ts = F in Equation (1.3.4) we get,
s 1 1
Therefore a range of – 2  f  + 2 implies that Fs  2F
x(n) = A cos 2 n + 

ge
F
 …(1.3.5)
 Fs 
 Hence, an Analog signal x(t) = A cos (2F t + ) can be
io led 1.4 Aliasing and Nyquist Rate
n
converted to a digital signal by replacing t = nT= F .  Aliasing is a phenomenon where high frequencies look
s
identical to low frequencies because of low sampling
F frequency.
ic ow

Let f =
Fs
 In order to represent a continuous signal faithfully, we
 Substituting this value of f in Equation (1.3.5), we get need to take as many samples as possible. This is done
n
the standard representation of a discrete time signal. by an Analog to Digital Converter having a high
bl kn

x(n) = A cos (2 fn + ) …(1.3.6) sampling.

F What is the ideal sampling required ?
at

where, f =
Pu ch

Fs
 Sampling theorem was introduced in 1949 by Shannon
1 1 and is also called as Shannon sampling’s theorem.
We know that the range of f is – 2  f  2
Te

 It is stated as ‘A continuous time signal x(t) can be

F completely represented in its sampled form and
Since f =
Fs recovered back from the sampled form if the sampling
1 F 1 frequency Fs  2F where F is the maximum frequency in
 –  
2 Fs 2 the continuous signal’.
– Fs Fs  A minimum sampling rate of 2F is called the Nyquist
  F+
2 2 rate. Low sampling frequency gives rise to aliasing.
1 1 – Fs Fs  We will see the effect of aliasing by solving a couple of
Hence – 2  f  + 2 basically means 2  F  + 2
examples.
 To understand this range we consider the positive
Solved Example
frequency.
Fs Ex. 1.4.1 : Consider the give continuous time signals having
F 
2 frequencies 5 Hz, 10 Hz, 15 Hz, 45 Hz, 50 Hz and 130 Hz.
i.e., Fs  2F …(1.3.7) x1 (t) = 5 cos 2  (5) t

Here F is the frequency of input signal (in Hz) while Fs x2 (t) = 5 cos 2 (10) t
is the sampling frequency (in Hz). x3 (t) = 5 cos 2 (15) t

 From Equation (1.3.7) we realize that sampling x4 (t) = 5 cos 2 (45) t

frequency Fs should be greater than or equal to twice x5 (t) = 5 cos 2 (50) t
the maximum frequency of the signal being sampled. x6 (t) = 5 cos 2 (130) t
Obtain the discrete time signals by using a sampling
frequency of 40 Hz.
 DSIP (MU-Sem. 7-COMP) 1-10 Introduction to Digital Signal Processing

Soln. : 9
Since 8 > 1, it can be written as,
We will first solve this example mechanically and then
x4 (n) = 5 cos 21 +
analyze the results obtained. 1 ∵ 9 = 1 + 1
n
8  8 8
The given signals have the following frequencies.
x4 (n) = 5 cos 2 + 2  n
1

x1 (t) = 5 Hz, x2 (t) = 10 Hz, x3 (t) = 15 Hz, x4 (t) = 45 Hz,  8
x5 (t) = 50 Hz and x6 (t) = 130 Hz, Since cos (2 + ) = cos  we write,

x4 (n) = 5 cos 2  n

A continuous time signal is given by the equation, 1
x (t) = A cos (2Ft + ) 8
1
We convert this to a discrete time signal by replacing  f4 = = 0.125 …(6)
8
n
t = nTs=F x5 (t) x5 (t) = 5 cos 2 50t
s

x5 (n) = 5 cos 2  n

50
x(n) = A cos 2 n +  
F

 Fs 
…(1) 40

ge
x5 (n) = 5 cos 2
5
 x(n) = A (cos 2 fn + ) …(2) 
4 n
F
Here = f
Fs
io led 5
Since > 1, it can be written as,
4
In our case Fs = 40 Hz
x5 (n) = 5 cos 21 +
1 ∵ 5 = 1 + 1
We will convert each of the given continuous signals into n
 4  4 4
ic ow

discrete time signals using Equation (1).

x5 (n) = 5 cos 2 + 2
1
x1 (t)  x1 (t) = 5 cos 2 5t  n
 4
n
x1 (n) = 5 cos 2  n
5
bl kn


40 Since cos (2 + ) = cos  we write,

x1 (n) = 5 cos 2  n x5 (n) = 5 cos 2

1 1

at

8 4 n
Pu ch

Comparing this with Equation (1) we have 1

 f5 = = 0.25 …(7)
1 4
Te

f1 = = 0.125 …(3)
8
x6 (t) x6 (t) = 5 cos 2130t
x2 (t)  x2 (t) = 5 cos 210t
x6 (n) = 5 cos 2
130 13

x2 (n) = 5 cos 2  n
10

40  40  n = 5 cos 2 4  n
13
x2 (n) = 5 cos 2  n
1
 Since > 1, it can be written as,
4 4

= 5 cos 23 +
1 1 ∵ 5 = 3 + 1
 f2 = = 0.25 …(4) n
4  4  4 4
x3 (t)  x3 (t) = 5 cos 215t
x6 (n) = 5 cos 6 + 2
1
 n
 4
x3 (n) = 5 cos 2  n
15

40
Since cos (6 + ) = cos  we write,
x3 (n) = 5 cos 2  n
3

8 x6 (n) = 5 cos 2
1
3 4 n
 f3 = = 0.325 …(5)
8 1
 f6 = = 0. 25 …(8)
x4 (t)  x4 (t) = 5 cos 245t 4

x4 (n) = 5 cos 2  n

45 We have thus been successful in converting the

40 6 continuous time signals into discrete time signals.
x4 (n) = 5 cos 2  n
9 Let us now analyze what we have got. We form a table

8 of the results obtained.
 DSIP (MU-Sem. 7-COMP) 1-11 Introduction to Digital Signal Processing

Sr. Continuous time Value of Is the Nyquist Discrete time signal Value of Is f distinct
No. signal analog condition satisfied ? Discrete
frequency Fs  2F frequency(f)
(Fs = 40 Hz)

x1 (n) = 5 cos 28 n

1. x1 (t) = 5 cos 25t F1 = 5 Hz Yes 1 1 Yes
 f1 = 8 = 0.125

x2 (n) = 5 cos 24 n

2. x2 (t) = 5 cos 210t F2 = 10 Hz Yes 1 1 Yes
 f2 = 4 = 0.25

x3 (n) = 5 cos 28 n

3. x3 (t) = 5 cos 215t F3 = 15 Hz Yes 3 3 Yes
 f3 = 8 = 0.375

x4 (n) = 5 cos 28 n

4. x4 (t) = 5 cos 245t F4 = 45 Hz No 1 1 No (f4 same as f1)
 f4 = 8 = 0.125

ge
x5 (n) = 5 cos 24 n
5. x5 (t) = 5 cos 2 50t F5 = 50 Hz No 1 1 No (f5 same as f2)
 f5 = 4 = 0.25

x6 (n) = 5 cos 24 n

1 1
6. x6 (t) = 5 cos 2130t
io led
F6 = 130 Hz No No (f6 same as f2)
 f6 = 4 = 0.25
ic ow

We observe that, x4 (n) looks identical to x1 (n). sampled all the signals at a sampling frequency of
Similarly x5 (n) and x6 (n) are identical to x2 (n). This is Fs  2  130 Hz
n
known as Aliasing. It is important to note that the  Fs  260 Hz
bl kn

continuous time signals x1 (t), x2 (t), x3 (t), x4 (t), x5 (t) and x6

Try solving the earlier example with Fs = 300 Hz and
(t) were all distinct having different frequencies. However
you will observe that x1 (n), x2 (n), x3 (n), x4 (n), x5 (n) and
at
Pu ch

on sampling, x4 (n) becomes identical to x1 (n). Similarly x5

x6 (n) are all distinct.
(n) and x6 (n) resemble x2 (n).
Given below is the MATLAB code for the above
Te

On performing digital to analog conversion again, we

example.
would not be able to distinguish between x4 (t) and x1 (t)
similarly we would not be able to distinguish between x5 (t), Aliasing
x6 (t) and x2 (t).
clc
In other words, while converting the discrete time
signal back to a continuous time signal, 45 Hz signal (x4 (t)) clear all
will look the same as a 5 Hz signal (x1 (t)). Similarly 50 Hz
signal(x5(t)) and 130 Hz signal (x6 (t)) will look the same as
Fs=40; %sampling Frequency
10 Hz signal (x2 (t)). This distortion has taken place because
the Nyquist condition was not satisfied while sampling F1=5;F2=10;F3=15;F4=45;F5=50;F6=130 ;
th
x4 (t), x5 (n) and x6 (n). (Look at the 4 column on the table). % various frequencies present
This phenomena of higher frequencies resembling low For n=1:1:50;
frequencies is known as aliasing. A simple way to avoid this
is to ensure that the sampling frequency Fs is always greater x1(n)=5*cos(2*pi*(F1/Fs)*n);
than or equal to twice the maximum frequency that needs x2(n)=5*cos(2*pi*(F2/Fs)*n);
to be sampled. x3(n)=5*cos(2*pi*(F3/Fs)*n);
Fs  2 Fmax
x4(n)=5*cos(2*pi*(F4/Fs)*n);
In the above example, x6 (t) had the maximum
x5(n)=5*cos(2*pi*(F5/Fs)*n);
frequency of 130 Hz. To avoid aliasing, we should have
x6(n)=5*cos(2*pi*(F6/Fs)*n);
 DSIP (MU-Sem. 7-COMP) 1-12 Introduction to Digital Signal Processing

end
subplot(6,1,1)
stem(x1)
ylabel(‘x1(n)’)
subplot (6,1,2)
stem(x2)
ylabel (‘x2(n)’)
subplot(6,1,3)
stem(x3)
ylabel(‘x3(n)’)

ge
subplot(6,1,4)
Fig. P.1.4.1(a)
stem(x4)
Ex.1.4.2 : Consider the analog sinusoidal signal
io led
ylabel(‘x4(n)’)
x (t) = 5 sin (500 t)
subplot(6,1,5)
If the signal is sampled at Fs = 1500 Hz,
ic ow

stem(x5) (i) What is the discrete time signal obtained after sampling.
ylabel(‘x5(n)’) (ii) Find the frequency of discrete time signal.
n
subplot(6,1,6) (iii) If sampling frequency Fs = 300 Hz then obtain discrete
bl kn

time signal.
stem(x6)
at

Soln. :
Pu ch

ylael(‘x6(n)’)
Given : x (t) = 5 sin (500 t)
(i) Fs = 1500 Hz
Te

Now discrete time signal is obtained by replacing

n
t = nTs =
Fs
500 n
 x (n) = 5 sin 
 1500 

 x (n) = 5 sin 
3n …(1)

This is the discrete time signal obtained after sampling.

(ii) The standard equation of the Discrete time signal is,
x (n) = A sin n …(2)
Comparing Equations (1) and (2) we get,

Fig. P. 1.4.1
 =
3
We observe that, x4(n) is identical to x1(n) where as ∵  = 2f
x5(n) and x6(n) are identical to x2(n). 1
 f =
If we simply change the sampling frequency (Fs) to 6
300 Hz in the program, each discrete time signal will be 
Hence the frequency of discrete time signal is  = 3
distinct. This is shown in Fig. P.1.4.1(a).
1
or f =6
 DSIP (MU-Sem. 7-COMP) 1-13 Introduction to Digital Signal Processing

(iii) Here Fs = 300 Hz

1.4.1 Anti Aliasing Filter
Discrete time signal is obtained by replacing
The minimum sampling frequency which is essential is
n
t = nTs = 2F. where F is the maximum frequency. This Fs = 2Fmax is
Fs
called the Nyquist rate.
500  n
x (n) = 5 sin 
 300  Note : Most of the naturally occurring signals are not
band limited. This means that they have infinite
 x (n) = 5 sin 
5n
 3  frequencies present in them (The extremely high

Now
5
= 2 –  frequencies have very small amplitudes, but they exist).
3  3 Nyquist criteria of Fs  2 Fmax will never be satisfied in

 x(n) = 5 sin 2 –  n such cases and we will always encounter aliasing. These
 3
Band unlimited signals are converted to Band limited
 
 x(n) = 5 sin – n signals by passing them through a low pass filter. The sole
 3 

ge
purpose of this low pass filter is to remove very high

x(n) = – 5sin  frequencies thereby making the signal band limited and
io led3  n hence removing the affects of aliasing. These low pass
This is the discrete time signal. filters are called anti-aliasing filters.

1.5 Basic Elements of a DSP System

ic ow

We have discussed the basics DSP systems in this chapter. We had drawn a basic block diagram of a DSP system in
n
Fig. 1.5.1. We will now draw a modified block diagram of a DSP system and explain each block.
bl kn
at
Pu ch
Te

Fig.1.5.1
1. Input signal : A signal is defined as any physical quantity that varies with time, space or any other independent
variable. Anything that carries some information can be called a signal.
2. Anti-Aliasing Filter : Most of the naturally occurring signals are not band limited. This means that they have infinite
frequencies present in them. Anti-Aliasing filters are basically low pass filters which are used to remove the high
frequencies from the input signal thereby making them band limited.
3. Analog to Digital Converter (ADC) : As the name suggests, ADC converts a continuous time signal x(t) to a discrete
time signal x(n).
4. Digital Signal Processor (DSP) : This is the heart of the DSP system. It processes the discrete time input signal x(n) and
gives us the required discrete time output signal y(n). One could either use a computer or dedicated DSP’s
manufactured by companies such as Motorola and Texas Instruments.
5. Digital to Analog Converter (DAC) : The output of the DSP is in digital form but most of the output devices only work
in continuous time signals. DAC converts a discrete time signal y(n) to a continuous time signal y(t) which can be then
fed to the out device.
6. Reconstruction filter: The output of the DAC might contain high frequencies due to the interpolation of sinc functions.
These spurious high frequencies are eliminated using the reconstruction filter which is basically a low pass filter.
We will now solve a few examples to understand the sampling concepts.
 DSIP (MU-Sem. 7-COMP) 1-14 Introduction to Digital Signal Processing

Thus discrete signal is obtained by substituting

1.5.1 Solved Examples
n n
t = =
Fs 80
Ex. 1.5.1 : An analog signal is given as :
= sin 
5 
 x (n) = sin  250
n 
x (t) = sin (10 t) + 2 sin (20 t) + 2 cos (30 t).
 80   4 n
(i) What is the Nyquist rate of this signal ?
Now,
5  2 – 3 
=
(ii) If the signal is sampled with sampling frequency of 20 Hz, 4  4 

 x (n) = sin  2 –
what is the discrete time signal obtained after sampling ? 3 
Soln. :
 4 

x (n) = – sin 
3 
Given :
 4 n
x (t) = sin (10 t) + 2 sin (20 t) + 2 cos (30 t)
Since Ideal interpolation is used, analog signal is
(i) This signal consists of following frequencies, obtained by replacing n by t Fs.
2F1 = 10, 2 F2 = 20 and 2F3 = 30  n = 80 t

ge
 F1 = 5 Hz, F2 = 10 Hz, and F 3= 15 Hz
 x (t) = – sin  
3
Thus the maximum frequency present in the signal is,  4 80 t 
 x (t) = – sin (60 t)
io led
Fmax = W = 15 Hz
Hence a 50 Hz signal gets reconstructed as a 60 Hz
 Nyquist rate = 2 Fmax = 30 Hz
signal.
(ii) The discrete time signal is obtained by substituting
ic ow

Ex. 1.5.3 : What should be the sampling frequency to avoid

n n
t = F = 20 in the given equation. aliasing for an analog signal represented as :
n
s
x (t) = cos (150 t) + 2 sin (300 t) – 4 cos (600 t)
bl kn

 x (n) = sin  10 
n 
+ 2 sin  20  
n
 20   20  Obtain the discrete time signal if this sequence is sampled at

+ 2 cos  30  
at

n FS = 400 Hz.
Pu ch

 20 
Does aliasing occur ? If yes, calculate the aliased
  3
= sin  frequencies from the original frequencies.
 2  n + 2 sin ()n + 2 cos  2  n
Te

Soln. :
but, sin (n) = 0
3  Given :
also = 2 –
2 2 x (t) = cos (150 t) + 2 sin (300 t)
 
 x (n) = sin   n + 2 cos  2 – n – 4 cos (600 t) ...(1)
2  2 
Here 2F1t = 150t, 2F2t = 300t and 2F3t = 600t
  – 
 x (n) = sin 
 2  n + cos  2  n  F1 = 75 Hz, F2 = 150 Hz, and F3= 300 Hz
but,  Maximum frequency Fmax = 300 Hz
  – 
cos  Thus the sampling frequency, required to avoid
 2  n = cos  2  n =0 aliasing is,

 x (n) = sin   n Fs  2 Fmax
2
 Fs  600 Hz
Ex.1.5.2 : A signal x (t) = sin (t ) of frequency 50 Hz is
The given signal is sampled at Fs = 400 Hz.
sampled using a sampling frequency of 80 Hz. Obtain the
recovered signal if ideal reconstruction is used.
Thus discrete signal is obtained by substituting
n n
Soln. : Given : t = = in Equation (1)
Fs 400
x (t) = sin t of frequency 50 Hz
n
 x (t) = sin (2 50t)  t =
400
Now , Fs = Sampling frequency = 80 Hz
 DSIP (MU-Sem. 7-COMP) 1-15 Introduction to Digital Signal Processing

x (n) = cos  150

n 
+ 2 sin  300
n  Ex. 1.5.5 : Consider the analog signal

 400   400 
xa (t) = 3 cos 2000t + 5 sin 6000 t + 10 cos 12,000t.
– 4 cos  600
n 
 400  (i) What is Nyquist rate for this signal ?

x (n) = cos 
3  3 (ii) If this analog signal is sampled at
 8  n + 2 sin  4  n Fs = 5000 samples/sec, what is the discrete-time signal
– 4 cos   n
3
...(2) obtained after sampling ?
 2
(iii) What is the analog signal ya(t) we can reconstruct from
3
The third terms of 2 can be written as the samples if we use ideal interpolation ?

 Soln. :
– 4 cos 2 – 2 n.
  Given :

= – 4 cos –  n xa(t) = 3 cos 2000t + 5 sin 6000t + 10 cos 12000t …(1)
 2
(i) The given signal consists of following frequencies,

but, cos   n = cos 
– 

ge
2  2  n =0 2 F1 = 2000t, 2 F2 = 6000t, 2 F3 = 12000 t
 F1 = 1000 Hz, F2 = 3000 Hz and F3 = 6000 Hz
x(n) = cos   n + 2 sin   n
3 3

 8  4
io led Thus Fmax = 6000 Hz = 6 kHz
Hence aliasing does occur.  Nyquist rate = 2 Fmax = 12 kHz
Ex.1.5.4 : An analog signal contains frequencies upto (ii) A discrete signal is obtained by substituting
ic ow

10 kHz n n
t = = in Equation (1)
Fs 5000
n
(i) What range of sampling frequencies allows exact
2000  n   6000 n 
 x(n) = 3 cos 
bl kn

reconstruction of this signal from its samples.

 5000  + 5 sin  5000 
(ii) If the signal is sampled with a sampling frequency
12000 n 
+ 10 cos 
at

Fs = 8 kHz
Pu ch

 5000 
What is the folding frequency ?
x(n) = 3 cos   n + 5 sin   n
2 6

Examine what happens to the frequency F1 = 5 kHz 5 5
Te

+ 10 cos 
Examine what happens to the frequency F2 = 9 kHz
12
 5 n
Soln. :
The second term  5  can be written as,
6
(i) From the Nyquist criteria we know that,  
Fs  2W
6 = 2 – 4
 Fs  2  10 5  5

5 sin   n = 5sin 2 –  n

 Fs  20 kHz 6 4

5  5
This is the range of sampling frequencies which allows
exact reconstruction. Since sin (2) = 0, we obtain

5 sin   n = 5 sin –  = – 5 sin  

6 4 4
(ii) The signal is sampled at 8 kHz.
5  5  5
Hence the folding frequency is, 12
Fs 8 kHz The third term, 5 can be written as,
Ffold = = = 4 kHz
2 2 12
= 2 + 2
F1 = 5 kHz (given) 5  5
Since Ffold is less than F1 aliasing takes place. 5 KHz will
 10 cos 2 +
2
= 10 cos  
2
look like 1kHz (5  4) = 1.  5 5
x(n) = 3 cos 
F2 = 9 kHz (given) 2 4 2

Since Ffold is less than F1 aliasing takes place. 9KHz will
 5  n – 5 sin  5  n + 10 cos  5 
x(n) = 13 cos   n – 5 sin   n
2 6
look like 1 kHz (9  (4+4)) = 1 kHz. 
5 5
 DSIP (MU-Sem. 7-COMP) 1-16 Introduction to Digital Signal Processing

(iii) The reconstructed signal is obtained by substituting Since sin(2) = 0, we obtain

 
x (n) = – cos   n + 2 sin  – n
n = t Fs in the discrete time signal.
Since Fs = 5000  2  2 
 n = 5000 t Ex. 1.5.7 : An analog signal is given as :

x(t) = 13 cos 
2 4 xa (t) = 15 cos(1250 t) + 17 cos (2170 t) + 33 cos (4750 t).

 5  5000 t – 5 sin  5  5000 t is converted to a discrete time signal. Determine the Nyquist
 x(t) = 13 cos 2000 t – 5 sin 4000t. rate, Folding frequency, resulting discrete time signal x(n) if
the sampling frequency is 625 Hz. Also write the discrete
Ex. 1.5.6 : An analog signal is given by :
frequencies in radians
x (t) = 3cos (100 t) + 2 sin (300 t) – 4 cos (100 t)
Soln. :
(1) What is the Nyquist rate of the sigal.
Given
(2) Write the equation of the sampled signal
xa (t) = 15 cos(1250 t) + 17 cos (2170 t) + 33 cos (4750 t).
(3) If this analog signal is sampled at
This signal consists of following frequencies,

ge
Fs = 200 samples/sec, what is the discrete-time signal
obtained after sampling ? 2F1 = 1250, 2 F2 = 2170 and 2F3 = 4750
 F1 = 625 Hz, F2 = 1085 Hz, and F 3= 2375 Hz
io led
Soln. : Given :
Thus the maximum frequency present in the signal is,
x (t) = 3cos (100 t) + 2 sin (300 t) – 4 cos (100 t)
Fmax = W = 2375 Hz
ic ow

This can be written as,

x (t) = – cos (100 t) + 2 sin (300 t) ...(1)  Nyquist rate = 2 Fmax = 4750 Hz
n
Here 2F1t = 100t and 2F2t = 300t Given : The signal is being sampled at 625 Hz
bl kn

 F1 = 50 Hz and F2 = 150 Hz, Fs 625

 The folding frequency = Ffold = = = 312.5 Hz.
2 2
Maximum frequency Fmax =150 Hz
at
Pu ch

 Nyquist rate = 2 Fmax = 300 Hz The discrete time signal is obtained by substituting

(2) The discrete signal is obtained by substituting n n

Te

t = = in the given equation.

Fs 625
n n
 x (n) = 15 cos  1250 
n 
+ 17 cos  2170
T = =
Fs 300
in Equation (1) n 
 625   625 
x (n) = – cos  100
n 
+ 2 sin  300
n 

+ 33 cos  4750 
n 
 300   300 
 625 

 x(n) = – cos   x (n) = 15 cos(2 )n + 17 cos (2 1.736 ) n
 3 n + 2 sin ( n)
Since sin ( n) = 0, we get + 33 cos (2 3.8)n

x(n) = – cos   n
This equation is of the form x(n) = A cos (0n)
 3
 x (n) = 15 cos(2)n + 17 cos (3.47 )n + 33 cos (7.6 )n
(3) Given : Fs = 200 Hz.
Hence the discrete frequencies in terms of  are,
The discrete signal is obtained by substituting
n n 1 = 2
t = F = 200 in Equation (1)
s 2 = 3.47 
3 = 7.6 
x (n) = – cos  100
n 
+ 2 sin  300
n 

 200   200 
 Ex. 1.5.8 : Consider the following analog signal
x (n) = – cos   n + 2 sin   n
3
 2  2 x(t) = 2 sin (100t)
3  The signal is sampled at x(t) is sampled with a sampling rate
Now = 2 –
2 2
Fs = 50 Hz. determine the discrete time signal.
  
x (n) = – cos   MU - Dec. 2016, 10 Marks
 2 n + 2 sin  2 – 2  n
 DSIP (MU-Sem. 7-COMP) 1-17 Introduction to Digital Signal Processing

Soln. : Given : x(t) = 2 sin (100t) …(1) Since Fs = 4000

The discrete signal is obtained by substituting  n = 4000 t

x(t) = 15 cos 
n n
t = = in Equation (1) 
Fs 50 2 4000t

 x (n) = 2 sin  100
n  x(t) = 15 cos 2 1000t
 50 
We observe that the reconstructed signal has only one
 x (n) = 2 sin 2n frequency of 1000 Hz.
We vary n and realize that x(n) = 0 for all values of n
 x (n) = { 0,0,0,…….} 1.6 Applications of DSP
Ex. 1.5.9 : Consider the analog signal Digital signal processing systems are used in a variety
x (t) = 5 cos 2(1000)t + 10 cos 2(5000)t is to be sampled.
of applications. Look around and we realize that every
(i) What is Nyquist rate for this signal ?
electronic system has a DSP in it. Some of the applications
are listed as follows :
(ii) If this analog signal is sampled at 4 kHz, will the signal be
1. Speech Recognition, Synthesis, Analysis.

ge
recovered from its samples ? MU - May 2018, 10 Marks
2. Image processing applications.
Soln. : (i)Nyquist rate
3. Robotic vision.
io led
Given : x (t) = 5 cos 2(1000)t + 10 cos 2(5000)t …(1) 4. Biomedical signal application like processing and
The given signal consists of following frequencies, analyzing ECG, EMG, CT, MRI, etc.
F1 = 1000 Hz and 5. Data compression.
ic ow

6. Military application like Radar signal processing, secure

F2 = 5000 Hz
communications.
Thus Fmax = 5000 Hz = 5 kHz
n
7. Industrial application like robotics, CNN.
 Nyquist rate = 2 Fmax = 10 kHz
bl kn

8. Communication application like voice commands,

(ii) Analog signal is sampled at 4 kHz . cellular telephones.
9. Robotic vision, vibration analysis.
at
Pu ch

A discrete signal is obtained by substituting

n n Summary
t = = in Equation (1)
Fs 4000
Te

This chapter deals with the basics of Digital signal

 x(n) = 5 cos 2 
1000 n   5000 n 
 4000  + 10 cos 2 4000  processing. A basic classification of signals is done. We have
 discussed the concept of frequency and explained the
x(n) = 5 cos  n + 10 cos   n
5

 
2  2 Nyquist criteria. Examples have been solved to understand
Aliasing. A block diagram of a DSP system is discussed. The
Now, in the second term  2  can be written as,
5
  chapter ends with listing down the applications of DSP.
5 = 2 +  Review Questions
2  2
  Q. 1 Explain advantages of Digital signal Processing over
10 cos   n = 10 cos 2 +  n = 10 cos   n
5

2  2  2 Analog Signal Processing

Hence we observe that 10 cos   n is identical to

5 Q. 2 Explain aliasing effect.
2 Q. 3 Define : Aliasing

10 cos   n. This is called aliasing.
2  Q. 4 What is aliasing observed in sampling process ? How
this can be avoided ?

 x(n) = 15 cos   n
2  Q. 5 Define a Nyquist rate.

The continuous time signal had two different Q. 6 Draw and explain basic elements of DSP.
frequencies while the discrete time signal has only one. Q. 7 Explain Concept of Frequency in Discrete Time
(iii) The reconstructed signal is obtained by substituting Signals.
n = t Fs in the discrete time signal


 2 Discrete Time Signal and System

Syllabus :
Classification of Discrete-Time Signals, Classification of Discrete- Systems
Linear Convolution formulation for 1-D and 2-D signal (without mathematical proof), Auto and Cross Correlation formula
evaluation, LTI system, Concept of Impulse Response and Step Response, Output of DT system using Time Domain
Linear Convolution

2.1 Introduction Solved Example

ge
Ex. 2.1.1 : Draw the graphical representation of the given
In chapter 1, we discussed the basics of digital signal signal x(n) = {1, 2, 2, 2, 3}
io led
processing. We also understood the method of obtaining a
discrete time signal form an analog signal. In this chapter Soln. : Since there is no arrow, the first element is the
we discuss various aspects of discrete time signals and origin.
ic ow

systems. We will also learn how to perform operations on

discrete time signals.
n
2.1.1 Representation of Discrete Time
bl kn

Signals
There are various methods of representing a discrete
at
Pu ch

time signal.

(1) Graphical representation

Fig. P. 2.1.1
Te

Consider a signal x(n) with values x(– 2) = 2, x(– 1) = 1,

An infinite sequence is represented by dots on both sides.
x(0) = 1.5, x(1) = 1, x(2) = 2, x(3) = 3, x(4) = 1.5. This signal
can be represented graphically as shown in Fig. 2.1.1. x(n) = {…, 1, 2, 4, 3,….}

(3) Functional representation :

The discrete time signal can also be represented as,
1 n = 1,3
x(n) = –4 n=0
0 else where

This can be graphically drawn as shown in Fig. 2.1.2.

Fig. 2.1.1

(2) Sequence representation

A more concise way of representing the same signal is
by representing it as a sequence shown as follows,
x(n) = {2, 1, 1.5 , 1, 2, 3, 1.5}

The arrow at the bottom indicates the origin.
If the arrow is not mentioned, then the first element is
the origin. Fig. 2.1.2
 DSIP (MU - Sem.7- COMP) 2-2 Discrete Time Signal and System

Solved Example

Ex. 2.1.2 : Represent the following signals graphically :

(i) x ( n ) = {1, 2, 0 , – 1, 1}

(ii) x ( n ) = {0, 0, – 1, 2, 3 }

(iii) x ( n ) = {0 , 1, – 1, 1, – 1)}

Soln. : These signals are as shown in Figs. P.2.1.2(a), (b) and (c) respectively.

ge
io led
ic ow
n
(a) (b) (c)
bl kn

Fig. P. 2.1.2 : Graphical representation of given sequences

at

2.1.2 Elementary Discrete Time Signals

Pu ch

There are a number of basic signals that we encounter regularly and which play an important role in signal processing.
Te

1. Delta or a unit impulse function

A unit impulse signal is denoted by (n) and is defined as,

1 ; n=0
(n) =
0 ; otherwise

The graphical representation of the unit impulse signal is shown in

Fig. 2.1.3.
Fig. 2.1.3
A unit impulse function is a signal that is zero everywhere except at
n = 0 where its value is unity.

2. Unit step signal

A unit step signal is denoted by u(n) and is defined as,

1 ; n0
u(n) =
0 ; otherwise

The graphical representation of the unit step signal is shown in

Fig. 2.1.4.
Fig. 2.1.4
 DSIP (MU - Sem.7- COMP) 2-3 Discrete Time Signal and System

3. Unit ramp signal

A unit ramp signal is denoted by ur(n) and is defined
as,
n ; n0
ur(n) =
0 ; otherwise

The graphical representation of the unit ramp signal is

shown in Fig. 2.1.5.

(b)

ge
io led
ic ow

Fig. 2.1.5
n
4. Exponential signal
bl kn

A exponential signal is of the form (c)

at

n
Pu ch

x(n) = a for all n

The graphical representation of the exponential signal

Te

is shown in Fig. 2.1.6. Four different ranges of ‘a’ are taken

to study the change in the exponential pattern.
Fig. 2.1.6 illustrates different type of discrete-time
exponential signals. When the value is 0 < a < 1, the
sequence decays exponentially, when value of a > 1, the
sequence grows exponentially. Note that when a < 0, the
(d)
discrete-time exponential signal takes alternating signs.
Fig. 2.1.6 : Exponential sequences

5. Sinusoidal signal
A discrete time sinusoidal signal is given by the
formula

x(n) = A cos ( n + )
Where A is the amplitude,  is the frequency in
radians and  is the phase.
An example of discrete-time sinusoidal signal is shown
in Fig. 2.1.7.

Fig. 2.1.6 (a)

 DSIP (MU - Sem.7- COMP) 2-4 Discrete Time Signal and System

ge
io led Fig. 2.1.7 : A sinusoidal sequence

Given below is the MATLAB code for generating x3(n) = a^n;

various elementary discrete time signals. Run the above end
program and see the results.
ic ow

figure (3)
%Plot Unit Step, Unit Ramp, Exponential and Sinusoidal
stem (x3)
Functions
n
bl kn

%SINUSOIDAL SIGNAL
clc
at

A = input (‘enter the amplitude for sinusoidal function :’)

Pu ch

clear all
w = input (‘enter the value of frequency for sinusoidal
function :’)
Te

%UNIT STEP
%Take w = 0.1*pi, 0.2*pi,….0.9* pi, pi
for n = 1 : 1 : 20% We take the length of the sequence to be
20 for n = 1 : 1 : 20

x1(n) = 1; x4(n) = A*cos (w*n) ;

end end
figure(1) figure(4)
stem(x1) stem(x4)

2.2 Classification of Discrete Time

%UNIT RAMP
Signals
for n = 1 : 1 : 20
x2 (n) = n ; Discrete time signals have different characteristics. We
end now classify the discrete time signals based on these
figure(2) characteristics. The discrete time signals are classified as
stem(x2) follows :
1. Even and odd signals (Symmetric and Anti-symmetric
% EXPONENTIAL SIGNAL signals)

a = input (‘enter the value of a for exponential signal :’) 2. Periodic and Aperiodic signals
for n = 1 : 1 : 20 3. Energy and power signals
 DSIP (MU - Sem.7- COMP) 2-5 Discrete Time Signal and System
2.2.1 Even and Odd Signals
A discrete time signal is said to be even (symmetric) if
it satisfies the following condition.

x(n) = x(– n) …(2.2.1)

A discrete time signal is said to be odd (Anti-

symmetric) if it satisfies the following condition.

x(n) = – x(– n) …(2.2.2)

For an even signal, x(1) = x(– 1), x(2) = x(– 2)…

(c) Odd signal
While for a odd signal, x(1) = – x(– 1), x(2) = – x(– 2)…
Given below are examples of even and odd signals.

ge
io led
ic ow

(a) Even signal

n
bl kn

(d) Odd signal

at
Pu ch

Fig. 2.2.1

A sine wave is a odd signal while a cos wave is a even

Te

signal,
Note : sin (50) = – sin (– 50)
cos (50) = cos (– 50)
Most of the signals that we encounter are neither
purely even or odd but are a combination of both. In other
(b) Even signal
words any arbitrary signal can be expressed as a sum of
even signals and odd signals.

Fig. 2.2.2
 DSIP (MU - Sem.7- COMP) 2-6 Discrete Time Signal and System
The even component is extracted by adding x(n) to
x(– n) and dividing by 2 i.e.
x(n) + x(– n)
xe(n) = …(2.2.3)
2
Similarly the even component is extracted by adding
x(n) to – x(– n) and diving by 2.
x(n) – x(– n)
Xo (n) = …(2.2.4)
2
Clearly Equation (2.2.3) satisfies the symmetric
condition in Equation (2.2.1) while Equation (2.2.4) satisfies (c)
the antisymmetric condition in Equation (2.2.2).
x (n) – x (– n)
Let us solve an example to understand this better. xo (n) =
2
Note : x (– n) is simplify the mirror image of x (n).

Solved Example

ge
Ex. 2.2.1 : Find the even and odd components of the
given discrete time signal
io led
x (n) = {1, 2, 3, 4 , 1, 2, 2}

Soln. : We shall solve this example graphically. To find the
ic ow

even and odd components, we need x (n) as well as x (– n).

(d)
We draw them one below the other.
Fig. P. 2.2.1
n
As stated earlier, x (– n) is simply the mirror image of
Note : If we add xe (n) and xo (n), we get back the original
bl kn

x (n).
signal ! !
at
Pu ch
Te

(a)
Fig. P. 2.2.1 (e)

(b)
Fig. P. 2.2.1 Fig. P. 2.2.1 (f)
x (n) + x (– n) x (n) = xe (n) + xo (n)
Now, xe (n) =
2
 DSIP (MU - Sem.7- COMP) 2-7 Discrete Time Signal and System

Solved Examples

Ex. 2.2.2 : Few discrete time sequences are given below

(i) cos (0.01  n) (ii) cos (3  n)

(iii) sin (3 n) (iv) cos ( n8 ) cos 8n 

Determine whether they are periodic or aperiodic. If a
sequence is periodic, determine its fundamental period.

Fig. P.2.2.1 (g) Soln. :

(i) Given :
Take any sequence and this property will always work.

2.2.2 Periodic and Aperiodic Signals x(n) = cos (0.01  n) ...(1)

A signal is periodic if there exist N such that We have the standard equation,

ge
x (n + N) = x (n) io led …(2.2.5) x(n) = cos n …(2)

Let us take an example. Check whether Comparing Equations (1) and (2) we get,
x(n) = A cos (n + ) is periodic.
 = 0.01 
ic ow

x (n) = A cos (n + )

But  = 2f
n
Here,  = 2f
 2  f = 0.01 
bl kn

 x (n) = A cos (2fn + ) 0.01  0.01

 f = =
2
at

x (n + N) = A cos (2f (n + N) + ) 2
Pu ch

Now,
For periodicity x (n + N) = x (n) 1
 f = cycles per sample …(3)
200
Te

 A cos (2f (n + N) +  ) = A cos (2fn + )

 A cos (2fn + 2fN + ) = A cos (2fn + ) Since frequency ‘f ’is expressed as the ratio of two integers;
this sequence is periodic. Now we have the condition of
The above equation will be true only if
periodicity,
2fN = 2k where, k is an integer.
k
k f = …(4)
 f = N
N
Here, N and k are both integers. Here ‘N’ indicates, the fundamental period.

Hence a discrete time signal is periodic only if it’s Comparing Equations (3) and (4)
frequency is a rational number (ratio of two integers). Fundamental period = N = 200 samples
If the input signal is expressed as a summation of two (ii) Given :
signals i.e., x (n) = x1 (n)+ x2 (n), we calculate f1 and f2
k1 k2 x (n) = cos (3  n) …(5)
where, f1 = N and f2 = N
1 2 Comparing with Equation (2) we get,
N1
The resultant signal x (n) is periodic if N is a ratio of  = 3
2

two integers.
 2f = 3
The period of x (n) will be the LCM of N1 and N2.
3
A few examples would make things clear.  f = cycles / sample …(6)
2
 DSIP (MU - Sem.7- COMP) 2-8 Discrete Time Signal and System

Since `f' is ratio of two integers; the given sequence is A discrete time complex exponential is periodic if its
periodic. Comparing Equations (4) and (6) we get, relative frequency is a rational number. We have the
standard equation of a discrete time complex exponential.
Fundamental period = N = 2 samples
2kn
j N
(iii) Given sequence is, x (n) = e ...(2)
x (n) = sin 3 n …(7) Comparing Equations (1) and (2) we get,
Comparing with Equation (2) we get, 2kn  2k 1
= n  =
 = 3 N 4 N 4
 2f = 3 k 1
3  =
N 8
 f =
2
Since it is ratio of two integers; the given sequence is
Here 2 is not an integer. That means ‘f ’ cannot be
periodic. It’s fundamental period is N= 8 samples.

ge
expressed as the ratio of two integers. In this case f is
x2 ( n ) = 3 sin  n
1
(ii) Given ...(3)
irrational. Thus the given sequence is non-periodic. 8 
(iv) Given sequence is,
io led We have the standard equation,

x (n) = cos
n   n x ( n ) = A sin  n ...(4)
 8  cos 8  ...(8)
Comparing Equations (3) and (4) we get,
ic ow

The standard equation can be expressed as,

1 1
 = 2f=
n
8 8
x (n) = cos1n cos2 n ...(9)
bl kn

1
 f =
Comparing Equations (8) and (9) we get, 16 
at
Pu ch

1 
1 = and 2 = Here  is an irrational number hence f is not the ratio of two
8 8
integers. Thus given sequence is non-periodic.
Te

but,  = 2  f
2.2.3 Energy and Power Signals
1 
 2  f1 = and 2  f2 = For a discrete time signal x (n), the energy is defined
8 8
as,
1 1
 f1 = and f2 =
16  16 +

2
E = | x (n) |
Hence f1 is ratio of non-integer values. So it is non-periodic.
n=–
While f2 is ratio of two integer values. So it is periodic. But
the total signal is multiplication of non-periodic and The reason for using magnitude square is so that the
periodic signals. So the resultant signal x(n) is non-periodic. definition applies to real valued as well as complex valued
signals. If E is finite, then x (n) is called an energy signal.
Ex. 2.2.3 : Find if the following sequences are periodic or
Many signals that possess infinite energy have finite power.
not. If yes find its fundamental time period. The average power of a discrete time signal x (n) is defined
n
(i) 1
()
x (n) = e 4
j
(ii) x2 (n) = 3 sin (18 n )
as
N


Soln. : lim 1 2
P = | x (n) |
N  2N + 1
 n
e (4 )
j
n=–N
(i) Given : x1 (n) = ...(1)
A signal x (n) is said to be a power signal if the average
power is finite.
 DSIP (MU - Sem.7- COMP) 2-9 Discrete Time Signal and System

Solved Example  n

E =   
1
…(3)
Ex. 2.2.4 : Obtain energy for the signal x ( n ) = a u ( n )
n 4
n=0
where | a | < 1.
We have the standard geometric series formula,
Soln. : For D.T. signal, energy is given by,

 1
 A
n
= 1+A+A +…=
2

E =  |x(n)|
2
...(1) n=0
1–A
n=–
1 4
The given equation of x ( n ) is,  E = =
1 3
1–
n 4
x(n) = a u(n) ...(2)
Here u( n ) represents unit step. It’s value is one for i.e. E is finite.
the range 0 to . Hence multiplication by u ( n ) does not
n n Note : If 0 < E < then the signal is energy signal. Since
change the amplitude of a ; but it indicates that a is

ge
present only from n = 0 to n = . the calculated value of energy is finite; the given signal is
Thus Equation (1) reduces to, energy signal.
io led

Ex. 2.2.6 : Determine the energy and power of signal
E =  | an |2 ...(3)
given by,
n=0
ic ow

Rearranging the term we get, x(n) = ( ) n0

1
2

n
n
= 3 n< 0
 2 n
bl kn

E = [a ] ...(4)
n=0 Soln. :The energy of signal is given by,

at

We use the geometric series formula,

Pu ch

 E = 
|x(n)|
2
… (1)
1 n=–
 A = 1 + A + A + A + …. =
n 2 3
if | A | < 1.
Te

1–A n
n=0
The given signal is, x ( n ) =   for the range n 0 and
1
2
Here A = a . Thus Equation (4) reduces to,
2
n
x ( n ) = 3 for n < 0. Thus Equation (1) becomes,
1
E =
2
2 if | a | < 1  n 2 –
1 
E =   
1–a
+  [3 ]
n 2

n
2 
Ex. 2.2.5 : x ( n ) = ( 0.5 ) u ( n ). State whether it is an n=0 n=–1
energy or power signal. Justify.  2 n –
1 
E = 

2  +  [ 3 ]
2 n
Soln. : First we will calculate the energy of signal x ( n ). It
n=0 n=–1
is given by,
 n –

=   +  (9)
1 n
…(2)
E = 
|x(n)|
2
… (1) 4
n=0 n=–1
n=–
Consider the first summation term. Using geometric series
The given signal is,
n formula we can express it as follows :
x ( n ) = ( 0.5 ) u ( n ) … (2)
 n
Since it is multiplied by unit step; this signal is present from  1
 4 =
1
=
4
…(3)
n = 0 to n = . Thus Equation (1) becomes, 1 3
n=0 1–
4
  2n

E =  [ ( 0.5 ) ] =   
1 n 2 Now consider the second summation term. To make
2 the limits of the summation positive, put n = – m. The limits
n=0 n=0
change as follows.
 DSIP (MU - Sem.7- COMP) 2-10 Discrete Time Signal and System

When n=–1 –m=–1 m=1 From the geometric progression formula,

When n = –   – m = –  m =  We know
–    1
 an =
m
; |a| < 1
  (9) =
n
 ( 9 ) =  9
–m  1
n=D
1–a
n=–1 m = +1 m=1 1
  E = = 1.125
Now use the standard summation formula, 1
1–
9

A Since E < , the given signal is a energy signal.
 A
n
=
1–A
n=1 Since the signal has finite energy, the power of the
 m
1 signal is zero.

 9
1 1/9 9 1
 = = = …(4)
1–
1 8/9 8 2.3 Operation on Signals
m=1 9

ge
Putting Equations (3) and (4) in Equation (2) we get, It is imperative for students to understand simple
4 1 manipulation techniques on discrete time signals. Signal
E = + processing involves modifying the original signal. This
3 8
io led
modification is achieved by performing different operations
35
 E = on discrete time signals. The mathematical transformation
24
from one signal to another is represented as,
ic ow

Since the energy is finite, the given signal is an Energy

signal. y (n) = T [x (n)]
n
If the energy of signal is finite then its power is zero.
bl kn

Here x (n) is the input, y (n) is the output and T is the

Thus the power of given signal is zero. operation that one performs.
at

The basic operations are :

Pu ch

 P = 0
1. Time shifting
Ex. 2.2.7 : Determine the energy and power of the signal 2. Time reversal
Te

given by, 3. Time scaling

n
1 4. Scalar multiplication
x(n) = 3 u(n)
  5. Signal addition and multiplication
MU - May 2018, 10 Marks
We shall discuss each one in detail.
Soln. :
n 2.3.1 Time Shifting
x(n) =   u(n)
1
3 The shift operation takes the input signal x (n) and
We know shifts the signal, resulting in either a delay or an
 advancement of the signal.
E = 
n = –
|x(n)|2
Mathematically it is represented as,
 1n 2 y(n) = x (n – k)
=  3 u(n)
n = –  
 Here k is the amount of shift required.

Since there is u(n), we change the limits of summation Let us see what this equation amounts to.
 n2
1 Solved Examples
  E = 
n=0
3
Ex. 2.3.1 : x (n) = {1, 2, 3, 4, 5}
 n 
 9
1
  E =
n=0 Find y (n) = x (n – 2)
 DSIP (MU - Sem.7- COMP) 2-11 Discrete Time Signal and System

Soln. :We draw x (n) Soln. :

Fig. P.2.3.1

y (n) = T [x (n)] Fig. P.2.3.2

ge
 y (n) = x (n – 2) io led y (n) = T [x (n)]

We take different values of n  y (n) = x (n + 2)

y (0) = x (– 2) = 0 (∵ the value of x(n) at n = –2 is 0) We take different values of n.
y (1) = x (– 1) = 0
ic ow

y (– 2) = x (0) = 1
y (2) = x (0) = 1 (∵ the value of x(n) at n = 0 is 1)
y (– 1) = x (1) = 2
n
y (3) = x (1) = 2
bl kn

y (0) = x (2) = 3
y (4) = x (2) = 3
y (1) = x (3) = 4
at

y (5) = x (3) = 4
Pu ch

y (6) = x (4) = 5 y (2) = x (4) = 5

Te

y (7) = x (5) = 0 y (3) = x (5) = 0

We plot these values

Fig. P. 2.3.2(a)
Fig. P. 2.3.1(a)
We hence note that y (n + 2) shifts the output to the left.
Hence we state that y (n) = x (n – 2) is a right shifted This is known as time advance.
version of x (n) i.e., y (n) is a delayed version of x (n). Hence for y (n – k), if k is > 1  Right shift  Time delay
Ex. 2.3.2 : x (n) = {1, 2, 3, 4, 5} if k is < 1  Left shift  Time advance

Find y (n) = x (n + 2)
 DSIP (MU - Sem.7- COMP) 2-12 Discrete Time Signal and System
2.3.2 Time Reversal Hence y (n) is a mirror image of x (n). It is a folded version
of x (n).
This is also known as the folding operation. It is We have used the time reversal operation while computing
mathematically given by the equation, the odd and even components of a signal.
y (n) = T [x (n)] 2.3.3 Time Scaling
y (n) = x (– n) This operation is related to change in time scale.
This can be very easily understood by taking the same There are two types of time scaling operations :
example. (a) Down scaling

Solved Example (b) Up scaling

Ex. 2.3.3 : x (n) = {1 , 2, 3, 4, 5} find y(n).

(a) Down scaling : It is mathematically represented as,

Soln. : y (n) = T [x (n)]

ge
y (n) = x (2n)
We shall explain this with an example.
io led
Solved Example

Ex. 2.3.4 : x (n) = {5, 4, 3, 2, 1 , 2, 3, 4, 5}

ic ow


Find y (n) = x (2n)
n
bl kn

Fig. P. 2.3.3
at
Pu ch

y (n) = T [x (n)]
y (n) = x (– n)
Te

We take different values of n

y (– 4) = x (– (– 4)) = x (4) = 5
y (– 3) = x (3) = 4
Fig. P. 2.3.4
x (– 2) = x (2) = 3
Soln. :
x (– 1) = x (1) = 2
y(n) = T[x(n)]
y (0) = x (0) = 1
y (n) = x (2n)
We take different values of n
y (0) = x (0) = 1
y (1) = x (2) = 3
y (2) = x (4) = 5
y (3) = x (6) = 0
y (– 1) = x (– 2) = 3
y (– 2) = x (– 4) = 5
y (– 3) = x (– 6) = 0
Fig. P.2.3.3(a)
 y (n) = {5, 3, 1, 3 , 5}

 DSIP (MU - Sem.7- COMP) 2-13 Discrete Time Signal and System

Fig. P. 2.3.5

Fig. P.2.3.4(a) y (n) is an expanded version of x (n). It is important to note

Hence we note that the signal gets compressed. It is that the shape of x(n) and y(n) remain the same. This is
also important to note that the shape of x (n) and y (n) known as up-sampling.
remain the same. This is known as down sampling.
2.3.4 Scalar Multiplication

ge
(b) Up scaling : It is mathematically represented as,
In this operation, the entire signal is multiplied by a
y (n) = x (n / 2)
io led scalar. This is mathematically represented by,
We take the same example. y (n) = T [x (n)]
Solved Example i.e., y (n) =  x (n)
This results in amplification of the signal.
ic ow

Ex. 2.3.5 : x (n) = {5, 4, 3, 2, 1 , 2, 3, 4, 5} find y(n)

 Solved Example
n
Soln. : Let x (n) = {1 , 2, 3, 4, 5}, find y (n) = 2  x (n)
bl kn

Ex. 2.3.6 :
y(n) = T[x(n)] 
at

y (n) = x (n / 2)
Pu ch

Soln. : y(n) = T[x(n)]

We take different values of n
 y (n) = 2  x (n)
Te

y (0) = x (0) = 1
y (1) = x (1/2) = No sample We take different values of n

y (2) = x (1) = 2 y (0) = 2 x (0) = 2  1 = 2

y (3) = x (3/2) = No sample y (1) = 2 x (1) = 2 2 = 4
y (4) = x (2) = 3
y (2) = 2 x (2) = 2  3 = 6
y (5) = x (5/2) = No sample
y (3) = 2 x (3) = 2  4 = 8
y (6) = x (3) = 4
y (4) = 2 x (4) = 2  5 = 10
y (7) = x (7/2) = No sample
y (8) = x (4) = 5
y (– 1) = x (– 1/2) = No sample
y (– 2) = x (– 1) = 2
y (– 3) = x (– 3/2) = No sample
y (– 4) = x (– 2) = 3
y (– 5) = x (– 5/2) = No sample
y (– 6) = x (– 3) = 4
y (– 7) = x (– 7/2) = No sample
y (– 8) = x (– 4) = 5 Fig. P. 2.3.6
 DSIP (MU - Sem.7- COMP) 2-14 Discrete Time Signal and System
2.3.5 Signal Addition and Multiplication All the operations discussed so far would be used to
build up complex operations like the convolution operation.
In these operations, signals are added or multiplied
taking into account of their instances of time (their values With classification of signals and operation on signals
of n). complete, we now move to discuss various systems.

Let us solve an examples to make things clear. Ex. 2.3.8 : Draw u(n) – u(n – 5)
Soln. :
2.3.6 Solved Examples on Signals
u(n) is a unit step while u(n – 5) is a unit step shifted
Ex. 2.3.7 : Given two signals x1 (n) = {1, 1, 0 , 1, 1} and
by 5 to the right. We draw them one below the other and
 perform subtraction.
x2 (n) = {2, 2, 0 , 2, 2}.

Perform (i) x1 (n) + x2 (n) (ii) x1 (n)  x2 (n)

Soln. : The first step is to place the two signals exactly one

ge
below the other.

Fig. P.2.3.8
io led
ic ow
n
Fig. P.2.3.7
bl kn

Fig. P.2.3.8(a)
at
Pu ch
Te

Fig. P.2.3.7(a)

Fig. P. 2.3.8(b)

  u(n) – u(n – 5) = {1, 1, 1, 1, 1}

Ex. 2.3.9 :
For x(n)= {3, 2, 1, 6, 4, 5}, plot the following discrete time
signals.
(i) x(n + 1)
Fig. P.2.3.7(b)
(ii) x(–n) u (–n)
(iii) x(n – 1) u(n – 1)
(iv) x(n – 1) u(n)
(v) x(n – 2) MU - May 2018, 10 Marks

Soln. :
We begin with drawing x(n).
Fig. P. 2.3.7(c)
Bear in mind that we add / multiply only the
corresponding terms.
 DSIP (MU - Sem.7- COMP) 2-15 Discrete Time Signal and System

Fig. P. 2.3.9

(i) x(n+1) Fig. P.2.3.9(d)

This is x(n) shifted to the left by 1.

ge
(iii) x(n – 1)  u(–n – 1)
 x(n + 1) = {3, 2, 1 , 6, 4, 5}

io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. P.2.3.9(e)
Fig. P.2.3.9(a)

(ii) x(–n)  u(–n)

Fig. P.2.3.9(f)

Fig. P.2.3.9(b)

Fig. P.2.3.9(g)

All values are zero.

Fig. P.2.3.9(c)
 DSIP (MU - Sem.7- COMP) 2-16 Discrete Time Signal and System

(iv) x(n – 1) u(n) Ex. 2.3.10 : For x(n) = {1, 2, –1, 5, 0, 4}, plot the following
discrete time signals.
(i) x(n + 3) (ii) x(–n – 2)
(iii) x(n – 2) (n – 2) (iv) x(2n)
MU - Dec. 2018, 10 Marks
Soln. : We begin with drawing x(n)

Fig. P.2.3.9(h)

ge
io led
Fig. P.2.3.10
Fig. P.2.3.9(i)
ic ow

(i) x(n + 3)
n
This is x(n) shifted to the left by 3.
bl kn
at
Pu ch
Te

Fig. P.2.3.9(j)

(v) x(n – 2)
Fig. P.2.3.10(a)
This is x(n) shifted to the right by 2.
(ii) x(–n – 2)

Fig. P.2.3.9(k)
Fig. P.2.3.10(b)
 DSIP (MU - Sem.7- COMP) 2-17 Discrete Time Signal and System

(iii) x(n – 2) (n – 2) 3. Linear and Non-linear system

4. Time Variant and Time invariant system
5. Stable and Unstable system
We shall discuss each of them.

1. Static and Dynamic systems

 A system is said to be static if the output of the system

depends only on the present input and not on the past
or future input.
 Given below are a few examples of static system,
y (n) = T [x (n)]
Fig. P.2.3.10(c)
(i) y (n) = 4 x (n)

ge
(ii) y (n) = log x (n)
(iii) y (n) = A cos x (n)
 In each case, y (n) requires only the present value of
io led
input x (n)
Fig. P.2.3.10(d) i.e. from y (n) = log x (n), we write
ic ow

y (1) = log x(1)

y (2) = log x(2)
n
 Static system are also called memoryless system as we
bl kn

do not need to store pervious input values


at

A dynamic system is one in which the output y (n)

Pu ch

Fig. P.2.3.10(e)
depends on the present as well as past inputs.
(iv) x(2n)
For example, y (n) = x (n) + x (n – 2)
Te

This is time compression also known as down  In this case, to find out y (4), we need x (4) which is the
sampling. present input as well as x (4 – 2) = x (2) which is the
past input. Hence to find out entire y (n), we need
present inputs as well as part inputs.
Some examples of Dynamic systems are :
x (n) + x (n – 12)
1. y (n) = 2
2. y (n) = x (n) – x (n – 1)
Fig. P.2.3.10(f)
 Since past input values need to be stored and
2.4 Discrete Time Systems retrieved, the system requires a certain amount of
memory. Hence Dynamic systems are also called
memory systems.
As defined earlier, a discrete time system is a device or
an algorithm that operates on a discrete time signal. It is 2. Causal and Non-causal system
mathematically represented as y (n) = T [x (n)]
 A system is said to be causal if the output of the
Discrete time systems are classified according to their system y (n) at any time n depends only on present
general characteristics. and past inputs but does not depend on future inputs.
1. Static and Dynamic system (Memory less and Memory i.e. y (n) = x (n) + x (n – 2) + x (n – 16)
systems)  If the output of a system also depends on future inputs
2. Causal and Anti-causal system it known as a Non-Casual systems.
 DSIP (MU - Sem.7- COMP) 2-18 Discrete Time Signal and System

For example, y (n) = x (n) + x (n + 1) Here the system multiples the input x(n) with n.
 In this case, to find out say y (4), we need x (4) which is We use two inputs, x1(n) and x2(n)
the present input as well as x (5) which is the future We check if T [x1 (n) + x2 (n)] = T [x1 (n)] + T [x2 (n)]
input. (1) T[x1(n) + x2(n)]
 Hence in an non-causal system, we need to predict the
future inputs to get the present input. Given below are
a few examples of Causal and Non-Causal systems.
x (n) + x (n – 1) + x (n – 2)
1. y (n) = 3  causal

2. y (n) = log (x (n))  causal

Fig. P. 2.4.1
3. y (n) = x (2n)  Non-causal
4. y (n) = x (n) – x (n + 2)  Non-causal  y1 (n) = n [x1 (n) + x2 (n)]

3. Linear and Non-Linear Systems Now,

ge
 A system is said to be linear if it satisfies the
superposition principle.
io led
 Superposition principle states that the response of the
system to a weighted sum of signals is equal to the
corresponding weighted sum of output of the system
ic ow

to individual input signals.

i.e. T [ax1 (n) + bx2 (n)]= aT [x1 (n)] + bT [x2 (n)]
n
A block diagram will make things clearer.
bl kn
at
Pu ch

Fig. P. 2.4.1(a)

 y2 (n) = n [x1 (n) + x2 (n)]

Te

Fig. 2.4.1

and  y1 (n) = y2 (n), hence system is linear.

x(n)
(2) y (n) = e
We know,

Fig. 2.4.2 y (n) = T [x (n)]

Note : The system T is said to be linear if y1 (n) = y2 (n). Hence the system in this case computes the exponential
of the input.
Solved Example
We use two inputs, x1 (n) and x2 (n)
Ex. 2.4.1 : Check whether the following systems are linear
We check for T [x1 (n) + x2 (n)] = T [x1 (n)] + T [x2 (n)]
or non linear.
1. y (n) = n x(n)
x(n)
2. y (n) = e
2
3. y (n) = x (n)

Soln. : 1. y (n) = n x (n)

We know
Fig. P.2.4.1(b)
y (n) = T [x (n)]
y1 (n) = e[x1 (n) + x2 (n)]
y (n) = n[x (n)]
 DSIP (MU - Sem.7- COMP) 2-19 Discrete Time Signal and System
Now, 4. Time Variant and Time Invariant Systems
A system is Time invariant (Also called shift invariant) if
its input-output characteristics do not change with time. If
y (n) is the response of the system to input x (n), then
y (n – k) will be the response of the system to input x (n – k)
i.e. if the input gets shifted by k samples, then the output
also shifts by k samples.

(a) (b)
Fig. P.2.4.1(c) Fig. 2.4.3
x1(n) x2(n)
y2 (n) = e To check whether the system in Time invariant or not,
Hence y1 (n)  y2 (n); Hence system is unstable. we perform the following steps :
2 Step 1 : Delay input x (n) by k samples and observe the

ge
(3) y (n) = x (n)
output. We call this output y (n, k).
We know y (n) = T [x (n)]
Step 2 : Delay the output y (n) by k samples i.e., in the
io led
Hence the system computes the square of the input
given equation y (n), replace n by (n – k). We call
x(n).
this output y (n – k).
We use to inputs x1 (n) and x2 (n) and check it
If y (n, k) = y (n – k) the system is said to be time
ic ow

T [x1 (n) + x2 (n)] = T [x1 (n)] + T [x2 (n)] invariant or else it is called time variant.
Solved Example
n
bl kn

Ex. 2.4.2 : Check whether the following systems are time

invariant or time variant.
at

x(n)
Pu ch

1. y (n) = e 2. y (n) = x (n) + x (n – 1)

Fig. P. 2.4.1(d) 3. y (n) = n x (n)

Te

Soln. :
 y1(n) = [x1(n) + x2(n)]2
x(n)
1. y (n) = e
Now,
We know y (n) = T [x (n)]

Fig. P.2.4.2

Step 1 : Delay the input

Fig. P. 2.4.1(e)
Fig. P.2.4.2(a)
2 2
y2 (n) = x1 (n) + x2 (n)
 y (n, k) = e
x (n – k)

 y1(n)  y2 (n) Step 2 : Delay the output

Hence the system is Non-Linear.

Fig. P.2.4.2(b)
 DSIP (MU - Sem.7- COMP) 2-20 Discrete Time Signal and System

 y (n – k) = e
x (n – k)

Since y (n, k) = y (n – k), system is Time Invariant.

2. y (n) = x (n) + x (n – 1)
y (n) = T [x (n)] Fig. P.2.4.2(c)

Step 1 : Delay the input

Fig. P.2.4.2(d)
Step 2 : Delay the output

ge
io led Fig. P.2.4.2(e)
Since, y (n, k) = y (n – k), the system is Time Invariant.
3. y (n) = n x (n)
y (n) = T [x (n)] Fig. P.2.4.2 (f)
ic ow

Step 1 : Delay the input

n
bl kn

Fig. P.2.4.2 (g)

at

= n  x (n – k)
Pu ch

y (n, k)

Step 2 : Delay the output

Te

y (n – k) = (n – k) x (n – k)
Since y (n, k)  y (n – k), the system is Time Variant. Fig. P.2.4.2 (h)

5. Stable and Unstable systems 2.4.1 Representation of a Signal in Terms of

Weighted Sum of Shifted Discrete
A system is said to be stable if for a Bounded input, the Impulse
system produces a bounded output.
Any arbitrary signal can be represented as a
Let Mx be a finite number i.e., Mx<. Input is said to be summation of shifted and scaled impulses,  (n). Let us
bounded if | x (n) |  Mx . show this with a simple example. Consider an input signal
Similarly let My is finite number i.e., My<. Output is x(n).
said to be bounded if | y (n) |  My . x(n) = {0.5, 2, 1, 0.5 , 1, 2, 0.5}

Solved Example

Ex. 2.4.3: Check whether the following systems are stable.

x(n)
i) y (n) = e ii) y (n) = x (2n)
Soln. :
In each of the cases, if x (n) is finite, y (n) will also be
finite. Hence both systems are stable.
Fig. 2.4.4
 DSIP (MU - Sem.7- COMP) 2-21 Discrete Time Signal and System

The sample x(0) can be obtained by multiplying x(0),  In general, we have

the magnitude, with a unit impulse function (n). 
x(0) ; for n = 0
i.e., x(0) . (n)=  x(n) =  x(n)  (n – k)
 0 ; for n  0
k =–
Hence a signal can be viewed as a summation of scaled
and shifted impulses. This equation will be used in deriving
an important equation known as the Convolution sum.

2.5 Impulse Response and Convolution

Fig. 2.4.5  As discussed earlier, a simple block diagram of a

system is shown in Fig. 2.5.1.

ge
Corel 2 Corel 3
io led Fig. 2.5.1

 If the input x (n) is a unit impulse  (n), then the output

of the system is known as the impulse response of the
system and in denoted as h (n).
ic ow

Fig. 2.4.6

We could get the other components of x(n) by shifting

n
the delta function. Fig. 2.5.2
bl kn

x(–1) ; for n = –1
i.e., x(–1) . (n + 1) =   h (n) = T [ (n)]
 0 ; for n  –1
at
Pu ch

 The impulse response completely characterizes the

system.
Te

 We have shown in the earlier section that any

arbitrary input x (n) can be represented as a weighted
sum of discrete impulses.
Fig. 2.4.7 +

In a similar manner
i.e. , x (n) =  x (k)  (n – k)
k=–
x(–2) ; for n = –2
x(–2) . (n + 2) = 
 0 ; for n  –2 We know,
x(–3) ; for n = –3
x(–3) . (n + 3) = 
 0 ; for n  –3 y (n) = T [x (n)]
x(1) ; for n = 1
x(1) . (n – 1) = 
 0 ; for n  1
 + 
= T x (k)  (n – k)
 
y (n) …(2.5.1)
x(2) ; for n = 2 
x(2) . (n – 2) =  k = –  
 0 ; for n  2
x(3) ; for n = 3  If we assume the system to be linear, then
x(3) . (n – 3) = 
 0 ; for n  3 Equation (2.5.1) can be written as,

 x (n) = x(–3) (n + 3) + x(–2) (n + 2) + x(–1) +

    (n + 1) + x(0) (n) + x(1) (n – 1) + x(2)

y (n) =  x (k) T [  (n – k) ]
k=–
(n – 2) + x(3) (n – 3)
Let, T [  (n – k) ] = h (n, k)
 DSIP (MU - Sem.7- COMP) 2-22 Discrete Time Signal and System
i.e., Let the response to the shifted impulse sequence It involves the folding operation, shifting operation,
be denoted by h(n, k). multiplication operation and the summation operation.
+ In the original formula of convolution, the range of
 y (n) =  x (k)  h (n, k) …(2.5.2) k is –   k +  and the range of n is –   n  + .
k=– However this ranges depend on the length of x(n) and
h(n).
 If we now assume the system to be time invariant,
Range of n
h (n, k) = h (n – k) …(2.5.3)
 Substituting Equation (2.5.3) in Equation (2.5.2), Consider the following notations,
we get , y1 = Lowest range of y(n)
+ yh = Highest range of y(n)
y (n) =  x (k)  h (n – k)
xl = Lowest range of x(n)
k=–

ge
xh = Highest range of x(n)
 This equation is called the convolution sum and is one
of the most important formula in Digital Signal hl = Lowest range of h(n)
io led
Processing and Image Processing applications. hh = Highest range of h(n)
 It states that for a Linear Time-Invariant (LTI) system, if
The range of n is y(n) will be equal to
ic ow

the input sequence x(n) and impulse response h(n) are

known, y(n) can be found out from the convolution yl = xl + hl and yh = xh + hh
n
sum.
Let us consider an examples to understand this better.
bl kn

The convolution sum is represented as,

2.5.2(A) Solved Examples on Graphical
Method
at
Pu ch

y (n) = x (n) * h (n) where, * denotes the

convolution operation. Ex. 2.5.1 : What will be the range of y(n) if x(n) = {1,2, 3,1}

Te

2.5.1 Computation of Linear Convolution h(n) = {1, 2, 2, – 1}


There are two methods of computing linear
Soln. :
convolution.
1. Graphical method Here x(n) = { 1 , 2 , 3 , 1 }

2. Tabular method x(0) x(1) x(2) x(3)
 xl = 0, xh= 3
We will begin with discussing the graphical method
and then move on to the tabular method which is very easy Similarly,
to perform. h(n) = { 1 , 2 , 2 , – 1 }

2.5.2 Linear Convolution using Graphical
h(–1) h(0) h(1) h(2)
Method
 hl = – 1, hh= 2
Graphical method gives us a pictorial representation of
how complex the convolution operation is Linear Now yl = xl + hl
convolution is given by the formula.
+ = 0–1=–1

y(n) =  x(k) h(n – k) …(2.5.4) and yh = xh + hh

k=–
= 3+2=5
 DSIP (MU - Sem.7- COMP) 2-23 Discrete Time Signal and System

Hence the range of n will be – 1 n 5. (1) Draw x(k) and h (– k)

 The output will have values,
{y(–1), y(0), y(1), y(2), y(3), y(4), y(5)}

Range of k

The range of k will be the same as that of the

sequence x(n). In the above example, since xl = 0 and xh= 3,
the range of k will be 0  k  3.
Hence for the above example the formula of Fig. P. 2.5.2
convolution reduces to,
3
y(n) =  x(k) h (n – k); –1n5
k=0

ge
Ex 2.5.2 : Consider an audio signal x(n) = {1, 2, 4}. This
signal is filtered using a low pass filter having an impulse
io led Fig. P. 2.5.2(a)
response h(n) = {1, 1, 1}. Obtain the filtered output signal.
Soln. : Filtering a signal implies the convolution operation. y(0) =  x(k) h(–k) = 1  1 = 1
ic ow

Always remember this. The filtered output is given by the (2) Shift h(– k) to the right
formula.
n
y(n) = x(n) * h(n)
bl kn

+
i.e. y(n) = ∑ x(k)h(n–k)
at
Pu ch

k=–

When n = 0, the formula reduces to

Te

Fig. P. 2.5.2(b)
+
y(0) = ∑ x(k)h(–k)
k=–
i.e. we need x(k) and h(– k)
The steps involved are as follows :

(1) Draw x(k) and h(– k) one below the other.

(2) Multiply corresponding elements of x(k) and h (– k) Fig. P. 2.5.2(c)

and add the resultants.
y(1) =  x(k) h(1 – k)
(3) Shift h(– k) to the right and repeat step (2). Continue
shifting to the right till there is no overlap. y(1) = 1(1) + 2(1) = 3

(4) Shift h(– k) to the left and repeat step (2).Continue

shifting to the left till there is no overlap.

x(k) = {1, 2, 4}, h(k) = {1, 1, 1}

Since there are no arrows, the first elements of x(k) and

h(k) are taken as the origin.
Fig. P. 2.5.2(d)
 DSIP (MU - Sem.7- COMP) 2-24 Discrete Time Signal and System

y(2) =  x(k) h(2 – k)

y(2) = 1(1) + 2(1) + 4(1) = 7

Fig. P. 2.5.2(i)

Fig. P. 2.5.2(e)

y(3) =  x(k) h(3 – k)

y(3) = 1(0) + 2(1) + 4(1) = 6

ge
io led Fig. P. 2.5.2(j)

y(–1) = 0
y(1) =  x(k) h(–1 – k) = 0
ic ow

We see there is no overlap between x(k) and h(–1 – k)

hence the resultant is 0. All subsequent left shifts will give
Fig. P. 2.5.2(f)
n
us zero values.
bl kn

y(4) =  x(k) h(4 – k)  y(n) = {y(0), y(1), y(2), y(3), y(4)}

at

y(4) = 1(0) + 2(0) + 4(1) = 4

Pu ch

 y(n) = {1, 3, 7 , 6, 4}
Now for the 5thshift we have
Hence the filtered audio signal is y(n) = {1, 3, 7, 6,4}.
Te

There are two ways to check the result.

(i) Length of y(n) = Length of x(n) + Length of h(n) – 1

= 3+3–1=5

We observe that length of y(n) obtained is also 5.

Fig. P. 2.5.2(g) (ii) Sum of elements of x(n) × sum of elements of
h(n) = sum of elements of y(n).
i.e. x(n) h(n) = y(n)
in our case
x(n) = 1 + 2 + 4 = 7
h(n) = 1 + 1 + 1 = 3
Fig. P. 2.5.2(h)  x(n) h(n) = 7  3 = 21
y(n) = 1 + 3 + 7 + 6 + 4 = 21
y(5) =  x(k) h(5 – k) Now
 x(n) h(n) = y(n)
We see there is no overlap between x(k) and h(5 – k)
Hence the result obtained is correct.
hence the resultant is 0. All subsequent right shifts will give
in zero values. Ex 2.5.3 : Compute linear convolution of the following
sequences x(n) = { 1,2, 3, 1}, h(n) = {1, 2, 2, – 1}
We now shift h(–k) to the left  
 DSIP (MU - Sem.7- COMP) 2-25 Discrete Time Signal and System
+
Soln. : y(n) = ∑ x (k) h (n – k)
k=–

We draw x(k) and h(– k) and obtain y(0).

Fig. P. 2.5.3(d)
y(3) = 2 (– 1) + 3(2) + 1(2) = 6

Fig. P. 2.5.3

ge
Fig. P. 2.5.3(e)
io led
y(4) = 3(– 1) + 1(2) = – 1
ic ow

Fig. P. 2.5.3(a)
n
y(0) = 1(2) + 2(1) = 4
bl kn

We shift h(–k) to the right till there is no overlay

Fig. P. 2.5.3(f)
at
Pu ch

y(5) = 1(– 1) = – 1
beyond this there will be no overlap between x(k) and
Te

h(n – k). Hence we stop shifting h(– k) to the right.

We now shift h(–k ) to the left.

Fig. P. 2.5.3(b)

y(1) = 1(2) + 2(2) + 3(1) = 9

Fig. P. 2.5.3(g)

Fig. P. 2.5.3(c)

y(2) = 1(–1) + 2(2) + 3(2) + 1(1) = 10

Fig. P. 2.5.3(h)

y(–1) = 1(1) =1
 DSIP (MU - Sem.7- COMP) 2-26 Discrete Time Signal and System

We assume arrows are at x(1) and h(0).

We could interchange the positions of x(n) and h(n).
Step 2 : Multiply corresponding elements of x(n) and
h(n)

Fig. P. 2.5.3(i)

y(–2) = 0

Since there is no overlap, we stop shifting h(–k) to the left.

 y(n) = {1, 4, 9,10,6,–1, –1}

We will check if this result is correct

ge
Fig. 2.5.4
length of y(n) = length of x(n) + length of h(n) – 1
Step 3 : We separate the elements diagonally
= 4+4–1=7
io led
We shade the section where the arrows intersect. In
We observe that y(n) obtained has 7 elements this case the arrows intersect at x(1)h(0) hence we have
shaded the entire diagonal section.
ic ow

Also
x(n) h(n) = y(n)
n
x(n) = 1 + 2 + 3 + 1 = 7
bl kn

h(n) = 1 + 2 + 2 – 1 = 4
at

x(n) h(n) = 7  4 = 28
Pu ch

Now y(n) = 1 + 4 + 9 + 10 + 6 – 1 – 1 = 28
Te

Hence the result obtained is correct.

2.5.3 Linear Convolution using Tabular

Method Fig. 2.5.5
We have learnt how to perform linear convolution
using the graphical method. There is a much easier way to
Step 4 : We add the elements in each section to obtain
performing linear convolution. This is the tabular method.
y(n). The sum of elements in the shaded section
In this method we follow the given steps. gives us y(0)
Step 1 : Form a matrix as shown in Fig. 2.5.3.
 y(–1) = x(0)  h(0)

y(0) = x(0) h(1) + x(1) h(0)

y(1) = x(0) h(2) + x(1) h(1) + x(2)h(0)

y(2) = x(1) h(2) + x(2) h(1)

y(3) = x(2) – h(2)

 y(n) = {y(–1), y( 0), y(1), y(2), y(3)}

Fig. 2.5.3
We will solve a couple of example to understand the
tabular method better.
 DSIP (MU - Sem.7- COMP) 2-27 Discrete Time Signal and System
2.5.3(A) Solved Examples on Tabular Method Soln. : Determine the response simply means obtaining
Ex. 2.5.4 : Perform linear convolution of x(n) = {1,1,0,1, 1} y(n).
 Given x(n) and h(n), obtaining y(n) means performing
h (n) = {1, – 2, – 3, 4}
linear convolution.

Soln. :We form a matrix of x(n) and h(n) y(n) = x(n) * h(n)

We use the tabular method.

ge
io led
Fig. P. 2.5.5
 y(n) = {1, 4, 8, 8, 3, – 2, – 1}
ic ow

This is the output response of the system.

n
Fig. P. 2.5.4 Length of y(n) = Length of x(n) + Length of h(n) – 1
bl kn

We add elements from each diagonal section. Since the = 4+4–1=7

at

arrows intersect at 0, we shade that entire segment. The sum

Pu ch

Also, x(n) h(n) = y(n)

of this section will be the origin value.
x(n) = 1 + 2 + 3 + 1 = 7
 y(n) = {1, – 1, – 5, 2, 3, – 5, 1 , 4}
Te

 h(n) = 1 + 2 +1–1 = 3
 x(n)h(n) = 73 = 21
We check whether this result is correct.
Now y(n) = 1 +4 +8 +8 + 3–2 – 1
Length of y(n) = Length of x(n) + Length of h(n) – 1 = 21
 Our result is correct.
= 5+4–1=8

The length of y(n) obtained = 8.

2.6 Correlation
Correlation is basically used when we want to
Also, x(n) h(n) = y(n)
compare two signals. It is a very important operation in
x(n) = 1 + 1 + 0 + 1 + 1 = 4 signal processing and image processing. Correlation is a
h(n) = 1 – 2 – 3 + 4 = 0 measure of the degree to which two signals are similar.
Correlation as a mathematical operation closely
 x(n) h(n) = 4  0 = 0
resembles convolution.
Now y(n) = 1 – 1 – 5 + 2 + 3 – 5 + 1 + 4 Correlation of two separate signals is known as cross
=0 correlation, while correlation of the signal with itself is
known as auto correlation.
 Our result is correct.
1. Cross correlation : Correlation between two signals
Ex. 2.5.5 : The impulse response of a linear time invariant
x(n) and y(n), is called cross correlation. It is given by
system is h(n) = {1, 2, 1, – 1}. Determine the response of the
 the formula
system to the input x(n) = {1, 2, 3, 1}
 DSIP (MU - Sem.7- COMP) 2-28 Discrete Time Signal and System
+  rxy = {1, – 1, – 5, 2 , 2, – 3, 4}
r xy (l) =  x (n)  y (n – 1) 
n=–
Thus we get
2. Auto correlation : Many times it is necessary to
correlate the signal with itself for example rxy(l) = x (n) * y (– n)
determination of time delays between transmitted
= { 1, – 1 ,– 5, 2 , 2, – 3, 4 }
and received signal.

Hence the two signals are generated from the same
source. It is given by the formula Ex. 2.6.2 : Determine the auto-correlation of the following
signal :
+
x ( n ) = { 1 , 2, 1, 1 }
rxx (l) =  x (n) x (n – l) 
n=–
Soln. :
Simple method to calculate correlation Let x1 ( n ) = { 1 , 2, 1, 1 },

ge

We will use one of the important properties of
x2 ( n ) = { 1 , 2, 1, 1 }
correlation. It is as follows :
io led 
rxy ( l ) = x ( n ) * y ( – n ) …(2.6.1) We use the tabular method.
ic ow

Observe Equation (2.6.1), The L.H.S. indicates cross

correlation of x ( n ) and y ( n ). R.H.S. term indicates the
n
convolution (‘*’) operation. It is the convolution of x ( n )
bl kn

and y ( – n ). Here y ( – n ) is the folded version of y ( n ).

So this equation implies :
at
Pu ch

(i) Take signal x ( n ) as it is.

(ii) Fold sequence y ( n ) to obtain y ( – n ).
Te

(iii) Obtain the convolution of x ( n ) and y ( – n ).

(iv) The result of convolution will be same as the
correlation.
Fig. P.2.6.2
Let us solve the same example using the method.
 rxx = {1, 3, 5, 7 , 5, 3, 1}
2.6.1 Solved Examples on Correlation 
Ex. 2.6.1 : x ( n ) = { 1, 1, 0 ,1} and y ( n ) = { 4, – 3, – 2, 1 }
  Summary
Soln. : In the tabular form simply flip y (n).
In this chapter we discussed representation of Discrete
time signals. We classified discrete time signals as well as
discrete time systems. How to operate on discrete time
signals was understood. In this chapter we derived one the
most important formula which is known as convolution.
Two methods of computing the convolution were explained
and examples were solved. We also discussed correlation in
this chapter

Fig. P. 2.6.1
 DSIP (MU - Sem.7- COMP) 2-29 Discrete Time Signal and System
Q. 6 With example explain the time scaling and time
Review Questions
reversal operations performed on discrete time
signals.
Q. 1 Define discrete time system. Explain any three
properties with suitable example. Q. 7 With example, explain how linear convolution can be
obtained using graphical method.
Q. 2 Define a periodic signal
Q. 8 Explain in brief different properties of convolution.
Q. 3 Define : Even signal and Odd signal
Q. 9 Explain why the result of linear and circular
Q. 4 Define power signal. convolution is not same.
Q. 5 Sketch and define two standard DT signals (n) and Q.10 How N point circular convolution can be obtained
u(n). using linear convolution ?



ge
io led
ic ow
n
bl kn
at
Pu ch
Te
 3 Discrete Fourier Transform (DFT)

Syllabus :
Introduction to DTFT, DFT, Relation between DFT and DTFT, IDFT, Properties of DFT without mathematical proof
(Scaling and Linearity, Periodicity, Time Shift and Frequency Shift, Time, Reversal, Convolution Property and Parsevals‘
Energy Theorem). DFT computation using DFT properties. Transfer function of DT System in frequency domain using
DFT. Linear and Circular Convolution using DFT, Convolution of long sequences, Introduction to 2-D DFT Circular
Convolution (without mathematical proof), Linear convolution using Circular Convolution. Spectral Analysis using FFT

ge
 Since colours are nothing but wavelengths and
3.1 The Fourier Transform
io led wavelengths are related to frequency, splitting of
white light into different colours is actually a form of
 Fourier transform is named after the French
frequency analysis. Newton then placed another prism
mathematician Jean Baptiste Joseph Fourier who was
ic ow

upside down and showed that constituent colours

born in 1768 in Auxerre, France. The Fourier transform
blended back to give white light again. Fig. 3.1.2.
is one of the most important transforms in
n
engineering. Most of the branches of engineering owe
bl kn

their growth to the wonderful Fourier transform. The

at

Fourier transform gave a new direction and

Pu ch

understanding to the field of Digital Signal Processing.

So what exactly is the Fourier transform?
Te

 To understand this, Let us look at a simple experiment

that we have all conducted in school. Consider Fig. 3.1.2
Fig. 3.1.1.
Let us summarize what is stated.
 The first prism splits white light into different colours
(wavelengths) and the second prism takes these
different colours and reproduces white light.
 Joseph Fourier studied the works of Isaac Newton and
showed that what the prism does to light, the Fourier
transform does to signals.

Fig. 3.1.1  White light is replaced by a signal, the first prism by

the Fourier transform and the second inverted prism
 White light when passed through a prism gets split
by the Inverse Fourier transform.
into different wavelengths (colours). Isaac Newton
who was born in 1643, in England submitted a paper in  The Fourier transform breaks up a signal of any shape
1672 to the Royal Society where he used the word into pure sinusoidal functions. The inverse Fourier
Spectrum to describe the continuous band of colours transform combines the pure sinusoidal functions to
produced by this apparatus. Newton was the first to give back the original signal. To put it simply, the
note that white light can be split into its constituent Fourier transform tells us the frequencies that are
colours using a prism. present in a given signal.
 DSIP (MU-Sem.7-COMP) 3-2 Discrete Fourier Transform (DFT)

Fig. 3.1.3

3.2 Discrete Time Fourier Transform

 Discrete time Fourier transform, as the name suggest

ge
is used when the signal is discrete in time and is a
periodic. It is given by the formula.
+
io led Fig. P. 3.2.1

– jn
X () = x (n) e
n=  1 ; n=0
(n) = 
 0 ; otherwise
ic ow

 Since x(n) is discrete, we use the summation ()

operation. The DTFT is given by the formula,
n
 Hence x(n) is the discrete time signal and X() are the +
bl kn


– jn
frequency components present in the signal. X () = x (n) e
n=–

at

We can obtain x(n) from X () by using the Inverse

Pu ch

+


Discrete Time Fourier Transform (I.D.T.F.T), It is given – jn
 X () =  (n) e
by the formula, n=–
Te

+
Since  (n) exists only at n = 0.
∫ X()
1 + jn
 x (n) = e d
2
–  X () =  – jn
1. e =1e
jo

n=0
 Here we use the integral operation because X() is
continuous.  X () = 1
FT
 The DTFT does not exist for every periodic sequence. A   (n)  1
sufficient condition for the existence of the DTFT for a
Ex. 3.2.2 : Obtain DTFT of a unit step signal u(n)
periodic signal x(n) is.
+ Soln. : A unit step signal is shown in Fig. P. 3.2.2.
 | x (n) | < 
n=–

 In other words, The DTFT exists only if the signal x(n) is

absolutely summable.

3.2.1 Solved Examples on DTFT of Standard

Signals
Fig. P. 3.2.2
Ex. 3.2.1 : Obtain DTFT of unit impulse,  (n)
1 ; n≥0
u (n) =  0 ; otherwise
Soln. : A unit impulse is shown in Fig. P. 3.2.1. 
 DSIP (MU-Sem.7-COMP) 3-3 Discrete Fourier Transform (DFT)
L–1
The DTFT is given by the formula,

– j n
+ = A (e )


– jn n=0
X () = x(n) e
n=– We know from the geometric of progression formula.
N
Since u(n) exists only for n ≥ 0, we change the limits of the 2
aN1 – a0N2 – 1

n
summation. a =
1–a
 K=N
1

– jn
X () = e
Here N1 = 0, and N2 = L – 1
n=0

X () = A 
 1 – e j L
 

– j n
= (e )  1 – e – j 
n=0
FT  1 – e –jL
 Rectangular pulse A  A  –j 
We know from the geometric progression formula,  1–e 


ge

n 1 Ex. 3.2.4 : Obtain DTFT and sketch the magnitude spectrum
 a = |a|<1
1–a
n=0 for,
1
io led
 x () = X (n) = u (n) – u (n – 4)
1 – e– j
FT 1 Soln. :
 u (n) 
ic ow

1 – e– j Here u(n) is a step input while u(n – 4) is a unit step shifted

by 4.
n
Ex. 3.2.3 : Obtain DTFT of rectangular pulse
bl kn

x (n) = A Hence,
0  n  L–1 = 0 ; otherwise
at
Pu ch

Soln. : The rectangular pulse is an shown in Fig. P. 3.2.3.

Te

Fig. P. 3.2.4

Fig. P. 3.2.3

This is similar to a unit step except that a unit step is infinite

in length.
The DTFT is given by the formula,
+ Fig. P. 3.2.4(a)

– jn
 X () = x(n) e
n=–

We change the limits of the summation,

L –1
– jn
 X () = 
A e
n=0
 Fig. P. 3.2.4(b)

– jn
 A e
n=0  x (n) = {1,1,1,1}


 DSIP (MU-Sem.7-COMP) 3-4 Discrete Fourier Transform (DFT)

The DTFT is given by the formula, We sample  such that

+ 2k
 =

– jn … (3.3.2)
 X () = x(n)e N
n=–
Where, N = Number of sample.
In our case
3 Substituting Equation (3.3.2) is Equation (3.3.1), we get,


– jn
 x () = 1. e N –1 – j 2kn
n=0 X (k) =  x(n)e 
… (3.3.3)
n=0
Since the limits of the summation are small, we open up the
summation. N is normally equal to the length of the input signal x (n)
– jo – j1 – j2 – j3 Equation (3.3.3) is called the DFT and is one of the most
 X () = e +e +e +e
– j – j2 – j3
important transforms in engineering.
 X () = 1 + e

ge
+e +e
3.4 Discrete Fourier Transform (DFT)
We need to sketch the magnitude spectrum, io led
– j
e = cos  – j sin   The DFT is a modification of the Fourier transform so
– j
 |e | = cos2  + sin2  = 1 as to use it for discrete time signals and to obtain a
discrete frequency spectrum. We shall now proceed to
ic ow

Similarly, explain the DFT.

– j2
2 2
n
|e | = cos 2 + sin 2 = 1  Remember, the objective of the DFT is to obtain the
bl kn

– j3
and |e | = cos2 3 + sin2 3 = 1 frequencies present in the signal.
 Consider a discrete time input signal x(n) of length N.
 | X () | = 1 + 1 + 1 + 1 = 4
at
Pu ch

The DFT of x(n) is denoted by X(k) and is given by,

 | X () | = 4 …for all  N–1 – j 2kn

 
Te

 The Magnitude Spectrum is shown in Fig. P. 3.2.4(c). X(k) = x(n) e

n=0

k = 0, 1, 2,  N – 1 (3.4.1)
 Since the input signal x(n) is of length N, the
summation varies from 0 to N – 1 and Equation (3.4.1)
is called as N-point DFT.
 As mentioned in the earlier section, Equation (3.4.1)
Fig. P. 3.2.4(c)
acts as the prism, breaking down the input signal into
Note : The spectrum is continuous. various frequency components.
 We can obtain x(n) from X(k) using the Inverse
3.3 Relationship between DTFT and
Discrete Fourier Transform IDFT which is given by the
DFT
formula,
N–1 + j 2kn


The DTFT is given by the formula, 1 
x(n) = X(k) e
+
N
k=0


– jn
X (k) = x (n) e … (3.3.1) n = 0, 1, 2,  N – 1 (3.4.2)
n=
Equation (3.4.2) is called as N–point IDFT.
Since the DTFT is continuous, we sample the  axis to
make it discrete.
 DSIP (MU-Sem.7-COMP) 3-5 Discrete Fourier Transform (DFT)

 This acts as the inverted prism which gives us the Ex. 3.4.2 : Obtain DFT of (n – n0).

original signal x(n) by combining the constituent Soln.: Here x(n) =  (n – n0)
frequencies.  (n – n0) is basically  (n) shifted to the right by n0.
1
Note :   (n – n0) is 1 at n0 and 0 at all other values of n.
N in Equation (3.4.2) is a scaling function.

 This book uses the scaling function in the IDFT formula

while other books choose to use it in the DFT formula.
1
Some other books use in both DFT and IDFT
N
formulae.
1
 MATLAB uses N in the IDFT formula and hence we too Fig. P.3.4.2

choose to use it there.

ge
The DFT is given by the equation,
 Unless specified the length of x(n) and X(k) are the N–1 – j 2kn

same.  X(k) =  x(n) e 

; k = 0, 1,  N – 1
io led
n=0
3.4.1 Solved Examples on DFT of Simple N–1 – j 2kn
Signals
 X(k) =   (n – n0) e 
ic ow

In most of the examples, we begin by changing the n=0

limits of the summation depending on the input signal. Since  (n – n0) exist at n = n0, for all other values of
n
bl kn

Ex. 3.4.1 : Obtain DFT of a unit impulse signal  (n). the summation, the result will be equal to zero.
Soln.: Here x(n) =  (n) – j 2 k n0
at
Pu ch

 X(k) =  (n – n0) e N
The DFT is given by the equation,
N–1 – j 2kn – j 2 k n0

 
; k = 0, 1,  N – 1
Te

X(k) = x(n) e  X(k) = 1 · e N

(˙.˙  (n – n0) = 1)
n=0
N –1 – j 2 k n0
– j 2kn DFT
  (n)  (n – n0)  e N

 X(k) = e
n=0
Similarly,
+ j 2k n0
DFT
 (n + n0)  e N

Ex. 3.4.3 : Obtain DFT of u (n) – u (n – 4).

Soln.: Here x(n) = u (n) – u (n – 4)
Let us draw u (n) – u (n – 4) to understand what the
input is.
Fig. P.3.4.1

Since  (n) exists only at n = 0, for all other values of

the summation the result will be equal to zero.
– j 2k·0

 X(k) =  (0) e =  (0) · e0
 X(k) = 1 (˙.˙ () = 1
DFT
 Hence,  (n)  1 Fig. P.3.4.3
 DSIP (MU-Sem.7-COMP) 3-6 Discrete Fourier Transform (DFT)

(ii) k = 1
3 – j21 · n
X(1) =  x(n) e 4

n=0
– j21 · 0 – j21 · 1
4 4
= x(0) e + x(1) e
– j21 · 2 – j21 · 3
4 4
+ x(2) e+ x(3) e
Fig. P.3.4.3(a)
–j
= x(0) e0 + x(1) e 2
–j3
+ x(2) e –j + x(3) e 2
–j –j3
= 1+1·e 2 + 1·e –j + 1·e 2
 

ge
= 1 + 1 cos – j sin 
Fig. P.3.4.3(b)  2 2

+ 1 (cos  – j sin ) + 1 cos – j sin 

3 3
io led
u(n) is a unit step, while u (n – 4) is a unit step shifted  2 2
to the right by 4. = 1 + (0 – j) + (– 1 – 0) + (0 + j)
From the Fig. P. 3.4.3(a) it is clear that,
ic ow

= 1–j–1+j
x(n) = { 1, 1, 1, 1} X(1) = 0

n
(iii) k = 2
bl kn

The DFT is given by the equation,

N–1
3 – j22 · n


– j 2kn 4
 
; k = 0, 1, 2, N – 1
X(2) = x(n) e
at

X(k) = x(n) e
Pu ch

n=0
n=0
– j22 · 0 – j22 · 1
We now change the limits of the summation as x(n) = x(0) e 4
+ x(1) e 4
Te

exists only from 0 to 3. – j22 · 2 – j22 · 3

4 4
N = 4 + x(2) e + x(3) e

Range of k is also 0 to 3. = x(0) e + x(1)–j + x(2) e–j2 + x(3) e–j3

0

3 – j 2kn = 1 + 1 (cos  – j sin ) + 1 · (cos 2 – j sin 2)

 X(k) =  x(n) e 4
; k = 0, 1, 2, 3, + 1 (cos 3 – j sin 3)
n=0
= 1 + (– 1 – 0) + (1 – 0) + (– 1 – 0) = 0
For each value of k, we vary n from 0 to 3.
 X(2) = 0
(i) k = 0
3 – j20 · n
(iv) k = 3
X(0) =  x(n) e 4 3 – j23 · n
n=0 X(3) =  x(n) e 4

– j20 · 0 – j20 · 1 n=0

4 4
= x(0) e + x(1) e – j23 · 0 – j23 · 1
4 4
– j20 · 2 – j20 · 3 = x(0) e + x(1) e
4 4 – j23 · 2 – j23 · 3
+ x(2) e + x(3) e
4 4
+ x(2) e + x(3) e
= x(0) e0 + x(1) e0 – j3 – j9
0
+ x(2) e + x(3) e 0 = x(0) e0 + x(1) e 2 + x(2) e–j3 + x(3) e 2

= 1 + 1 cos – j sin 
=
3 3
x(0) + x(1) + x(2) + x(3)  2 2

+ 1 · (cos 3– j sin 3) + 1 cos – j sin 

= 1+1+1+1 9 9
 X(0) = 4  2 2
 DSIP (MU-Sem.7-COMP) 3-7 Discrete Fourier Transform (DFT)

= 1 + 1 · (0 + j) + 1 (– 1 – 0) + (0 – j)  X (0) = 1
= 1+j–1–j
= 0 For k = 1

  X(3) = 0 2 – j2n

  X(k) = { X(0), X(1), X(2), X(3)}

X (1) =  x(n) e 3

n=0
  X(k) = {4, 0, 0, 0} – j2 – j4
DFT  0
X (1) = x (0) e + x (1) e 3
+ x (2) e 3
Hence, {1, 1, 1, 1}  {4, 0, 0, 0} – j2 – j4
Let us draw the two sequences obtained.  X (1) = 1 – 1  e 3
+1e 3

 X (1) = 1 – (– 0.5 – j 0.866) + (– 0.5 + j 0.866)

 X (1) = 1

For k = 2

ge
2 – j2n
X (2) =  x(n) e 3
io led
Fig. P.3.4.3(c) n=0
– j4 – j8

Ex. 3.4.4 : Compute the DFT of the sequence given as  0

X (2) = x (0) e + x (1) e 3
+ x (2) e 3

x(n) = (– 1)n for

ic ow

 X (2) = 1 – (– 0.5 + j 0.866) + (– 0.5 – j 0.866)

(i) N=3  X (2) = 1
n
(ii) N = 4 Thus DFT is, X (k) = {1, 1, 1 }
bl kn

Soln.:
(ii) N = 4
at
Pu ch

(i) N=3
The range of ‘n’ is n = 0 to 3.
The range of ‘n’ is n = 0 to N – 1 that means n = 0 to 2. The sequence x(n) can be written as,
Te

Given sequence is , x (n) = (– 1)n .

x (n) = { 1, – 1, 1, – 1}
For n = 0  x (0) = (– 1)0 = 1 
n = 1  x (1) = (– 1)1 = – 1
According to the definition of DFT,
n = 2  x (2) = (– 1)2 = 1 N–1 – j 2k n
 x (n) = { 1, – 1, 1 } X (k) =  x(n) e N
… k = 0 to N – 1
 n=0
According to the definition of DFT 3 – j2k n
N–1 – j2kn  X (k) =  x(n) e 4
… k = 0 to 3
X (k) =  x(n) e N
…k = 0 to N – 1 n=0
n=0 3 – j kn
2 – j2kn  X (k) =  x(n) e 2

 X (k) =  x(n) e 3
…k = 0 to 2 n=0
n=0 3 3
For k = 0 For k = 0 X (0) =  x(n) e = 0
 x(n)
2 – j20n
n=0 n=0
X (0) =  x(n) e 3
 X (0) = x (0) + x (1) + x (2) + x (3)
n=0 = 1–1+1–1
X (0) = x (0) e0 + x (1) e0 + x (2) e0  X (0) = 0
 X (0) = x (0) + x (1) + x (2) = 1 – 1 + 1
 DSIP (MU-Sem.7-COMP) 3-8 Discrete Fourier Transform (DFT)
– j 2 k n0
For k = 1
DFT N
3 – jn
And  (n – n0)  e
X (1) =  x(n) e 2
DFT
– j 2 k·2
4
n=0  3  (n – 2)  3 e = 3e–jk
–j – j 2 k·3
–j 3 k
2 –j DFT 4
 0
X (1) = x (0) e + x (1) e + x (2) e and 2  (n – 3)  2 e = 2e
2

– j3
+ x (3) e 2  DFT of the given sequence is,

 X (1) = 1 – (– 0 – j 1) + (– 1 – j 0) – (0 + j 1) X(k) = 1 + 3e–j k + 2 e– j 3 k

Compute DFT of x (n) = cos  N k0 n .

 X (1) = 0 2
Ex. 3.4.6 :
 
For k = 2 Soln. :
3
e j  + e– j 
X (2) =  x(n) e– j  n We have, cos  =

ge
2
n=0
Thus given sequence can be written as,
 X (2) = x (0) e0 + x (1) e– j + x (2) e– j2 + x (3) e– j3
+ j 2  k0 n – j 2  k0 n
1  
io led
 X (2) = 1 – (– 1 – j 0) + (1) (1 – j 0) – 1 (– 1 – j 0) x (n) = e
N
+e
N

2
 X (2) = 4
According to the definition of DFT,
ic ow

For k = 3
N–1 –j2kn
3  N
n
– j3n X (k) = x(n) e
 2
bl kn

X (3) = x(n) e n=0

n=0 + j 2  k0 n – j 2  k0 n –j2kn
N–1
– j3 1  
 e
at


N N
 e N
Pu ch

 X (3) = x (0) e0 + x (1) e 2

+ x (2) e– j3 X (k) = +e
2
– j9 n=0
2
+ x (3) e N–1 + j 2  k0 n –j2kn
Te

1
 X (3) = 1 – (0 + j 1) + (– 1 – j 0) – (0 – j 1)  X (k) = 
2  e N e
N

 X (3) = 0 n=0

Thus DFT is, N–1 – j 2  k0 n –j2kn

1
X (k) = {0, 0, 4, 0} + 
2  e
N
e
N

n=0
Ex. 3.4.5 : Obtain the DFT of the sequence n
x(n) =  (n) + 3  (n – 2) + 2  (n – 3) 1  j 2 ( nk0 – k )
N–1

2 
 X (k) = e 
Soln. : The given sequence is shown below. The length of
n=0
the sequence is from 0 to 3.
N – 1 – j 2  ( k0 + k ) n
 1  
N = 4 +
2   e
N

DFT n=0
We know,  (n)  1
Now we have standard summation formula,

N2  a N1 – a N2 + 1
 a1
 an =  1–a
n = N1  N2 – N1 + 1 a = 1

Fig. P.3.4.5 Here N1 = 0, N2 = N – 1.

 DSIP (MU-Sem.7-COMP) 3-9 Discrete Fourier Transform (DFT)

 
0 N–1+1 – j 21 · 0 – j 21 · 1
j 2  (k0 – k)
  j 2 ( Nk0 – k ) 4 4
1 e  –e 
N = x(0) e + x(1) e
 X (k) = 
2 j2 (k0 – k)  – j 2 1 · 2 – j 21 · 3

 
4 4
N + x(2) e + x(3) e
1–e
– j

 
– j 2  (k0 + k) 0 – j 2  (k0 + k) N – 1 + 1
   = x(0) e0 + x(1) e 2
+ x(2) e–j
1 e  – e 
N N
+   – j 3
2 – j 2  (k0 + k) 2
+ x(3) e
 1–e
N
  
= 1 · 1 + 2 cos – j sin 
j 2  (k0 – k) – j 2  ( k0 + k )  2
1 1 – e  1 1 – e  2
= j2
 +  –j2
 + 3 [cos  – j sin ]
2  2
(k0 – k) ( k0 + k ) 
 1–e N   1–e N 
+ 4 cos – j sin 
3 3
 2 2
Ex. 3.4.7 : Compute the DFT of x(n) = {1, 2, 3, 4}. Draw
the corresponding magnitude and phase spectrum. = 1 + 2 (0 – j) + 3 (– 1 – 0) + 4 (0 + j)

ge
Soln.: x(n) = {1, 2, 3, 4} = 1 – 2j – 3 + 4j
io led X(1) = – 2 + 2j
(iii) k = 2
3 – j 22 · n
X(2) =  x(n) e 4

n=0
ic ow

– j 22 · 0 – j 22 · 1
4 4
= x(0) e + x(1) e
n
– j 22 · 2 – j 22 · 3
bl kn

4 4
+ x(2) e + x(3) e
Fig. P.3.4.7 = x(0) e0 + x(1) e–j + x(2) e–j2+ x(3)e–j3
at

= 1 · 1 + 2 (cos – j sin )
Pu ch

DFT is given by the equation,

+ 3 (cos 2 – j sin 2)
N–1 – j 2k · n

 4
; k = 0, 1, , N – 1 + 4 (cos 3 – j sin 3)
Te

X(k) = x(n) e
n=0 = 1 + 2 (– 1 – 0) + 3 (1 – 0) + 4 (– 1 – 0)
Since the length of x(n) is 4, N = 4 = 1–2+3–4
3 – j 2k n X(2) = – 2
X(k) =  x(n) e 4
; k = 0, 1, 2, 3, (iv) k = 3
n=0
3 – j 23 · n
For each value of k, we vary n from 0 to 3. X(3) =  x(n) e 4

n=0
(i) k = 0
– j 23 · 0 – j 23 · 1
3 – j 20 · n 4 4
 x(n) e
= x(0) e + x(1) e
X(0) = 4
– j 23 · 2 – j 23 · 3
n=0 4 4
+ x(2) e + x(3) e
– j 20 · 0 – j 20 · 1
3 9
4 4 –j –j
= x(0) e + x(1) e = x(0) e0 + x(1) e 2 + x(2) e–j3 + x(3) e 2
– j 20 · 2 – j 20 · 3
= 1 · 1 + 2 cos – j sin 
4 4
3 3
+ x(2) e + x(3) e  2 2
= e0 + 2 · e0 + 3 · e0 + 4 · e0
+ 3 (cos 3 – j sin 3) + 4 cos – j sin 
9 9
= 1+2+3+4  2 2
= 1 + 2 (0 + j) + 3 (– 1 – 0) + 4 (0 – j)
X(0) = 10
= 1 + 2j – 3 – 4j
(ii) k = 1
  X(3) = – 2 – 2j
3 – j 21 · n
X(1) =  x(n) e 4 Hence the final DFT of x(n) is,
n=0 X(k) = { 10, – 2 + 2j, – 2, – 2 – 2j }
 DSIP (MU-Sem.7-COMP) 3-10 Discrete Fourier Transform (DFT)
As is evident, the DFT is complex. (As in most practical cases) and hence can be split into magnitude and phase we
know,
Magnitude = (Real)2 + (Imaginary)2 and

Phase = tan–1 
Imaginary
 Real 
  X(k)  = { 10, 2.83, 2, 2.83}
 X(k) = { 0, 135 , 0, – 135}
We now draw the magnitude and phase. They are known as magnitude spectrum and phase spectrum.

ge
io led
ic ow
n
bl kn
at
Pu ch

Fig. 3.4.7(a)
Te

N N For k = 0
Note : We have drawn a dotted line at 2 (since N = 4, 2 = 2)
4 4
which we will explain later.
X (0) =  x (n) e
0
=  x (n)
Ex. 3.4.8 : Find N = 5 point DFT for x(n) = { 1,0,1,0,1 }. n=0 n=0
Soln. :  X (0) = x (0) + x (1) + x (2) + x (3) + x (4)
= 1+0+1+0+1
Given : x(n) = {1, 0, 1, 0, 1} and N = 5
 X (0) = 3
This is a five point DFT. For k = 1

According to the definition of DFT, 4 – j2n

N–1
X (1) =  x (n) e 5
– j2kn n=0
X (k) =  x (n) e N , k = 0 to N – 1
– j2
n=0
 X (1) = x (0) e + x (1) e 5
0

4 – j2kn – j4 – j6 – j8

 X (k) =  x (n) e 5 , k = 0 to 4 + x (2) e 5 + x (3) e 5 + x (4) e 5
n=0
 DSIP (MU-Sem.7-COMP) 3-11 Discrete Fourier Transform (DFT)

 X (1) = 1 + 0 + [– 0.81 – j0.59] + 0 + [0.31 + j0.95] Soln. :

According to the definition of DFT, We have
 X (1) = 0.5 + j0.36
N–1 – j2kn
For k = 2 X (k) =  x (n) e N
4 n=0
– j4n
X (2) =  x (n) e 5 L–1 – j2kn
n=0
– j4 – j8
 X (k) =  Ae N
n=0
 X (2) = x (0) e + x
0
(1) e 5 + x (2) e 5
n
– j12 – j16 L–1
e 5 e 5  – j2k 

+ x (3) + x (4)
= A e N 
 X (2) = 1 + 0 + [0.31 + j0.95] n=0
+ 0 + [– 0.81 + j0.59]

ge
We have standard summation formula,
 X (2) = 0.5 + j1.54
N2 N1 N +1
a –a 2
For k = 3
 ak =
io led 1–a
4 – j6n k = N1
X (3) =  x (n) e 5 – j2k
ic ow

n=0
Here N1 = 0, N2 = L – 1 and a = e N
– j6
 X (3) = x (0) e + x (1) e 5
0

 e 
– j2k0  – j2kL – 1 + 1
n
– j12 – j18 – j24
 N  – e N  
bl kn

+ x (2) e 5 + x (3) e 5 + x (4) e 5  X (k) = A

 – j2k

 X (3) = 1 + 0 + [0.31 – j0.95] + 0 + [– 0.81 – j0.59]
 
at

1–e N
Pu ch

 X (3) = 0.5 – j1.54

1–e N 
– j2kL
For k = 4
 
Te

 X (k) = A
4
 – j2k 
 1–e N 
– j8n
X (4) =  x (n) e 5
n=0
– j8 – j16
Ex. 3.4.10 : Derive DFT of the sample data sequence
 X (4) = x (0) e + x (1) e 5 + x (2) e 5
0

– j24 – j32
x(n) = {1, 1, 2, 2, 3, 3} and compute magnitude and
+ x (3) e 5 + x (4) e 5 phase spectrum.

 X (4) = 1 + 0 + [– 0.81 + j0.59]

+ 0 + [0.31 – j0.59] Soln. : According to the DFT equation,
N–1 2kn
 X (4) = 0.5 – j0.36 –j

 The 5 point DFT of the given sequence is

X (k) =  x (n)  e
N
k = 0, 1, 2, …N–1
n=0
X (k) = {3, 0.5 + j0.36, 0.5 + j1.54, 0.5 – Hence N = 6,
j1.54, 0.5 – j 0.36} 5 2kn
–j

Ex. 3.4.9 : Find the DFT of the following finite duration

X (k) =  x (n)  e
6
; k = 0, 1, 2, 3, 4, 5
n=0
sequence of length L.
 A For 0  n  L – 1
x (n) = 
 0 Otherwise
 DSIP (MU-Sem.7-COMP) 3-12 Discrete Fourier Transform (DFT)
2k 4k 6k 8k
–j –j –j –j
X (k) = x (0)  e –0
+ x (1)  e 6 + x (2)  e  6 + x (3)  e 6 + x (4)  e 6

10  k
–j
+ x (5)  e 6

For k = 0
X (0) = x (0)  e0 + x (1)  e0 + x (2)  e0 + x (3)  e0 + x (4)  e0 + x (5)  e0
  X (0) = 1 + 1 + 2 + 2 + 3 + 3 = 12
If k = 1
2 4 6 8 10 
–j –j –j –j –j
X (1) = x (0) + x (1)  e 6 + x (2)  e 6 + x (3)  e 6 + x (4)  e 6 + x (5)  e 6

 
X (1) = 1 + 1  cos – j sin  + 2  cos – j sin  + 2  (cos  – j sin )
2 2
 3 3  3 3

ge
+ 3 cos – j sin  + 3 cos – j sin 
4 4 5 5
 3 3  3 3
 X (1) = – 1.5 + j 2.6
io led
For k = 2
4 8 12 – 16 – 20 
–j –j –j
6 + x (2)  e 6 + x (3)  e 6 + x (4)  e 6 + x (5)  e 6
ic ow

X (2) = x (0) + x (1)  e

X (2) = 1 + 1 cos – j sin  + 2 cos – j sin  + 2 (cos 2  – j sin 2 ) + 3 cos   – j sin 

2 2 4 4 8 8
 3  3  3 3
n
3 3
bl kn

+ 3 cos
10 10
– j sin
 3 3 
at
Pu ch

 X (2) = – 1.5 + j 0.86

6 12  18  24  30 
–j –j –j –j –j
X (3) = x (0) + x (1)  e 6 + x (2)  e 6 + x (3)  e 6 + x (4)  e 6 + x (5)  e 6
Te

For k = 3
X (3) = 1 + 1 (cos  – j sin ) + 2 (cos 2 – j sin 2) + 2 (cos 3 – j sin 3 ) + 3 (cos 4 – j sin 4)
+ 3(sin 5 – j sin 5 )
 X (3) = 0
8 16  24  32  40 
–j –j –j –j –j
For k = 4 X (4) = x (0) + x (1)  e 6 + x (2)  e 6 + x (3)  e 6 + x (4)  e 6 + x (5)  e 6

8 8
X (4) = 1 + 1 cos   – j sin  + 2 cos
4 4
– j sin + 2 (cos 4  – j sin 4 )
   3 3   3 3 
16  16   20  20  
+ 3 cos – j sin + 3 cos – j sin
 3 3   3 3 
 X (4) = – 0.5 – j 0.86
10  20  30  40  50 
–j –j –j –j –j
For k = 5 X (5) = x (0) + x (1)  e 6 + x (2)  e 6 + x (3)  e 6 + x (4)  e 6 + x (5)  e 6

X (5) = 1 + 1 cos
5 5  10  10 
  3  – j sin 3  + 2 cos 3 – j sin 3  + 2 (cos 5  – j sin 5 )
20  20  25  25 
+ 3 cos – j sin + 3 cos – j sin
 3 3   3 3 
X (5) = – 1.5 – j 2.6
 DSIP (MU-Sem.7-COMP) 3-13 Discrete Fourier Transform (DFT)

 X(k) = {12, – 1.5 + j 2.6, – 1.5 j 0.86, 0, – 1.5 – j 0.86, – 1.5 – j 2.6}

Magnitude  | X(k) | = {12, 3, 1.73, 0, 1.73, 3}

Phase  X(k) = {0, 120, 150, 0, – 150, – 120}

We plot the magnitude and phase diagram as shown in Fig. P.3.4.10.

ge
io led
ic ow
n
bl kn
at

(a) Magnitude spectrum (b) Phase spectrum

Pu ch

Fig. P. 3.4.10
Te

Note : Observing the magnitude and phase plot of Ex. 3.4.7 Solved Examples
and Ex. 3.4.10 we note that
Ex. 3.5.1 : Given x(n) = {8, 2, 1, 5, 6, 2, 4, 3}
N
1. The magnitude plot is symmetric about the 2 point. If N is The DFT of this signal is
N+1 X(k) = { 31, 0.58 + 1.58 j, 9 + 4j, 3.41 – 4.4 j, 7, 3.41 + 4.4 j,
odd, then it is symmetric about the 2 point.
9 – 4 j, 0.58 – 1.58j }
N
2. The phase plot is anti-symmetric about the 2 point. If N is Verify this by a program or by using the fft(x) command in
N+1 MATLAB.
odd, then it is anti-symmetric about the 2 point.
Soln. : Since X(k) is complex, it has magnitude and phase.
This is always true. We plot its Magnitude.
|X(k)| = {31, 1.69, 9.84, 5.58, 7, 5.58, 9.84, 1.69}
3.5 The Fourier Spectrum

At this stage it is extremely important to understand

what the spectrum means. We stated that the Fourier
Transform gives us the frequency components present in a
signal. How do we get these.
We know how to solve a DFT example.
Let us take a 8 point example :
Fig. P.3.5.1
 DSIP (MU-Sem.7-COMP) 3-14 Discrete Fourier Transform (DFT)
The X-axis is the frequency axis. The following Since we take a 8-point DFT, N = 8
questions come to our mind  We get 0 Hz, 1250 Hz, 2500 Hz, 3750 Hz, 5000 Hz,
(1) What does 0, 1, 2, ...., 7 represent ? 6250 Hz, 7500 Hz and 8750 Hz.
(2) Is the X axis in Hz ? Hence the spectrum would be
(3) Does it mean that the original signal x(n) has
frequencies 0 Hz, 1 Hz,...., 6 Hz, 7 Hz ?
 These are very important questions that need to be
answered.
 Make sure you understand the next few lines.
 As we have seen in earlier chapters, the entire subject
of Digital Signal Processing carries no meaning if the
Sampling Frequency (Fs) is not known.

ge
 The DFT is basically sampling of the frequency axis.
 Suppose x(n) was obtained by sampling x(t) with a
Fig. P. 3.5.1(b)
io led
sampling frequency of 8000 Hz.
 Hence we state what has been stated earlier.
 A 8 point DFT will split this sampling frequency into 8
 The entire area of Digital Signal Processing makes
ic ow

equal parts.
sense only if we know the sampling frequency. Only if
Fs
k k = 0, 1, 2, …, 7 we know the sampling frequency, will we be able to
n
N
bl kn

find out the frequency components present in the

 Since it is a 8 point DFT, N = 8.
signal.
 We get 0 Hz, 1 kHz, 2 kHz, 3 kHz,….,7 kHz
at
Pu ch

 Hence the X-axis for the given example is,

3.6 Inverse Discrete Fourier Transform
(IDFT)
Te

We shall now learn how to compute the IDFT. It is

similar to the DFT except that the negative exponential is
replaced by a positive one. The IDFT is given by the
equation.
+ j 2kn
N–1 N
1
x(n) =
N  X(k) e ; n = 0, 1, 2,  N – 1
k=0

Fig. P.3.5.1(a)
Solved Example
 Hence we state that x(n) is made up of a DC value of
31, 1 Hz signal with magnitude 1.69, 2 Hz signal with Ex. 3.6.1 : Compute the IDFT of X(k) = {4, 0, 0, 0}
magnitude 9.84, 7 Hz signal with magnitude 1.69. Soln.: IDFT is given by the equation,
 If we change the sampling frequency, the X-axis N-1 + j 2kn
1
changes !! x(n) =
N  X(k) e
N
; n = 0, 1,  N – 1
k=0
 Suppose x(n) was obtained by sampling x(t) with a
sampling frequency of 10 kHz. Since length of X(k) is 4, N = 4
3 + j 2kn
 1
A 8-point DFT will split up the X-axis by the formula,
 x(n) =
4  X(k) e
4
; n = 0, 1, 2, 3
Fs k=0
k ; k = 0,1, 2, …, 7
N For each value of n, we vary k from 0 to 3
 DSIP (MU-Sem.7-COMP) 3-15 Discrete Fourier Transform (DFT)

1
+ j 20 · 3 + j 2 · 3
(i) n = 0
= X (0) e
4 4
+ X (1) e
3 + j 2k · 0 4
1
x(0) =
4  X(k) e
4


+ j 22 · 3 + j 23 · 3
k=0
+ X (2) e
4
+ X (3) e
4

1
+ j 20 · 0 + j 2 · 0

=
4
X (0) e 4
+ X (1) e
4
3
1 j
= [ 4e0 + 0 · e 2 + 0 ej3 +  ej9 ] = 1
4

+ j 22 · 0 + j 23 · 0

+ X (2) e
4
+ X (3) e
4

1  x(3) = 1
= [ X(0) e0 + X(1) e0 + X(2) e0 + X(3) e0] Hence, we have,
4
1 x(n) = [ x(0), x(1), x(2), x(3) }
= [4 + 0 + 0 + 0]
4  x(n) = {1, 1, 1, 1}

= 1 If we compare this example with Ex. 3.4.3, we can

ge
observe that,
 x(0) = 1 IDFT
X(k) = {4, 0, 0, 0}  x(n) = {1, 1, 1, 1}
(ii) n = 1
io led
1
3 + j 2k · 1 3.7 Computing DFT by Matrix Method
x(1) =
4  X(k) e
4
ic ow

k=0
We will now learn how to compute the DFT of a
1
+ j 20 · 1 + j 2 · 1
sequence by using the matrix method. This is a much easier
= X (0) e
4 4
n
+ X (1) e
4
bl kn

way as compared to the earlier technique which involved


+ j 22 · 1 + j 23 · 1
opening up the summation sign for each value of X(k).
+ X (2) e
4
+ X (3) e
4

at
Pu ch

2 j3 The DFT is given by the equation,

=
1
4
[4e + 0 · e
0
j
4
+0e + j
+ 0e 2 ] N–1 – j 2kn

 N
; k = 0, 1, 2, , N – 1
Te

X(k) = x(n) e
 x(1) = 1 n=0

"We define a new term WN which is called the twiddle

(iii) n = 2
3 + j 2k · 2 factor."
1
x(2) =
4  X(k) e
N
– j 2
k=0 N
WN = e
1
+ j 20 · 2 + j 2 · 2

=
4
X (0) e 4
+ X (1) e
4
Substituting this is the DFT equation we get,
N–1
  x(n) WN
+ j 22 · 2 + j 23 · 2 kn
X(k) = ; k = 0, 1, 2,  N – 1 (3.7.1)
+ X (2) e
4
+ X (3) e
4
 n=0

1
= [ 4e0 + 0 · ej + 0 ej2 + 0 ej3 ] = 1 Similarly the IDFT which is given by the equation,
4
N–1 + j 2kn
1
 x(2) = 1 x(n) =
N  X(k) e
N
; n = 0, 1,  N–1
k=0
(iv) n = 3
can be written in terms of WN as,
3 + j 2k · 3
1
 4 N–1
1
x(3) = X(k) e

–kn
4 x(n) = X(k) WN (3.7.2)
k=0 N
k=0
 DSIP (MU-Sem.7-COMP) 3-16 Discrete Fourier Transform (DFT)

Equation (3.7.1) can be represented in the matrix form as

DFT  X(k) = [WN] x(n) (3.7.3)

 x(1)   X(1) 
x(0) X(0)

where, x(n) =
 x(2):  ; X(k) =
 X(2) 
 x(N–1)   X(N–1) 
:
–j 2
N 1 N 1 N
We know WN = e
and
We calculate each value of the matrix,
2
0 –j
4
W4 = e = 1

ge
2
1 –j ·1
4
io led W4 = e = –j
2
2 –j ·2
4
W4 = e = –1
2
3 –j ·3
4
W4 = e = +j
ic ow

2
4 –j ·4
4
W4 = e = +1
n
bl kn

2
6 –j ·6
4
W4 = e = –1
2
at
Pu ch

9 –j ·9
4
W4 = e = –j

1 
(3.7.4) 1 1 1 1

 
Te

nk nk –j –1 j
Each value of WN
matrix is obtained by multiplying n  W4 = (1)
1 –1 +1 –1 
1 
and k. WN is called a N N DFT matrix.
j –1 –j
Note : Changing n and k co-ordinates will not make any
Hence a 4 4 DFT matrix is shown in Equation (1).
difference.
As mentioned earlier, WN is called the Twiddle
Similarly, Equation (3.7.2) can be written as, factor.
1 
IDFTx(n) = [WN] X (k) (3.7.5) The word twiddle means "Turn in a twisting or
N
spinning motion”. In other words, to move in a turning
Before we start solving examples of DFT using the
matrix method, let us try to generate WN matrix. (circular) fashion. A lot of us have a habit of twiddling our
tumbs while waiting for an oral exam.
3.7.1 Solved Examples of DFT Now WN is called a twiddle factor because it is cyclic
Ex. 3.7.1 : Generate a 4 4 DFT matrix. in nature.
Soln.: Since we require a 4 4 matrix N = 4. Shown below are W4 values mapped on a circle.

Therefore both, n and k will vary from 0 to 3.

 DSIP (MU-Sem.7-COMP) 3-17 Discrete Fourier Transform (DFT)

0 4 8
  W4 = W4 = W4 = 1
1 5 9
W4 = W4 = W4 = – j
2 6
W4 = W4 = – 1
3 7
W4 = W4 = + j

Fig. P. 3.7.1(a)

Similarly for W8, we can draw the values of the 8 8 matrix using a circle.

ge
io led
ic ow
n
bl kn
at
Pu ch

Fig. P. 3.7.1(b)
Te

2
Let us calculate each of these values and verify 9 –j ·9 9 9
 W8 = e 8
= cos – j sin
0 –j
2
·0 4 4
 W8 = e 8
= cos 0 – j sin 0 = 1
= 0.707 – 0.707 j
2
1 –j ·1   2
 W8 = e 8
= cos – j sin = 0.707 – 0.707 j  W8 = e
10 –j
8
· 10
= cos
5
– j sin
5
=–j
4 4 2 2
2
2 –j ·2   2
 W8 = e 8
= cos – j sin = – j  W8 = e
11 –j
8
· 11
= cos
11
– j sin
11
2 2 4 4
2
3 –j ·3 3 3 = – 0.707 + 0.707 j
 W8 = e 8
= cos – j sin = – 0.707 – 0.707 j
4 4 2
12 j · 12
4 –j
2
·4  W8 = e 8
= cos 3 – j sin 3 = – 1
 W8 = e 8
= cos  – j sin  = –1 2
13 j · 13 13 13
–j
2
·5 5 5  W8 = e 8
= cos – j sin
5 4 4
 W8 = e 8
= cos – j sin = – 0.707 + 0.707 j
4 4
= – 0.707 + 0.707 j
2
6 –j ·6 3 3 2
 W8 = e 8
= cos – j sin = +j 14 j · 14 7 7
2 2  W8 = e 8
= cos – j sin = +j
2 2
2
7 –j ·7 7 7 2
 W8 = e 8
= cos – j sin = 0.707 + 0.707 j 15 j · 15 15 15
4 4  W8 = e 8
= cos – j sin
4 4
2
8 –j ·8
 W8 = e 8
= cos 2 – j sin 2 = +1 = 0.707 + 0.707 j
2
16 j · 16
 W8 = e 8
= cos 4 – j sin 4 = 1
 DSIP (MU-Sem.7-COMP) 3-18 Discrete Fourier Transform (DFT)

Using these values we can easily generalize a 8 8 Since W8 is cyclic we have,
DFT matrix. 
0 8 16 24
W8 = W8 = W8 = W8 = W8 = W8 = 1
32 40

1 9 17 25 33 41 49
 W8 = W8 = W8 = W8 = W8 = W8 = W8
= 0.707 – 0.707 j
2 10 18 26 34 42
 W8 = W8 = W8 = W8 = W8 = W8 = j
3 11 19 27 35 43
 W8 = W8 = W8 = W8 = W8 = W8
= – 0.707 0.707 j
4 12 20 28 36 44
 W8 = W8 = W8 = W8 = W8 = W8 = 1
5 13 21 29 37 45
 W8 = W8 = W8 = W8 = W8 = W8
= – 0.707 + 0.707 j
6 14 22 30 38 46
 W8 = W8 = W8 = W8 = W8 = W8 = j

ge
7 15 23 31 39 47
and W8 = W8 = W8 = W8 = W8 = W8
= 0.707 + 0.707 j
io led
Using these values we generate a 8 8 DFT matrix.
ic ow
n
bl kn
at
Pu ch
Te

Let us now solve a few examples by using the DFT matrix method.
Ex. 3.7.2 : Compute the DFT of the sequence x(n) = { 1, 1, 1, 1}
Soln.: This is the same sequence which was solved in Ex. 3.4.3.

The DFT is given by the equation,

N–1


kn
X(k) = x(n) WN ,
n=0
– j 2kn
nk 
Where, WN = e
Since the length of x(n) = 4, the DFT equation reduces to,
3


kn
X(k) = x(n) W4 ; k = 0, 1, 2, 3
n=0

In matrix form it is written as,

X(k) = [W4] x(n)
 DSIP (MU-Sem.7-COMP) 3-19 Discrete Fourier Transform (DFT)

1 1 
1 1 1 1 1

Where, x(n) =
   
1
Step 1 :
1 1 
1 
W4 x (n)
1 
1 1 1 1
 1 –j 1
1 1 1 1 1
 –j 
 X(k) =
 –1 j   Step 2 :
1 
 1 –1 1 –1 1 1 –j 
1 j –1 –j 1
1 
1 1 1 1
0
4
 –j +j 
 X(k) =
  Step 3 :
1 
0 1 –j 
0 +j

ge
1 
1 1 1 1
Hence, X(k) = [ 4‚ 0‚ 0‚ 0 ]
Step 4 :
 –j j 
We note the answer is the same as that obtained in
1 
io led +1
Example 3.4.3. 1 j –j 
Ex. 3.7.3 : Compute the DFT of x(n) = {1, 2, 3, 4}
1 
1 1 1 1
ic ow

Soln. : This is the same as Example 3.5.7.

Step 5 :
 –j –1 +j 
Since x(n) is of length 4, N = 4 and we generate a DFT 1 –1 +1 –1 
n
matrix of size 4 4. 1 –1 –j 
bl kn

+j
  X(k) = [W4]4 4 x(n)
Ex. 3.7.4 : Compute the IDFT of
at

 1 –j –1  2 
1 1 1 1 1
Pu ch

X(k) = {10, –2 + 2j, –2, –2 – 2j}.

 X(k) =
 j   Soln. : The IDFT is given by the equation
 1 –1 1 –1  3  N–1
Te

 1 j –1 –j  4  x(n) =
1
N  X(k) WN
– nk
; n = 0, 1, 2,  N – 1
k=0

 –2 +2j 
10

 
In matrix form it is written as,
X(k) = 1
 
*
–2 x(n) = [WN] X(k)
N
 – 2 – 2j  Since X(k) is of length 4, N = 4 and we generate a IDFT
 X(k) = { 10‚ –2 +2j‚ –2‚ – 2 – 2j } matrix of size 4 4.
*
This is the same as obtained in Example 3.5.7. W4 X (k)

1 j   –2 + 2j 
You must have realized that the matrix method is much 1 1 1 1 10
easier that the conventional method of opening the
 x(n) =
1  –1 –j  
 1 –1 –1   –2 
summation. 4 +1
Note : The DFT matrix depends only on the length of x(n) and
not the values of x(n). Because of this for every 4 point
 1 –j –1 j   –2 – 2j 

 
input, we use the same 4 4 DFT matrix. 4

A 4 4 DFT matrix can be easily memorized using the 1 8 

x(n) =
 
4 12
 16 
5 steps.
 DSIP (MU-Sem.7-COMP) 3-20 Discrete Fourier Transform (DFT)

2
1 Soln. : The IDFT is given by the equation
  1
N–1
x(n) =

– nk
3 x(n) =
N
X(k) WN ; n = 0, 1, 2,  N – 1
4 k=0

In matrix form it is written as,

 x(n) = { 1 2 3 4 }
1 *
x(n) = [WN] X(k)
Ex. 3.7.5 : Compute the DFT of x(n) = {1, 1, 0, 0} N

MU - Dec. 2018, 10 Marks Since X(k) is of length 4, N = 4 and we generate a IDFT

matrix of size 4 4.
Soln. :
*
W4 X (k)
Since x(n) is of length 4, N = 4 and we generate a DFT
1 j  
1 1 1 1 2
matrix of size 4 4.
 X(k) = [W4]4 4 x(n)
 x(n) =
1  –1 –j  1–j 
 1 –j –1
1 1 1 1
 1 
1 4
 1 –1 +1 –1   0 

    1 –j j   1+j 

ge
j –1
 X(k) =
 1 –1 1  0   
4
–1
 1 j –1  0  1 4 
io led –j x(n) =
 1–j 
2  
4 0

  0
ic ow

X(k) =
 0  1
1
 1+j   
n
x(n) =
 X(k) = { 2, 1– j, 0, 1 + j} 0
bl kn

Ex. 3.7.6 : Compute the IDFT of 0

at

X(k) = {2, 1 – j, 0, 1 + j}.  x(n) = { 1 1 0 0 }

Pu ch

MU - Dec. 2016, 5 Marks

Te

Ex. 3.7.7 : For a continuous time signal equation, s(t) = sin [2  1000 t ] + 0.5 sin [ 2  2000 t ] Sample the given signal
at 8000 samples/sec and find out 8 point DFT. Plot magnitude and phase response.
Soln. :

Given :Sampling rate Fs = 8000 sample / sec. Discrete signal is obtained by replacing t by n/Fs

 x (n) = sin 2 

1000  2000     
 8000 n + 0.5 sin 2 8000 n = sin 4  n + 0.5 sin 2  n
For n = 0 to 7
 
x(0) = sin 0 + 0.5 sin 0 = 0 x(1) = sin + 0.5 sin = 1.207
4 2
 3 3
x(2) = sin + 0.5 sin  = 1 x(3) = sin + 0.5 sin = 0.207
2 4 2
5 5
x(4) = sin  + 0.5 sin 2  = 0 x(5) = sin + 0.5 sin = – 0.207
4 2
3 7 7
x(6) = sin + 0.5 sin 3  = – 1 x(7) = sin + 0.5 sin = – 1.207
2 4 2

 x(n) = {0, 1.207, 1, 0.207, 0, – 0.207 – 1, – 1.207}

Since the length of x(n) is equal to 8, we use a 8  8 DFT matrix,

 X (k) = [w8] x (n)

DSIP (MU-Sem.7-COMP) 3-21 Discrete Fourier Transform (DFT)

1 1 1 1 1 1 1 1 0
1 (0.707 – j –j (– 0.707 – j –1 (– 0.707 + j j (0.707 + j 0.707) 1.207
0.707) 0.707) 0.707)
1 –j –1 j 1 –j –1 j 1
X(k) = 1 (– 0.707 – j j (0.707 – j 0.707) –1 (0.707 + j 0.707) –j (– 0.707 – j 0.707) 0.207
0.707)
1 –1 1 –1 1 –1 1 –1 0
1 (0.707 + j –j (0.707 + 0.707) –1 (0.707 – j 0.707) j (– 0.707 – j 0.707) – 0.207
0.707)
1 j –1 –j 1 j –1 –j –1
1 (0.707 + j j (– 0.707 + j –1 (– 0.707 – j –j (0.707 – j 0.707) – 1.207
0.707) 0.707) 0.707)

 
0

ge
 
0 – j3.99


– 0.3 j

  j2
io led
 j0.0003 
–2j

 
0.006 j

 –j0.0003 
= =
0 0

 
ic ow

0.3 j

 2j
  j2 
n
– 0.006 j
 j3.99 
bl kn
at
Pu ch

X (k) = {0, – j3.99, – j2, j0.0003, 0,– j0.0003,

j 2, j 3.99}
Te

|X (k) | = {0, 3.99, 2, 0.0003, 0, 0.0003, 2, 3.99}

X (k) = {0, – 90, – 90, 0, – 90, 90, 90,}

Fig. P. 3.7.7(a)

Ex. 3.7.8 : Compute the DFT of the following frequency

n
x (n) = cos  0, 1, 2, 3
4

x (n) = cos 
n
Soln. :
 4  n = 0, 1, 2, 3
Fig. P. 3.7.7
We substitute values of n
0 
 x(10) = cos 
 4 =1
 DSIP (MU-Sem.7-COMP) 3-22 Discrete Fourier Transform (DFT)

 x(1) = cos  We generate a circular matrix of x2 (n)
4 = 0.707
  214 3  
y (0) 143 2 1
 x(2) = cos 
2
4=0

 y (1) 
=
  1 
 x(3) = cos   = – 0.707
3  y (2)   3 2 1 4  2 
4  y (3)  4321   2 
 x (n) = {1, 0.707, 0, –0, 707}
   
y (0) 15

A 4 – point DFT is matrix form is given by the equation  y (1) 

=
 17 
 y (2)   15 
X (k) = [W4] x (n)  y (3)   13 

 1 –j  
1 1 1 1 1
 y(n) = {15, 17, 15, 13}
y (k) =
 –1 j   0.707 
 1 –1 –1    Ex. 3.7.10 : Compute the 8-point DFT of the sequence

ge
1 0
1 j –1 –j   – 0.707  x(n) = {0, 1, 2, 3}. Also draw the magnitude and phase plot.
Soln. : Since we require a 8-point DFT, we append four
io led
 y (k) = {1, 1– j 1.414, 1, 1 + j 1.414} zeros to the original input sequence.
Ex . 3.7.9 : Compute circular convolution of x1 (n) = { 1,1,2,2}  x(n) = { 0, 1, 2, 3, 0, 0, 0, 0 }
ic ow

x2 (n) = {1, 2, 3, 4} As N = 8, we generate a 8 8 DFT matrix.

Soln. : X(k) = [W8] x(n)
n
bl kn

 x2 (n)
y (n) = x1 (n) O
at
Pu ch
Te

Solving the above matrix operation, we obtain

 –1.414 – j 4.83 
6

 –2 + j2 
 
 1.414 – j 0.83 
 1.414–2+ j 0.83 
X(k) =

 –2 – j2 
 – 1.414 + j 4.83 
 DSIP (MU-Sem.7-COMP) 3-23 Discrete Fourier Transform (DFT)

 X(k) ={6, –1.414 – j 4.83, –2 + j2, 1.414 – j 0.83, Ex. 3.7.11 : Determine 2-point and 4-point DFT of a
–2, 1.414 + j 0.83, –2 – j2, – 1.414 + j 4.83} sequence,
Since X(k) is complex, we plot the magnitude and phase. x(n) = u(n)–u(n–2)
|X(k)| = 2
(Real) + (Imaginary) 2 Sketch the magnitude of DFT in both the
 |X(k)| = {6, 5.03, 2.83, 1.64, 2, 1.64, 2.83, 5.03} cases.

X(k) = tan–1 
Imaginary Soln. :
 Real 
We begin with Identify the signal
  X(k) = {0, – 106.31, 135, –30.4, 0, 30.4,
x (n) = u(n) – u (n –2) is shown in Fig. P.3.7.11.
– 135, 106.31}
We now draw the magnitude and phase spectrum.

ge
io led
ic ow
n
bl kn

Fig. P.3.7.10
at
Pu ch
Te

Fig. P. 3.7.11

 x (n) = {1,1}

(i) Two point DFT

Since N = 2, we use a 2  2

DFT matrix

 X(k) = [W2] 2  2 x (n) … (1)

Fig. P.3.7.10(a) Generation of W2 is shown below n = 0 n=1

N
 W2 
A dotted line is drawn at = 4. k=0 0 0
2 W
2
Notice that the magnitude spectrum is symmetric about W
kn
2
=  0 
 W2 
1
N k=1 W
the point while the phase spectrum is anti-symmetric 2
2
N
about the point. This is always true.
2
 DSIP (MU-Sem.7-COMP) 3-24 Discrete Fourier Transform (DFT)
– j2
Now, W20 = e 2
·0
=1
 X (k) = {2, 1 – j, 0, 1 + j }

–j2
·1 – j
| X (k) | = (Real)2 + (Imaginary)2
W12 = e 2 = e =–1
 | X (k) | = {2, 1.414, 0, 1.414}
1 1 The magnitude plot is shown in Fig. P. 3.7.11(b).
 Wnk =  1 –1 
2
 

From Equation (1) we write W2 X(n)

1 11
X (k) =  1 –1   1 
 
Fig. P. 3.7.11(b)
 X (k) = 2
0
Ex. 3.7.12 : Compute the DFT of four point sequence

ge
x(n) = {0, 1, 2, 3}
 X (k) = {2,0}
Soln. :
| X (k) | =
io led
(Real)2 + (Img)2
The four point DFT in the matrix form is given by,
 | X(k) | = 22 = 2
X(k) = [ W4 ] · x ( n )
ic ow

 Magnitude plot is shown in Fig. P. 3.7.11(a).

  1 
1 1 1 1 0

 j  
n
1 –j –1
bl kn

=
 1 –1 1 – 1 2 
 1 j –1 – j 3 
at
Pu ch

 0 – j – 2 + 3j   – 2j + 2j 
0+1+2+3 6

 =  
Te

Fig. P.3.7.11(a)   X(k) =

 0–1+2–3   – 2 
(ii) Four point DFT  0 + j – 2 – 3j   – 2 – 2j 
Since N = 4, we use 4  4 DFT matrix   X(k) = {6, – 2 + 2j, – 2, – 2– 2j}

We append two zeros to x (n) to get x (n) = {1,1,0,0} Ex. 3.7.13 : Calculate 8 point DFT of
x ( n ) = { 1, 2, 1, 2 }
We have already learnt how to generate a 4  4 DFT
Soln. :
matrix
First we will make length of given sequence ‘8’ by
 X (k) = [W4] 4  4 x (n)
appending zeros to x(n).

 1
1 1 1 1 1

 X (k) =
 1 –j –1 j    x ( n ) = {1, 2, 1, 2, 0, 0, 0, 0}

 1 –1 1 – 1  0 
 1 j –1 –j  0 
Now

X (k) = [W8] x(n)

 1–j 
2

 X (k) =
 
 0 
 1+j 
 DSIP (MU-Sem.7-COMP) 3-25 Discrete Fourier Transform (DFT)

We have already learnt – how to compute W8, which is a 8  8 DFT matrix.

1 1 1 1 1 1 1 1 1

1 0.707 – j 0.707 –j – 0.707 – j 0.707 –1 – 0.707 + j 0.707 j 0.707 + j 0.707 2

1 –j –1 j 1 –j –1 j 1

1 – 0.707 – j 0.707 j 0.707 – j 0.707 –1 0.707 + j 0.707 –j – 0.707 + j 0.707 2

X (k)=
1 –1 1 –1 1 –1 1 –1 0

1 – 0.707 + j 0.707 –j 0.707 + j 0.707 –1 0.707 – j 0.707 j – 0.707– j 0.707 0

1 j –1 –j 1 j –1 –j 0

ge
1 0.707 + j 0.707 j – 0.707 + j 0.707 –1 – 0.707 – j 0.707 –j 0.707– j 0.707 0
io led
1+2+1+2+0+0+0+0
1 + 1.414 – j 1.414 – j – 1.414 – j 1.414 + 0 + 0 + 0 + 0
ic ow

1–j2–1+j2+0+0+0+0
 X(k) =
n
1 – 1.414 – j 1.414 + j + 1.414 – j 1.414 + 0 + 0 + 0 + 0
bl kn

1–2+1–2+0+0+0+0
1 – 1.414 + j 1.414 – j + 1.414 + j 1.414 + 0 + 0 + 0 + 0
at
Pu ch

1+j2–1–j2+0+0+0+0
Te

1 + 1.414 + j 1.414 + j – 1.414 + j 1.414 + 0 + 0 + 0 + 0

 1 – j3.828 
6

 0 
 
 1 – j 1.828  This is the required DFT.
 1 + –j 1.828 
X(k) =
2

 0 
 1 + j3.828 
 The 8 point DFT of the sequence is
 X (k) ={6,1 – j3.828, 0, 1 – j 1.828, – 2, 1 + j 8. 28, 0, 1 + j 3.828}

Ex. 3.7.14 : Compute 8 point DFT of the sequence x (n) = {0, 1, 2, 3}. Sketch the magnitude and phase plot also
Soln. : Given sequence is, x (n) = {0, 1, 2, 3}
It is asked to calculate 8 point DFT.
We append four zero to x (n) to make its length equal to 8,

 x (n) = {0, 1, 2, 3, 0, 0, 0, 0}
 DSIP (MU-Sem.7-COMP) 3-26 Discrete Fourier Transform (DFT)

Now X (k) = [ W8 ] x(n)

1 1 1 1 1 1 1 1 0

1 0.707 – j 0.707 –j – 0.707 – j 0.707 –1 – 0.707 + j 0.707 j 0.707 + j 0.707 1

1 –j –1 j 1 –j –1 j 2

1 – 0.707 – j 0.707 j 0.707 – j 0.707 –1 0.707 + j 0.707 –j – 0.707 + j 0.707 3

X(k) = 1 –1 1 –1 1 –1 1 –1 0

1 – 0.707 + j 0.707 –j 0.707 + j 0.707 –1 0.707 – j 0.707 j – 0.707– j 0.707 0

1 j –1 –j 1 j –1 –j 0

1 0.707 + j 0.707 j – 0.707 + j 0.707 –1 – 0.707 – j 0.707 –j 0.707– j 0.707 0

ge
 –1.414 – j 4.828 
6

–2+j2 
io led
 X (k) =
 1.414 – j 0.828 
 –1.414 
ic ow

–2–j2 
+ j 0.828
n
 – 1.414+ j 4.828 
bl kn
at
Pu ch

Magnitude response = | X (k) | = (Real)2 + (Imaginary)2

 | X (k) | = {6, 5.03, 2.83, 1.64, 2, 1.64, 2.83, 5.03}

Te

 Imaginary
Phase response = tan –1  
 Real 

 X(k) = {0, – 106.3, 135, – 30.36, 0, 30.36, – 135, 106.32}

The magnitude plot and phase plot are shown in Fig. P. 3.7.14.

Fig. P. 3.7.14
 DSIP (MU-Sem.7-COMP) 3-27 Discrete Fourier Transform (DFT)

ge
io led
ic ow

Fig. P. 3.7.14(a)
n
We draw the, magnitude and phase plot
bl kn

Ex. 3.7.15 : Plot the magnitude and phase spectrum of the

sampled data sequence = {2, 0, 0, 1} which was obtained
at

using a sampling frequency of 20 Hz.

Pu ch

Soln. : To obtain magnitude and phase response we will

Te

calculate DFT as follows,

X(k) =[W4] x(n)

     
1 1 1 1 2 3
Fig. P. 3.8.15

 X(k) =
 1 –j –1 j   0 
=
 2+j 
 1 –1 1 –1   0   1 
 1 j –1 –j   1   2–j 
 X(k)= { 3, 2 + j, 1, 2 – j }

Magnitude plot

| X (k ) | = Re2 + Im2

 | X (k) | = {3, 2.24, 1, 2.24}

 X (k) = tan
–1
 Img 
 Real 

 X (k) = {0, 26.56, 0, – 26.56}

Fig. P. 3.7.15(a)
 DSIP (MU-Sem.7-COMP) 3-28 Discrete Fourier Transform (DFT)

Ex. 3.7.16 : Compute the sequence x(n) for which the DFT is – 1.4142 – 4.8284i
{20, – 4 + 4j, – 4, – 4 – 4j} – 2.0000 + 2.0000i
Soln. : Since we need to find the sequence x(n), we 1.4142 – 0.8284i
compute the IDFT. – 2.0000 – 0.0000i
The IDFT is given by the equation 1.4142 + 0.8284i
1 *
 x(n) = [W4] X(k) – 2.0000 – 2.0000i
N
–1.4142 + 4.8284i
Since N = 4, we use a 4 4 IDFT matrix.

   – 4+ 4j 
1 1 1 1 20 Ex. 3.7.17 : Find the DFT of following sequence and check
1  1 j –1 –j   your result by using IDFT x (n) = {0, 1, 2, 3}
x(n) =
4
 1 –1 +1 –1   – 4  Soln. :
 1 –j –1 j   – 4 – 4j 
We have already obtained DFT of x (n) of Ex. 2.7.2
 
8

ge
1  16  It is X (k) = {6, 2j – 2, – 2, – 2j – 2}
x(n) =
 
4 24
 32 
io led
Calculation of IDFT :

x(n) = { 2, 4, 6, 8 } 1
x (n) =
N [ W] X
N N
ic ow

Given below is a MATLAB code to compute the DFT of a

  1  
sequence of any length. x(0) 1 1 1 1 6

 =  – j  
n
Program to Compute the DFT using Matrix Method x (1) 1 j –1 2j−2

bl kn

clc  x (2)  4 1 –1 1 – 1  −2 
clear all
 x (3)   1 –j –1 j  −2j−2 
at
Pu ch

  1   
x={0 1 2 3 0 0 0 0] ; 6 + 2j – 2 – 2 – 2j – 2 0 0
N=length (x); 1  6 – 2 – 2j + 2 – 2 + 2j =  4 = 1 
Te

=
4
 6 – 2j + 2 – 2 + 2j + 2  4 8   2 
for k = 0 : 1 : N–1  6 + 2 + 2j + 2 + 2 – 2j   12   3 
for n = 0 : 1 : N–1
 x (n) = {0, 1, 2, 3}
W (n+1, k+1) =exp (–j*2*pi*k*n/N);
end Ex. 3.7.18 : Show that the basis matrix for DFT is unitary.
end
Consider N = 4.
X=W*x'
X_Mag=abs (X); Soln. : A matrix is said to be unitary is

X_Ph=rad2deg (angle (X)); WN = W*T =I

N
subplot (2, 2, 1)
stem (X_Mag) We need to check is
title ('Magnitude plot') W4 · W*T = I
N
subplot (2, 2, 2)
stem (X_Ph) We have,
title ('Phase plot')
 
1 1 1 1

W4 =
 1 –j –1 j
X=
 1 –1 1 – 1
6.0000 + 0.0000i  1 j –1 – j
 DSIP (MU-Sem.7-COMP) 3-29 Discrete Fourier Transform (DFT)
3.8.1 Linearity
 
1 1 1 1


and W =
 1 j –1 – j
MU - Dec. 2015, Dec. 2017, May 2018)
4 1 –1 1 – 1
 1 –j –1 j Q. State the DFT Properties : Linearity.
(Dec. 2015, Dec. 2017, May 2018, 2 Marks)
Now

 
1 1 1 1 DFT
x1(n)  X1(k)
 j
If and
1 –j –1
W4  W4
T
= DFT
 1 –1 1 – 1 x2(n)  X2(k) then
 1 j –1 – j DFT
a x1(n) + bx2 (n)  a X1(k) + bX2(k)

 
1 1 1 1


 1 j –1 
–j Proof : According to the definition of DFT,

 –1

ge
N–1 – j2nk
1 –1 1
 1 –j –1 j X(k) =  x(n) e N

n=0
io led
 
1 0 0 0 Let x(n) = ax1(n) + bx2(n)

4 =
 0 1 0 0   4 [I]
N–1 – j2nk

 0 0 1 0  X(k) =  [ax1(n) + bx2(n)] e N

 0
ic ow

n=0
0 0 1
N–1 – j2nk N–1 – j2nk
n
Hence DFT matrix is unitary =  a x1(n) e N
+  bx2(n) e N
bl kn

n=0 n=0
3.8 Discrete Fourier Transform (DFT) N–1 – j2nk N–1 – j2nk
at

Properties  N
 N
Pu ch

= a x1 (n) e +b x2 (n) e
n=0 n=0
MU - May 2016, Dec. 2016, May 2017 = a X1 (k) + b X2 (k)
Te

Hence
Q. State any five DFT properties.
DFT
(May 2016, Dec. 2016, May 2017, 10 Marks) If x1(n)  X1 (k)
DFT
Like any other transform, the DFT has several important and x2(n)  X2(k), then
properties. These properties help us in computing the DFT DFT
easily. ax1(n) + bx2(n)  aX1(k) + bX2(k)
The properties that we study here are listed below. Solved Example
1. Linearity 2. Periodicity
Ex. 3.8.1 : x1 (n) = {1, 2, 3, 4} and x2 (n) = {5, 6, 7, 8}
3. Circular time shift 4. Circular frequency shift
Compute the DFT of the sequence
5. Time reversal 6. Symmetry property x3 (n) = 2x1 (n) + 3x2 (n)
7. Complex Conjugate property Soln. : From the Linearity property we know
8. Parseval's theorem DFT
If x1 (n)  X1 (k)
9. Multiplication
DFT
10. Circular Convolution property x2 (n)  X2 (k)
Though the Proof of each of the properties is not a DFT
part the syllabus, we shall prove them for the sake of then, a x1 (n) + bx2 (n)  aX1 (k) + bX2 (k)
continuity. Students can choose to ignore them. Now, x3 (n) = 2x1 (n) + 3x2 (n)
We will also solve examples using them.  DFT { x3 (n)} = DFT { a x1 (n)} + DFT {b x2 (n) }
 DSIP (MU-Sem.7-COMP) 3-30 Discrete Fourier Transform (DFT)
DFT
is x3 (n)  aX1 (k) + bX2 (k) Proof : According to the definition of the DFT
N–1 – j2nk
We calculate DFT of x1 (n) and x2 (n)
X(k) =  x(n) e N

 1 –j –1 2
1 1 1 1 1
n=0

X1 (k) =
 j   We replace k by (k + N)
 1 –1 1 –1 3 – j2n(k + N)
 1 j –1 4
N–1
–j  X(k + N) =  x(n) e N

 –2 + 2j 
10 n=0


  N–1 – j2nk – j2nN
X1 (k) =
 –2  =  x(n) e N
e N

 –2 –2j  n=0
N–1 – j2nk

 1 –j –1  6 
1 1 1 1 5
=  x(n) e N
 e – j2n
 j   n=0

ge
Similarly X2(k) =
 1 –1 1 –1  7  Now, e – j2n
= cos 2n – jsin 2n = 1
 1 j –1 io led –j  8  N–1 – j2nk

  N
 –2 +2j 
26 X(k + N) = x(n) e

 X2 (k) =
  n=0

 –2  = X(k)
 –2 –2j 
ic ow

 X(k + N) = X(k)
 DFT {x3 (n)} = X3 (k) = 2 X1 (k) + 3X2 (k) Hence the DFT is periodic with period N.
n
bl kn

 –2 + 2j   –2 +2j 
10 26 Circular shift of a sequence :

= 2
 +3  Before we move to the next properties, let us
at

 –2   –2 
Pu ch

 –2 –2j   –2 –2j  understand how to visualize a periodic signal.

Consider a sequence x(n),
 –4 + 4j   –6 +6j 
20 78
Te

=
 +  
 –4   –6 
 –4 –4j   –6 –6j 
 –10 + 10 j 
98

 X3 (k) =
 
 –10 
 –10 –10j 
 X3 (k) = {98, –10 + 10 j, –10, –10 –10j} Fig. 3.8.1(a)
3.8.2 Periodicity A periodic extension of a sequence x(n) can be written
MU - Dec. 2015, Dec. 2017, May 2018) as,
+

Q. State the DFT Properties : Periodicity. xp(n) =  x(n – lN); where N is the period.
l = –
(Dec. 2015, Dec. 2017, May 2018, 2 Marks)
This is shown in Fig. 3.8.1 (b).
A simple way of visualizing a periodic signal is to
DFT
If x(n) X(k) imagine x(n) on a circle in the counter clockwise direction.
then X(k + N) = X(k) The Fig. 3.8.1(b) will make this clear.
i.e., The DFT is periodic with a period N. xp (n) = X((n))N is the periodic extension of x(n).
 DSIP (MU-Sem.7-COMP) 3-31 Discrete Fourier Transform (DFT)

Fig. 3.8.1(e)
Fig. 3.8.1(b)
Here ((n))N denotes n modulo N.
As stated earlier, the periodic signal xp (n) can be
 x((n – 2))N denotes x(2 modulo N)
viewed as lying on a circle in the counter-clockwise
A general way of representing a periodic signal is,
direction.


ge
xp (n) = x [(n modulo N)] or x((n))N.
io led If n is between 0 and N – 1, then leave it as it is. If not,
then add or subtract multiplies of N from n until the result
is between 0 and N – 1.
Example : x(– 3 Modulo 4) = x((–3))4 = x(1)
ic ow

x(11 Modulo 8) = x((11))8 = x(2)

n
The shifted version of xp(n) shown in Fig. 3.8.1(d) can
bl kn

be written as,
Fig. 3.8.1(c)
+


at

x'p (n) = xp (n – k) = x(n – lN – k)

Pu ch

As we move around the circle in the counter-clockwise

l = –
direction we get,
x(0), x(1), x(2), x(3), x(0), x(1), x(2), x(3),  In general if
Te

 xp (n) = x [(n modulo N)] = xp ((n))N x(n) = {x(0), x(1), x(2), , x(N – 1)}

If we now shift the periodic signal by say 2. then x((n – 1)N = {x(N – 1), x(0), x(1), , x(N – 2)}

We get, x((n – 2))N = {x(N – 2), x(N – 1), x(0), x(1), , x(N – 3)}
: :
: :
: :
x((n – N))N = {x(0), x(1),  , x(N – 1)}
 x((n – N))N = x(n)

In a similar way,
x((n – m))N = x(N – m + n)

We now prove the circular shift property of the DFT.

Fig. 3.8.1(d)
3.8.3 Circular Time Shift
This is nothing but a circular shift of the circle as MU - Dec. 2015, Dec. 2017, May 2018)
shown in Fig. 3.8.1(e)
Q. State the DFT Properties : Time Shift.
(Dec. 2015, Dec. 2017, May 2018, 2 Marks)
 DSIP (MU-Sem.7-COMP) 3-32 Discrete Fourier Transform (DFT)
DFT – j2mk N – 1 – j2k
If x(n)  X(k)
– j2mk
= e N
 x(l) e N
DFT l=0
N
then x((n – m))N  ee  X(k) DFT {x((n – m))N} = e– j2km/N  X(k)
N–1 – j2nk
– j2mk
DFT {x(n)} = X(k) =  x(n) e N

DFT
x((n – m))N e N
 X(k)
n=0
– j2nk
m–1 Solved Examples
 DFT {x((n – m))N}=  x((n – m))N e N

n=0 Ex. 3.8.2 : (i) Given x(n) = {1, 2, 3, 4}

N–1 – j2nk
Find the DFT of x(n).
+  x((n – m))N e N
…(3.8.1)
(ii) Using results obtained in (i) and not otherwise, obtain
n=m
the DFT of the following sequences.
We see that the above equation is made up of two
x1(n) = {4, 1, 2, 3}
parts. We work on the first part.

ge
x2(n) = {3, 4, 1, 2}
Since, x((n – m))N = x(N – m + n), we have
m–1 – j2nk m–1 – j2nk Soln. :
io led
 x((n – m))N e N
=  x(N – m + n) e N (i) We first find the DFT of x(n) using the matrix notation
n=0 n=0 nk
X = WN  x
Let N – m + n = l
ic ow

    
X(0) 1 1 1 1 x(0)

   j   x(1) 
n = l–N+m
X(1) 1 –j –1
n
– j2nk N–1 – j2k(N + m + l) =
   X(2)   1 –1 1 – 1   x(2) 
bl kn

   – j   x(3) 
N N
 x((n – m))N e = x(l) e
l= N–m X(3) 1 j –1

     
X(0) 1 1 1 1 1
at
Pu ch

N–1 – j2k(l + m)

=  x(l) e N
 e+ j2k  X(1)  =
 1 –j –1 j   2 
l= N–m  X(2)   1 –1 1 –1   3 
   –j   4 
Te

X(3) 1 j –1
Since e+ j2k = 1 for k = 0, 1, 2, …. we have,
 X(k) = {10, – 2 + 2j, – 2, – 2 –2j}
– j2nk N–1 – j2k(l + m)
(ii) We now assume that x(n) is periodic, we plot x(n)
 x((n – m))N e N
=  x(l) e N

l= N–m
…(3.8.2)
Similarly for the second part of Equation (3.8.1) we
have,
N–1 – j2nk N–1–m – j2k(m + l)

 x((n – m))N e N
=  x(l) e N

n=m l=0

…(3.8.3)
Substituting Equations (3.8.2) and (3.8.3) in
Equation (3.8.1) we get,
N–1 – j2k(l + m)
Fig. P. 3.8.2
DFT {x((n – m))N} =  x(l) e N

l= N–m Now x1(n) is x(n) shifted to the right by one.

N–m–1 – j2k(m + l) Note, x1(n) can also be x(n) shifted to the left by 3.
+  x(l) e N
i.e. x1(n) = x (n – 1) OR x1(n) = x(n + 3)
l=0
 DSIP (MU-Sem.7-COMP) 3-33 Discrete Fourier Transform (DFT)
Let’s take x1(n) = x (n – 1)
(Both will give same answer)

Fig. P. 3.8.2(b)

Fig. P. 3.8.2(a) DFT pk

Since x(n – p)  WN  X(k)

ge
We know DFT 2k
DFT  x (n + 2)  W4 X (k)
If x(n)  X(k)
io led 2k
– j2pk
X2(k) = W4 X(k)
DFT N 0
x(n – p)  e  X(k) X2(0) = W4 X(k) = (1)  (10) = 10
ic ow

DFT pk 2
i.e., x(n – p)  WN X (k) X2(1) = W4 X (k) = (– 1) (– 2 + 2j) = 2 – 2j
4
n
DFT k X2(2) = W4 X (k) = (1) (– 2) = –2
 x(n – 1)  WN  X(k)
bl kn

6
DFT k X2(3) = W4 X (k) = (– 1) (– 2 – 2j) = 2 + 2j
 x1(n) = x(n – 1)  WN  X(k)

at

DFT {3, 4, 1, 2} = {10, 2 – j2, – 2, 2 + j2}

Pu ch

We already know X(k) = {10, – 2 + 2j, – 2, – 2 – 2j}  X2(k) = {10, 2 – 2j, –2, 2 + 2j}
k
i.e., X1(k) = W4 X (k)
Te

Ex. 3.8.3 :
0
X1(0) = W4 X (0) = (1) (10)
(i) x (n) = {1, 2, 3, 4} find DFT X (k)
= 10 (ii) Using results obtained in part (i) and not otherwise find
1
X1(1) = W4 X (1) = (– j) (– 2 + 2j) = 2 + j2 the DFT of following sequences
2 x1 (n) = {4, 1, 2, 3} x2 (n) = {2, 3, 4, 1}
X1(2) = W4 X (2) = (– 1) (– 2) = 2
3 x3 (n) = {3, 4, 1, 2} x4 (n) = {4, 6, 4, 6}
X1(3) = W4 X (3) = (j) (– 2 – 2j) = 2 – j2
Soln. :
 DFT {4, 1, 2, 3} = {10, 2 + j2, 2, 2 – j2}
 X (1)    
X (0) 1 1 1 1 1

   j  
X1(k) = {10, 2 + j2, 2, 2 – j2}
1 –j –1 2
(i) =
Similarly for x2(n), we observe that x2(n) is simply x(n)
 X (2)   1 –1 1 – 1  3 
shifted to the left by 2 or to the right by 2.  X (3)   1 j –1 – j  4
 x2(n) = x(n + 2) = x (n – 2)
 1 – 2j – 3 + 4j   
1+2+3+4 10
Let us take x2(n) = x (n + 2)
=
 = – 2 + 2j 
(Both will give the same answer)  1 – 2 + 3 – 4   –2 
 1 + 2j – 3 – 4j   – 2 – 2j 
 X (k) = {10, – 2 + 2j, – 2, – 2, – 2j}
 DSIP (MU-Sem.7-COMP) 3-34 Discrete Fourier Transform (DFT)
(ii) (a) Given x1 (n) = {4, 1, 2, 3} and – j2

x (n) = {1, 2, 3, 4}. Thus x1 (n) is obtained by For k = 1  X2 (1) = X (1) e 4

= (– 2 + 2j) cos + j sin 

delaying x (n) by one position 2 2
 4 4
 x1 (n) = x (n – 1)
  X2 (1) = (– 2 + 2j) (0 + j)
According to circular time shift property
  X2 (1) = – 2 – j2
– j2kl
DFT – j4
x (n – l) – X (k) e N
For k = 2  X2 (2) = X (2) e 4
Here l = 1
– j2k = (– 2) [cos  + j sin ]
 X2 (2) = (– 2) [– 1]
  X1 (k) = X (k) e N
 X2 (2) = 2
For k = 0  X1 (0) = X (0) e0 = 10
– j6
– j2
For k = 3  X2 (3) = X (3) e 4

ge
k=1 X1 (1) = X (1) e 4
= (– 2 – 2j) cos + j sin 
6 6
= (– 2 + 2j) cos – j sin 
2 2
 4 4
 4
io led 4
 X2 (3) = (– 2 – 2j) [– j]
 X1 (1) = (– 2 + 2j) (– j)
 X2 (3) = – 2 +2j
 X1 (1) = 2 + j2
ic ow

– j4  X2 (k) = {10, – 2, – j2, 2, – 2 + j2}

For k = 2  X1 (2) = X (2) e 4 = X (2) e– j (c) Given x3 (n) = {3, 4, 1, 2} and x (n) = {1, 2, 3, 4}.
n
That means x3 (n) is obtained by delaying x (n) by 2
  X1 (2) = (– 2) [cos  – j sin ]
bl kn

positions.
= (– 2) (– 1)
 x3 (n) = x (n – 2)
at

 
Pu ch

X1 (2) = 2
– j6 According to circular time shifting property,
4 – j2kl
For k = 3 
Te

X1 (3) = X (3) e DFT

x (n – l) – X (k) e N
= (– 2 – 2j) cos – j sin 
6 6
 4 4 DFT
– j4k

 X1 (3) = (– 2 – 2j) (+ j) x (n – 2) – X (k) e 4

 X1 (3) = 2 – j2   X3 (k) = X (k) e–jk

 X1 (k) = {10, 2 + j2, 2, 2 – j2} For k = 0 X3 (0) = X (0) e = 10

(b) Given x2 (n) = {2, 3, 4, 1} and x (n) = {1, 2, 3, 4}. For k = 1 X3 (1) = X (1) e–j  = (– 2 + 2j) [cos – j sin ]
Thus x2 (n) is obtained by advancing x (n) by   X3 (1) = (– 2 + 2j) (– 1)
 position   X3 (1) = 2 – 2j
 x2 (n) = x (n + 1) For k = 2  X3 (2) = X (2)e– j2
According to circular time shifting property, = (– 2) [ cos 2 – j sin 2  ]
DFT
– j2kl   X3 (2) = (– 2) (1)
x (n – l) – X (k) e N   X3 (2) = – 2
– j2k
DFT For k = 3  X3 (3) = X (3) e– j3
x (n + 1) – X (k) e N
– j2k
= (– 2 – j2) [ cos 3  – j sin 3  ]
  X3 (3) = (– 2 – j2) [– 1]
X2 (k) = X (k) e N
  X3 (3) = 2 + j2
For k = 0  X2 (0) = X (0) e = 10
  X3 (k) = {10, 2 – j2, – 2, 2 + j2}
 DSIP (MU-Sem.7-COMP) 3-35 Discrete Fourier Transform (DFT)

(d) Given x4 (n) = {4, 6, 4, 6} and x (n) = {1, 2, 3, 4}  X (k) = {3, 2 + j, 1, 2 – j}

(ii) The plot of x (n), xp (n) and xp (n – 3) are shown in
Thus x4 (n) and x (n) are related as,
Figs. P. 3.8.4(a), P. 3.8.4(b) and P. 3.8.4(c)
– respectively.
x4 (n) = x (n) + x (n + 2)

Using half period shift property,

–
X4 (k) = X (k) + (– 1)k X (k) …[∵ x (n + 2)  (– 1)k X (k)]

For k = 0 X4 (0) = X (0) + (– 1)0 X (0) = 10 + 10

X4 (0) = 20

For k = 1 X4 (1) = X (1) + (– 1)1 X (1) =– 2 + 2j + 2 – 2j

ge
  X4 (1) = 0 io led
For k = 2  X4 (2) = X (2) + (– 1)2 X (2) = – 2 + (– 2)

  X4 (2) = – 4
ic ow

For k = 3  X4 (3) = X (3) + (– 1)3 X (3) = – 2 – 2j + 2 + 2j

n
  X4 (3) = 0
bl kn

Fig. P. 3.8.4
  X4 (k) = {20, 0, – 4, 0}
at
Pu ch

(iii) In Fig. P. 3.8.4, consider the sequence in the basic

Ex. 3.8.4 : For a given sequence x (n) = {2, 0, 0, 1}, perform
range; that means n = 0 to 3.
following operations :
Te

(i) Find out the 4 point DFT of x (n)  x (n) = {0, 0, 1, 2}

(ii) Plot x (n), its periodic extension xp (n) and xp (n – 3)
(iii) Find out 4 point DFT of xp (n – 3) Comparing this to the original sequence,

(iv) Add phase angle in (i) with factor –  N 

2rk
  x (n) = {2, 0, 0, 1} we can write,

where N = 4, r = 3 k = 0, 1, 2, 3 x (n) = x ((n – 3))

(v) Comment to the result you had in point (i) and (ii)
According to circular time shifting property,
Soln. :
– j2kl
Given : x (n) = {2, 0, 0, 1}
x (n – l)N  X (k) e N
(i) X (k) = [W4] x4

  
1 1 1 1 2 Here l = 3, N = 4

=
 1 –j –1 j  0  and X (k) = {3, 2 + j, 1, 2 – j}
 1 –1 1 –1  0 
 1 j –1 –j  1  So, DFT of x (n) is,

   
2+0+0+1 3
k = 0  X (0) e– j0 = 3
=
 2+0+0+j  = 2+j 
 2+0+0–1   1  – j2 1 3 – j3

 2+0+0–j   2–j  k = 1  X (1) e 4 = (2 + j)  e 2

 DSIP (MU-Sem.7-COMP) 3-36 Discrete Fourier Transform (DFT)
DFT – j2 kl
= (2 + j) (0 + j) = – 1 + j2
 x ((n – l))5  X (k) e 5 …(1)
– j2 2 3 5
k = 2  X (2)  e 4 = 1  e– j3 = – 1 – j44
5
– j2 3 3 Given : Y(k) = e X (k) …(2)
k = 3  X (3) e 4 Comparing Equations (1) and (2) we note l = 2
– j9  y (n) = x (n – 2)5
= (2 – j)  e 2 = (2 – j) (0 – j) = – 1 – j2
This simply means y(n) is obtained by circularly delaying
 X (k) = {3, – 1 + j2, – 1, – 1 – j2}
x(n) by 2 position. We arrange the values of x(n) on a circle
(iv) We have, X (k) = {3, 2 + j, 1, 2 – j} in the counter otherwise direction.
– j2rk
The added phase angle is e N ; N = 4, r = 3 and

ge
k = 0, 1, 2, 3
io led – j23k – j3k
That means phase angle is e 4 = e 2
Due to this added phase angle DFT will be same as
DFTOF (ii). Fig. P. 3.8.5(a)
ic ow

(v) The DFTOF xp (n) is same as that of x (n); because

 y (n) = {1, 0, 1, 2, 2}
DFT is periodic in nature.
n
bl kn

Ex. 3.8.5 : Consider the finite length sequence x(n) shown in

Fig. P. 3.8.5. The five point DFT of x (n) is denoted by
at

X (k). Plot the sequence whole DFT is

Pu ch

4 k
–
5
Y (k) = e X(k)
Te

Fig. P. 3.8.5(b)

Ex. 3.8.6 :
1 0n3
x [n] = 
0 4n7
Fig. P. 3.8.5 (i) Find DFT X [ k ]
(ii) Using the result obtained in (i) find the DFT of the
Soln. : The given sequence contains 5 points. It can be
following sequences.
written as, n=0
1
x1 [n] =  0 1n4
x (n) = {1 , 2, 2, 1, 0 } 1 5n7

0 0n1
and x2 [n] =  1 2n5
According to circular time shifting property, 0 6n7
DFT Soln. :
kl
x ((n – l))N  X (k) WN
N  1 0  n  3
(i) Given, x(n) = 
 0 4  n  7
– j2
We have WN = e N Here N = 5  x(n) = {1, 1, 1, 1, 0, 0, 0, 0}
 DSIP (MU-Sem.7-COMP) 3-37 Discrete Fourier Transform (DFT)

The DFT of x(n) is calculated as follows.

 1 0.707 – j 0.707 – j – 0.707 – j 0.707 – 1 – 0.707 + j 0.707 j 0.707 + j 0.707  1

1 1 1 1 1 1 1 1 1

1 –j –1 j 1 –j –1 j  1
 X(k) =
 1 – 0.707 – j 0.707 j 0.707 – j 0.707 – 1 0.707 + j 0.707 – j – 0.707 + j 0.707  1
 11 – 0.707–+1j 0.707 –1j 0.707 –+ 1j 0.707 –11 0.707 –– 1j 0.707 1j – 0.707–– 1j 0.707   11 
1  1
 1 0.707 + j 0.707 j – 0.707 + j 0.707 – 1 – 0.707 – j 0.707 – j 0.707– j 0.707  1
j –1 –j 1 j –1 –j

 1 + 0.707 – j0.707 – j– 0.707 – j0.707 + 0 + 0 + 0 + 0 

1+1+1+1+0+0+0+0

 
ge
1–j–1+j+0+0+0+0

 X(k) =
 1 – 0.707 – j0.707 + j + 0.707 – j0.707 + 0 + 0 + 0 + 0 
io led
 1–1+1–1+0+0+0+0

 
1 – 0.707 + j0.707 – j + 0.707 + j0.707

 1 + 0.707 + j0.707 + j – 0.707 + j0.707 + 0 + 0 + 0 + 0 

ic ow

1+j–1–j+0+0+0+0
n
bl kn

 X(k) = {4, 1 – j2.414, 0, 1 – j0.414, 0, 1 + j0.414, 0, 1 + j2.414}

at

 20
Pu ch

 1 n = 0 – j10
8
2 –j
8 X(2) = 0
(ii) Given x1(n) =  0 1  n  4 X1(2) = e X(2) = e
 1 5  n  7 10 30
3
Te

–j –j
X1(3) = e 8 X(3) = e 8 X(3)
x1(n) = {1, 0, 0, 0, 0, 1, 1, 1}
Comparing with x(n), we can write = (0.707 + j0.707) (1 – j0.414) = 1 + j0.414
10
x1(n) = x ((n – 5)) –j 4
X1(4) = e 8 X(4) = 0
According to circular time shifting property,
10
DFT –j 5
kl X1(5) = e 8 X(5)
x1 ((n – 1))  X(k)  WN
N = (0.707 – j0.707) (1 + j0.414) = 1 – j0.414
DFT 10
 x1(n) = x((n – 5))  X(k)  WN
k.5 –j 6
X1(6) = e 8 X(6) = 0
N
10
– j2k –j 7
k X1(7) = e 8 X(7)
Now WN = e N
– j70
– j2
5k  5k = e 8 (1 + j2.414) = 1 – j 2.41
 WN = e 8
 X1(k) = {4, 2.41 – j, 0, 1 + j0.414, 0,
– j 10 k
= e = – 0.707 + j0.707
8 1 – j0.414, 0, 2.41 + j}
X1(0) = e  X(0) = 4
0
 0 0  n  1
10 Given x2(n) =  1 2  n  5
X1(1) = e
–j
8 X(1)  0 6  n  7
= (– 0.707 + j0.707) (1 – j2.414) = 1 + j 2.41
 DSIP (MU-Sem.7-COMP) 3-38 Discrete Fourier Transform (DFT)
– j2n(N + k – m)
 x2(n) = {0, 0, 1, 1, 1, 1, 0, 0} N–1

Thus x2(n) = x ((n + 2))

=  x(n) e N

n=0
Using circular time shifting property = X (N + k – m)
DFT
– kl = X ((k – m))N
 x2(n) = x((n + 2))  X(k)  WN
N DFT {x(n) e+ j2mn/N} = X ((k – m))N
– 2k
 X2(k) = X(k)  W8 This is the circular frequency shift property.
– j2.2k – jk
2k Solved Example
Now W8 = e 8 =e z
jk Ex. 3.8.7 : Given signal x(n) = {2, 2, 1, 4} has a 4-point DFT
–
 X2(k) = X(k)  e 2 X(k). Without performing DFT or IDFT, find out the sequence
x1(n) which would have a 4 point DFT X(k – 1).
 X2(0) = x(0)  e0 = 4

ge
j Soln. : This is what the problem states
–
X2(1) = x(1)  e 2 = (1 – j2.414) (0 + j1) DFT
x(n)  X(k)
io led
= – 2.414 – j then
j
X2(2) = X(2)  e = 0 DFT
??  X (k – 1)
ic ow

j3
X2(3) = X(3) e 2 = (1 – j0.414) (0 – j) = 0.414 + j The output in the frequency domain is shifted. Hence we
j2 use the circular frequency shift property.
n
X2(4) = X(4) e =0
DFT
bl kn

j5 i.e., if x(n)  X(k)

X2(5) = X(5) e2 = (1 + j0.414) (0 + j) = 0.414 – j j2mn
DFT
at

j3 N
 X(k – m)
Pu ch

X2(6) = X(6) e =0 then x(n) e

j7 In this case m = 1
X2(7) = X(7) e 2 = (1 + j2.414) (0 – j) = –2.414 – j
Te

j2n
N DFT
X2(k) = {4, –2.414 – j, 0, 0.414  x(n) e  X(k – 1)
j2n
+ j, 0, 0.414 – j, 0 – 2.414 + j}
 x1(n) = x(n) e 4
3.8.4 Circular Frequency Shift j20

DFT x1 (0) = x(0) e 4 = x(0) = 2

If x(n)  X(k)
j2.1
+j2mn
N DFT x1 (1) = x(1)e 4 = x(1)  (+ j) = 2j
then x(n) e  X ((k – m))N
j2.2
Proof : x1(2) = x(2) e 4 = x(2)  (– 1) = – 1
N–1 – j2nk
j23
DFT {x(n)} =  x(n) e N
x1 (3) = x(3) e 4 = x(3)  (– j) = – 4j
n=0
+j2mn
 x1 (n) = {2, 2j, –1, – 4j}
N
DFT {x(n) e } Ex. 3.8.8 : For a discrete time sequence x(n) = (1, 2, 3, 4),
N–1 + j2mn – j2nk DFT is given by X(k) = {10, – 2 + 2j, – 2, – 2 – 2j}.
=  x(n) e N
e N Compute the DFT of x^ (n) = {3, 4, 1, 2} using circular time
n=0 shift property of DFT.
N–1 – j2n(k – m)

=  x(n) e N

n=0
 DSIP (MU-Sem.7-COMP) 3-39 Discrete Fourier Transform (DFT)

Soln. : 3.8.5 Time Reversal

MU - Dec. 2015, Dec. 2017, May 2018)
Given : x (n) = {1, 2, 3, 4} and
Q. State the DFT Properties : Time Reversal
x^ (n) = {3, 4, 1, 2}.
(Dec. 2015, Dec. 2017, May 2018, 2 Marks)
That means x^ (n) is obtained by circularly delaying If x(n) is periodic,
x (n) by 2 positions. It is as shown in Fig. P. 3.8.8. x((– n))N = x(N – m) 0nN–1
DFT
If x(n)  X(k)
DFT
x(N – n)  X(N – k)
OR
DFT
x((– n))N  X(N – k)

ge
N–1 – j2nk

Fig. P. 3.8.8 Proof : DFT {x(m)} =  x(n) e N

n=0
io led
 x^ (n) = x ((n – 2)) N–1 – j2nk
According to circular time shifting property,  DFT {x (N – n)} =  x(N – n) e N
DFT
n=0
ic ow

kl
x ((n – l))N  X (k) WN
N Let, N–n = m n=N–m
jk(N  m)
n
Here, N = 4 and l = 2 N–1

  N
bl kn

DFT DFT {x (N – n)} = x(m) e

2k
 x^ (n) = x ((n – 2))4  X (k) W4 m=0
4
╱
at

– j2k N
Pu ch

N–1 j2km
– j2 ╱

N N
but, WN = e N = x(m) e e
m=0
Te

– j2 – j
but,
 W4 = e 4 =e 2
N–1 jkm
Let,
^
DFT of x^ (n) = X (k) =  x(m) e N

– j m=0
^ 2k  2k
 X (k) = X (k)  W4 = X (k)  e 2 – j2mN – j2mN
N N
^ We now insert e since e
 X (k) = X (k) e
– jk

 We get,
Given, X (k) = {10, – 2 + 2j, – 2, – 2 – 2j}
N–1 – j2km – j2mN
^

0
for k = 0  X (0) = X (0) e = X (0) = 10 DFT {x(N – m)} = x(m) e N
 e N

^ – j m=0
for k = 1  X (1) = X (1) e
– j2m(N  k)
= (– 2 + 2j) (cos  – jsin )
=  x(m) e N
= (– 2 + 2j) (– 1) = 2 – 2j
Comparing this with the standard DFT equation, we
^ – j2
for k = 2  X (2) = X (2) e have
= – 2 (cos 2 – jsin 2) = – 2 DFT
^ – j3
{x (N – m)}  X(N – k)
and for k = 3  X (3) = X (3) e
DFT
= (– 2 – 2j) (cos 3 – jsin 3) OR x((– m))N  X(N – k)
= (– 2 – 2j) (– 1) = 2 + 2j DFT
x((– n))N  X((– k) N
^
 X (k) = {10, 2 – 2j, – 2, 2 + 2j}
 DSIP (MU-Sem.7-COMP) 3-40 Discrete Fourier Transform (DFT)

Solved Example N–1 – j2kn

X (k) =  [ xR (n) + j xI (n) ] e N

…(3.8.7)
Ex. 3.8.9 : Find the DFT of the given sequence n=0
x(n)= {1, 2, 3, 4}. Using the results obtained in (i) and not
But according to Euler’s identify,
otherwise find the DFT of the signal x1(n) = {1, 4, 3, 2}
– j2kn
2 kn  2 kn
Soln. : e N
= cos 
 N  – j sin  N 
(i) We find the DFT using the matrix method
– j2nk Putting this value in Equation (3.8.7) we get,
N–1
X(k) = ∑ x(n) e N N–1
n=0 X (k) = ∑ [ xR (n) + j xI (n) ]
X(k) = [WN] x ; WN is the DFT matrix. n=0
cos 2 kn – j sin 2 kn
 X(1)   2 
X(0) 1 1 1 1 1
  N   N 
  =
 1 –j –1 j   N–1
 X(2)    3  ∑ xR (n) · cos 2 kn – j xR (n) sin 
2 kn

ge
1 –1 1 –1  X (k) =
 X(3)   1 j –1 –j  4  n=0
N   N 
 2 kn 2 kn
+ j xI (n) · cos 
X(k) = {10, – 2 + 2j, – 2, – 2 –2j}
io led 2
(ii) Here x1(n) = {1, 4, 3, 2}  N  – j xI (n) sin  N 
Here j2 = – 1; and writing summation for real and
On careful observation, we note that
ic ow

imaginary parts separately we get,

x1(n) = x(– n) (We assume x (n) to be periodic)
N–1
2 kn 2 kn
DFT
∑ xR (n) cos  N  + xI (n) sin  N 
n
i.e., if x(n)  X(k) X (k) =
bl kn

DFT n=0
then x(– n)  X (– k) N–1
2 kn 2 kn
∑ xR (n) sin 
at

We have already shown that

Pu ch

–j
  N  – xI (n) cos  N 
X(k) = {10, – 2 + 2j, – 2, – 2 – 2j} n=0
 X(– k) = {10, – 2 – 2j, – 2, – 2 + 2j} …(3.8.8)
Te

 DFT {x1(n)} = {10, – 2 – 2j, – 2, – 2 + 2j} Comparing Equations (3.8.8) and (3.8.5) we can write,
N–1
3.8.6 Symmetry Property 2 kn 2 kn
XR (k) = ∑ xR (n) cos 
  N  + xI (n) sin  n 
 The symmetry properties of DFT are derived in the n=0
similar way of DTFT symmetry properties. We know …(3.8.9)
that DFT of sequence x (n) is denoted by X (k).
and
 Now if x (n) and X (k) are complex valued sequence N–1
then it can be represented as follows : XI (k) = – ∑ xR (n) sin 2 kn – xI (n) cos 2 kn
  N   N 
x (n) = xR (n) + j x I (n), 0 ≤ n ≤ N – 1 …(3.8.4) n=0

and X (k) = XR (k) + j XI (k), 0 ≤ k ≤ N – 1 …(3.8.5) …(3.8.10)

Here ‘R’ stands for real part and ‘I’ stands for imaginary Equations (3.8.9) and (3.8.10) are obtained by using
part.
definition of DFT. Similarly we can obtain real and
According to the definition of DFT, imaginary parts of x (n) using definition of IDFT.
N–1 – j2kn N–1
∑ N
 XR (n) =
1
∑ XR (k) cos 2 kn – XI (k) sin 2 kn
X (k) = x (n) e …(3.8.6) n   N   N 
n=0 k=0

Putting Equation (3.8.5) in Equation (3.8.6) …(3.8.11)

and
 DSIP (MU-Sem.7-COMP) 3-41 Discrete Fourier Transform (DFT)
N–1
xI (n) =
1
∑ XR (k) sin 2 kn + XI (k) cos 2 kn Case (ii) : When x (n) is real and even
N   N   N 
k=0
Statement : When x (n) is real and even which means,
…(3.8.12) x (n) = x (N – n) then DFT becomes,
Now we will consider different cases as follows : X (k) = XR (k)
Case (i) : When x (n) is real valued Proof : Since imaginary part is zero; putting xI (n) = 0 in

Statement : If x (n) is real valued then Equation (3.8.9) we get,

N–1
X (N – k) = X (– k) = X* (k) XR (k) = ∑ XR (n) cos 2 kn
 N 
n=0
Proof : According to the definition of DFT,
Similarly IDFT can be written by putting XI (k) = 0 in
N–1
Equation (3.8.11).
X (k) = ∑ x (n) W 
kn
…(3.8.13)
N N–1
∑ XR (k) cos  2 kn 
n=0 1


ge
XR (n) =
Replacing k by N – k, N N 
n=0
N–1
Case (iii) : When x (n) is real and odd
X (N – k) = ∑ x (n) W N
(N – k) n
io led
n=0 Statement : When x (n) is real and odd which means,
N–1 x (n) = – x (N – n) then the DFT becomes,
X (N – k) = ∑ x (n) W N W N
Nn – kn

ic ow

…(3.8.14) N–1
n=0 X (k) = – j ∑ xR (n) sin 2 kn
j2  N 
n
– N n=0
bl kn

Now we have, twiddle factor WN = e Proof : Since x (n) is real, we will put xI (n) = 0 in
j2
Nn
Equation (3.8.8). Similarly x (n) is odd and ‘cos’ is even
 – 
at

= e 
Nn
2 kn
N
  W N
function so we can write, cos  N  = 0. Thus first
Pu ch

 
= e– j 2 n = cos 2n – j sin 2 n
summation in Equation (3.8.8) becomes zero. In the second
Te

Since n is an integer, cos 2 n = 1 and sin 2 n = 0 summation of Equation (3.8.8), putting xI (n) = 0 we get,
Nn N–1
 W N 2 kn
xR (n) sin 
= 1
X (k) = – j ∑  N 
Thus Equation (3.8.14) becomes, n=0

N–1 Similarly IDFT can be written as,

X (N – k) = ∑ x (n) W N 
– kn
…(3.8.15) N–1
n=0 x (n) = j
1
N
∑ XR (k) sin 2 kn
N 

Comparing Equation (3.8.15) with definition of DFT k=0
(Equation (3.8.13)) we get, Case (iv) : When x (n) is purely imaginary
X (N – k) = X (– k) …(3.8.16) sequence

Now using Equation (3.8.13) we can write, When x (n) is purely imaginary which means xR (n) = 0
N–1 and x (n) = j xI (n) then putting xR (n) = 0 in Equation (3.8.9)
X (k) = ∑ x (n) W N 
– kn
* …(3.8.17) we get,
n=0 N–1
Thus, from Equations (3.8.16) and (3.8.17) we get, XR (k) = ∑ xI (n) sin 2 kn
 N 
n=0
X (N – k) = X (– k) = X* (k)
And putting xR (n) = 0 in Equation (3.8.10) we get,
N–1
2 kn
XI (k) = ∑ xI (n) cos 
 N 
n=0
 DSIP (MU-Sem.7-COMP) 3-42 Discrete Fourier Transform (DFT)
Symmetry properties can be summarized as shown in Fig. 3.8.2.

Fig. 3.8.2 : Summary of symmetry property

This summary is shown in Table 3.8.1.

Table 3.8.1

ge
N point sequence x [ n ] 0 ≤ n ≤ N – 1
io led N point DFT

x* [n] X* [N – k]

x* [N – n] X* [k]
ic ow

1
xR [n] Xce [k] =
2
[ X [ k ] + X* [ N – k ] ]
n
1
[X [ k ] – X* [ N – k ] ]
bl kn

j xI [n] Xce [k] =

2
1
at

xce [n] = [x [ n ] + x* [ N – n ] ] XR [k]

Pu ch

2
1
xco [n] = [x [ n ] + x*[ N – n ] ] j XI [k]
Te

Solved Example Now we want remaining three samples namely X (5)

X (6) and X (7). Putting k = 5 in Equation (2),
Ex. 3.8.10 : The first five points of the 8 point DFT of a real
X (5) = X* (8 – 5) = X* (3)
valued sequence are,
{0.25, 0.125 – j0.3018, 0, 0.125 – j 0.0518, 0} Determine the We have X (3) = 0.125 – j 0.0518
remaining three points.
 X* (3) = 0.125 + j 0.0518
Soln. : Given DFT points are :
   X (5) = 0.125 + j 0.0518
X (0) = 0.25
X (1) = 0.125 – j 0.3018 Putting k = 6 in Equation (2),
X (2) = 0 X (6) = X* (8 – 6) = X* (2)
X (3) = 0.125 – j 0.0518
We have X (2) = 0, Thus X* (2) = 0
X (4) = 0
   X (6) = 0
Given sequence is a real valued sequence. According to
the symmetry property we have, Similarly putting k = 7 in Equation (2) we get,
X* (k) = X (N – k) X (7) = X* (8 – 7) = X* (1)
or X (k) = X* (N – k) …(1)
We have X (1) = 0.125 – j 0.3018
This is 8 point DFT. Thus N = 8
 X (k) = X* (8 – k) …(2)  X (7) = 0.125 + j 0.3018
 DSIP (MU-Sem.7-COMP) 3-43 Discrete Fourier Transform (DFT)

Ex. 3.8.11 : The first five DFT points of real and even Ex. 3.8.13 : For a 8-point DFT, the first five DFT coefficients
sequence x(n) of length eight are given below. Find are X(k) = X { 10, 2 + 3j, 1 + 2j, j, 4} Find the remaining
remaining three points. coefficients.

X ( k ) = {5, 1, 0, 2, 3, …} Soln. :
Given :
Soln. : Given DFT points are :
X(0) = 10, X(1) = 2 + 3j, X(2) = 1 + 2j,
X ( 0 ) = 5, X ( 1 ) = 1, X ( 2 ) = 0,
X(3) = j, X(4) = 4
X ( 3 ) = 2 and X ( 4 ) = 3. We know, X(k) = X* (N – k)

According to symmetry property we have, In this case, N = 8

 X(5) = X* (8 – 5) = X*(3) = + j
X* ( k ) = X ( N – k )
X(6) = X*(8 – 6) = X*(2) = 1 – 2j

ge
 X ( k ) = X* ( N – k )
X(7) = X* (8 – 7) = X*(1) = 2 – 3j
This is 8 point DFT. Thus N = 8 Therefore the entire DFT sequence is,
io led
 X ( k ) = X* ( 8 – k ) X(k) = {10, 2 + 3j, 1 + 2j, j, 4, j, 1 – 2j, 2 – 3j}
Note : The magnitude of the DFT is symmetric about the N/2
ic ow

 X ( 5 ) = X* ( 8 – 5 ) = X* ( 3 ) point. While the phase of the DFT is antisymmetric

about the N/2 point.
X ( 5 ) = 2
n
bl kn

Ex. 3.8.14 : The first five points of the 8-point DFT of a real
X ( 6 ) = X* ( 8 – 6 ) = X * ( 2 ) valued sequence are,
at
Pu ch

 X ( 6 ) = 0 X(k) = {10, 2.2 – j 0.3, 0, 1.2 – j0.13, 0}

Determine the remaining three points.

X ( 7 ) = X* ( 8 – 7 ) = X * ( 1 )
Te

Soln. : (We use the symmetry property here)

X ( 7 ) = 1 i.e. X(k) = X* (N – k)

Ex. 3.8.12 : Compute DFT of the sequence We have,

x1 (n) = {1, 2, 4, 2} using property and not otherwise compute X (0) = 10, X (1) = 2.2 – j0.3, X (2) = 0,
DFT of x2 (n) = {1 + j, 2 + 2j, 4 + 4j, 2 + 2j}
X (3) = 1.2 – j 0.13, X (4) = 0
Soln. : x1 (n) = {1, 2, 4, 2}
Since this is a 8-point DFT, N = 8
1 2
1 1 1 1 1

 X1 (k) =
 –j –1 j    X(k) = X* (8 – k)
1 –1 1 –1 4  X(5) = X* (8 – 5) = X* (3)
1 j –1 –j 2
X(6) = X* (8 – 6) = X* (2)
1+2+4+2   9
X1 (k) =  1 – 2 + 4 – 2  =  1 
1 – 2j – 4 + 2j –3 X(7) = X* (8 – 7) = X* (1)

   
 1 + 2j – 4 – 2j   – 3   X(5) = 1.2 + j0.13

Now x2 (n) = {1 + j, 2 + 2j, 4 + 4j, 2 + 2j} X(6) = 0

 x2 (n) = x1 (n) + j x1 (n) X(7) = 2.2 + j0.3

 X2 (k) = X1 (k) + j X1 (k)  The final 8-point DFT sequence is,
 X2 (k) = {9 + 9j, – 3 – 3j, 1 + j, – 3 – 3j} X(k) = {10, 2.2 – j 0.3, 0, 1.2 – j0.13, 0, 2.2 + j0.3}
 DSIP (MU-Sem.7-COMP) 3-44 Discrete Fourier Transform (DFT)

Ex. 3.8.15 : A sequence is given as x (n) = {1 + 2j, 1 + 3j, X(2) + X* (– 2) j – j

X1 (2) = = =0
2 + 4j, 2 + 2j}. From basic definition. Find X (k). 2 2
If x1 (n) = {1, 1, 2, 2} X (3) + X* (– 3)
X1 (3) =
2
x2 (n) = {2, 3, 4, 2}
– 2 – j3 + j
Find X1 (k) and X2 (k) by using DFT only. = =–1–j
2
Soln. : X1 (k) = { X1 (0), X1 (1), X1 (2), X1 (3)}
Given : x (n) = {1 + 2j, 1 + 3j, 2 + 4j, 2 + 2j} = {6, – 1 + j, 0, – 1 – j}

 X (1)   1 – j – 1 j  1 + 3j 
X (0) 1 1 1 1 1 + 2j X (k) – X* (– k)
(b) X2 (k) =
  =
   2j

 X (2)   1 – 1 1 – 1  2 + 4j  X (0) – X* (0)

 X (3)   1 j – 1 – j  2 + 2j 
X2 (0) =
2j
6 + 11j – 6 + 11j
 1 + 2j – j + 3 – 2 – 4j + 2j – 2 
1 + 2j + 1 + 3j + 2 +4j + 2 +2j = = 11
2j
 

ge
= X (1) – X* (– 1)
 1 + 2j – 1 – 3j + 2 + 4j – 2 – 2j  X2 (1) =
2j
 1 + 2j + j – 3 – 2 – 4j – 2j + 2 
io led – j + 2 – j3 2 – j4
= =
2j 2j
 –j 
6 + 11j

  j 2 – j4 2j + 4
=  = =–2–j
ic ow

= j 2j –2
 +j 
 – 2 – 3j  X2 (2) =
X (2) – X* (– 2) j + j
= =1
n
2j 2j
bl kn

 X (k) = {6 + +11j, – j, j, – 2 – 3j} X (3) – X* (– 3)

X2 (3) =
2j
Now, x = x1 (n) + j x2 (n)
– 2 – j3 – j j  – 2 – j4
at
Pu ch

= = =–2+j
x* (n) = x1 (n) – j x2 (n) 2j j  2j 
x (n) + x* (n) = 2 x1 (n) X2 (k) = { X2 (0), X2 (1), X2 (2), X3 (3)}
Te

x (n) + x* (n) = {11, – 2 – j, 1, – 2 + j}

x1 (n) =
2
3.8.7 Complex Conjugate Property
X (k) + X* (– k)
 X1 (k) = DFT
2 If x(n)  X(k)
X (k) – X* (– k) DFT
And X2 (k) =
2j then x*(n)  X*(N – k) = X* ((– k))N

X (k)+ X* (– k) Proof : From the DFT equation we have,

(a) X1 (k) =
2 N–1 – j2nk

X (k) = {6 + 11j, – j, + j, – 2 – 3j} DFT {x(n)} =  x(n) e N

N=0
X* (k) = {6 – 11j, + j, – j, – 2 + j3} N–1 – j2nk

X* (– k) = {6 – 11j, – 2, + j3, – j + j}  DFT {x*(n)} =  x*(n) e N

n=0
X (0) + X* (0)
Hence, X1 (0) = N–1
2
6 + 11j + 6 – 11j
=  [x(n) ej2nk/N]*
= =6 n=0
2 N–1

X1 (1) =
X (1)+ X* (– 1) =  [x(n) e– j2n(N – k)/N]*
2 n=0

– j – 2 + j3 = X*(N – k)
= =–1+j  DFT [x*(n)] = X*(N – k) = X*((– k))N
2
 DSIP (MU-Sem.7-COMP) 3-45 Discrete Fourier Transform (DFT)
3.8.8 Parseval's Theorem 1
= { 100 + 4 + 0 + 4 }
4
DFT
If x(n)  X(k) E = 27
Then Energy of the signal is,
N–1 N–1
Ex. 3.8.17 : Verify Parseval’s Theorem for sequence,
n

E =  |x(n)|2
n=0
=
1
N 
k=0
X(k)2 ()
x (n) = 2
1
u (n)

Proof : Energy of a signal is given by the equation, Soln. : According to Parseval’s Theorem,
N–1
N–1 N–1
E =  |x(n)|2 (3.8.18)  | x (n) |
2
=
1
N  | X (k) |
2
…(1)
n=0
n=0 k=0
N–1

E =  x(n) X*(n) (3.8.19) Step I : Need to prove this equality.

n=0

ge
From the IDFT equation we have, Consider value of N = 4. That means n = 0 to N – 1
N–1

 X(k) WN
1 +nK
Let,  n = 0 to 3
x(n) =
N
io led
k=0
N–1 Thus sequence x (n) is generated as follows,

1
X
+nK
 x (n) = (k) WN (3.8.20) 0
N  1
ic ow

For n = 0 x (0) =
k=0
 2  =1
We substitute Equation (3.8.20) in Equation (3.8.19) 1
x (1) = 
1 1
n
n = 1 =
N-1
1
W–1
  2 2
bl kn

 x(n) N  
+nk
 E = X* (k) WN 2
x (2) = 
1 1
  n =2 =
k=0 n=0
 2 4
at
Pu ch

Rearranging the terms, we have, 3

x (3) = 
1 1
n = 3 =
N-1

N-1
nk
 2 8
 X (k) ·   x(n) WN 
1
E =
Te

N 1 1 1 1
k=0 k = 0   x (n) =     
N-1
 2 4 8


1
 E = X*(k) · X(k) Step II : Consider L.H.S. of Equation (1).
N
k=0
N–1 3
1
E =
N
  X(k) 2 (3.8.21) L.H.S =  2
| x (n) | =  | x (n) |
2

n=0 n=0
 We have,
N-1 N-1 2 2 2 2
L.H.S = |x (0)| + |x (1)| + |x (2)| + | x (3)|
 
1
E =  x(n)  = 2
 X(k)  2
N 1 2  1 2  1 2
n=0 k=0  L.H.S = 1 + 
2 +4 +8
Solved Example
 L.H.S = 1.328125 …(2)
Ex. 3.8.16 : X(k) = {10, –2, 0, 2}.
Compute the energy of the signal x(n). Step III : Consider R.H.S term,
Soln. : From Parseval's Energy theorem, we have, N–1
1

N–1 N–1 2
R.H.S = | X (k) |
 
1 N
E =  x(n)  =
2
 X(k)  2
n=0
N
n=0 k=0
1 First we will calculate DFT X (k) using matrix method.
 E = { X (0) 2 +  X (1) 2 +  X (2) 2 +  X (3) 2}
4
X (k) = [W4] XN
 DSIP (MU-Sem.7-COMP) 3-46 Discrete Fourier Transform (DFT)

 
1 2 2 2
1 = 1 + (2) + (3) + (4)

1 –j –1 j  
 X (0)  1 1 1 1  2
  X (2) = 1 – 1 1 – 1
X (1) = 1 + 4 + 9 + 16 = 30
1
1 j – 1 – j  
    4
 X (3)
 
Step II : First we will calculate DFT of x (n) that is X (k)
1
8
X (k) = [W4] xN
 X (k)
  
= {1.875, 0.75 – j0.375, 0.625, 1 1 1 1 1
0.75 + j0.375}
 X (k) =
 1 –j –1 j   2 
 1 –1 1 –1   3 
Now are will calculate | X (k) |.  1 j –1 –j   4 
 | X (k) | = {1.875  0.838525, 0.625 0.838525}
 
10

=
 – 2 + 2j 

ge
Step IV :
 –2 
1
N–1  – 2 – 2j 
R.H.S =  | X (k) |
2
N
io led
k=0  X (k) = {10, – 2 + 2j, – 2, – 2 – 2j}
3
1 Step III :
 R.H.S =  | X (k) |
2
ic ow

4
k=0 N–1
1
1 R.H.S. =  | X (k) |
2
n

2 2 2
R.H.S = { | X (0) | + | X (1) | + | X (2) | + N
4 k=0
bl kn

2
| X (3) | }
Now | X (k) | = {10, 2.828, 2, 2.828}
 R.H.S = 1.328125

at
Pu ch

Since L.H.S= R.H.S ; Parseval’s theorem is verified. 3

1
 R.H.S. =  | X (k) |
2
Te

Ex. 3.8.18 : State and prove Parseval's theorem for the 4

k=0
following sequence x(n) = {1, 2, 3, 4}.
1 2 2 2 2
4[
Soln. : = | X (0) | + | X (1) | + | X (2) | + | X (3) | ]

Given : 1
= [100 + 7.997584 + 4 + 7.997584]
4
x (n) = {1, 2, 3, 4}
R.H.S. = 30
According to parseval’s theorem,
N–1 N–1 Step IV : Since L.H.S. = R.H.S, Parseval’s theorem is
1
 2
| x (n) | =
N  | X (k) |
2
verified.
n=0 k=0
3.8.9 Multiplication of Two Sequences
Step I :
DFT
If x1(n) X1(k)
Here N = 4
DFT
3 and x2(n)  X2(k)
L.H.S. =  | x (n) |
2 DFT 1
then x1 (n)  x2(n)  [X1(k) ⊛ X2 (k)]
n=0 N

2 2 2 2
= | x (0) | + | x (1) | + | x (2) | + | x (3) |
 DSIP (MU-Sem.7-COMP) 3-47 Discrete Fourier Transform (DFT)
3.8.10 Circular Convolution Property Given two periodic sequences x1(n) and x2(n), circular
convolution can be found out using the following two
MU - Dec. 2015, Dec. 2017, May 2018)
techniques.
Q. State the DFT Properties : Convolution . 1. Concentric circle method
(Dec. 2015, Dec. 2017, May 2018, 2 Marks) 2. Matrix multiplication method
DFT
If x1 (n)  X1 (k)
We shall explain each one in detail.
DFT
and x2 (n)  X2 (k) 3.9.1 Concentric Circle Method
DFT
 x2 (n)  X1 (k) · X2 (k).
then, x1 (n) O Given two periodic sequences x1(n) and x2(n)
X3 (k) = X1 (k) X2 (k) y(n) = x1(n) ⊛ x2(n)
 is the symbol used for circular convolution.
Hence, (O) Step 1 : Plot the samples of x1(n) evenly around the
We find x3 (n) for which DFT is X3 (k). Some books also outer circle in the counter clockwise direction.

ge
use the symbol N to denote circular convolution. Step 2 : Plot the samples of x2(n) evenly around the
inner circle in the clockwise. Make sure the
Proof : From the Periodic convolution we know, positions of x1(0) and x2(0) coincide.
N–1
io led
 x1p (m) x2p
Step 3 : Multiply all the corresponding samples of x1(n)
x3p (n) = (n – m)
and x2(n) and add. This will give us y1(0).
m=0
ic ow

The subscript p stands for periodic. Step 4 : Rotate the inner circle in the direction of the
N–1 outer circle (counter clockwise) by one step and
n
i.e., x3 ((n))N =  x1 ((m))N x2 ((n – m))N
repeat step 3.
bl kn

m=0 Step 5 : Repeat step 4 till x1(0) and x2(0) again coincide.
for 0  n  N – 1 (one period) We shall take an example to understand these steps.
at
Pu ch

x3 ((n))N = x3 (n),
3.9.1(A) Solved Example on Concentric
x1 ((m))N = x1 (m) Circle Method
Te

N–1

 x3 (n) =  x1 (m) x2 ((n – m))N (3.8.22) Ex. 3.9.1 : Perform circular convolution on the given two
periodic sequences.
m=0

Now equation RHS of Equation (3.8.22), represents x1(n) = {1, 2, 3, 4}, x2(n) = {4, 1, 1, 2}
circular convolution of x1 (n) and x2 (n) which is represented Soln. : y(n) = x1(n) ⊛ x2(n)

as (O). We take two concentric circles. Plot x1(n) on the outer
   x2 (n)
x3 (n) = x1 (n) O circle in the counter clockwise direction and x2(n) on the
 DFT {x3 (n)} = DFT { x1 (n) O
 x2 (n) } inner circle in the clockwise direction i.e.,
= X3 (k) = X1 (k) · X2 (k)
DFT
   x2 (n)  X1 (k) · X2 (k)
x1(n) O
Hence, circular convolution reduces to multiplication
while working with the DFT.
We will be solving a few examples after discussing one
more property of the DFT.
3.9 Circular Convolution
One of the most important formula of DFT is that of
circular convolution. We shall now see how to perform
circular convolution.
Fig. P. 3.9.1
 DSIP (MU-Sem.7-COMP) 3-48 Discrete Fourier Transform (DFT)

Fig. P. 3.9.1(a)
Fig. P. 3.9.1(d)
We multiply the corresponding samples and add. This
gives us y(0). y(3) =2  1 + 1  2 + 1  3 + 4  4 = 23

ge
 y(0) = 4  1 + 2  2 + 1  3 + 1  4 = 15 If we now rotate the inner circle, we come back to the
We now rotate the inner circle in the anticlock-wise starting point. Hence we stop here.
io led
direction by one step. This gives us,
 y(n) = {15, 19, 23, 23}
The second method of performing circular convolution
ic ow

is using the Matrix method.

As is evident circular convolution can be performed
n
only when x1(n) and x2(n) are of the same length.
bl kn

If they are not of the same length, then we need to

at
Pu ch

zero pad the shorter sequence and make it equal to the

longer sequence.
Te

Ex. 3.9.2 : Find the circular convolution of two sequences,

x1(n)= {1, – 1, 2, – 4} and x2(n) = {1, 2}
Fig. P.3.9.1(b)
Soln. : Since x2(n) is shorter than x1(n), we zero pad it so
 y(1) = 1  1 + 4  2 + 2  3 + 1  4 = 19 that their lengths become equal
We continue rotating the inner circle in steps of one.  x2(n) = {1, 2, 0, 0}
We now perform circular convolution.

Fig. P.3.9.1(c)

y(2) = 1  1 + 1  2 + 4  3 + 2  4 = 23
Fig. P. 3.9.2
DSIP (MU-Sem.7-COMP) 3-49 Discrete Fourier Transform (DFT)

y(3) = 0  1 + 0  – 1 + 2  2 + 1  – 4
= 0
 y(n) = {– 7, 1, 0, 0}
Ex. 3.9.3 : Determine
y(n) = x(n) N h (n)
Where
x(n) = {1, 2, 3, 1}; h(n) = {4, 3, 2, 2}
Soln. :

We generate a circular matrix of h(n)

Fig. P. 3.9.2(a)

   3 4 2 2  
y(0) = 1  1 + 0  – 1 + 0  – 2 + 2  – 4 = – 7 y (0) 4223 1

 y (1) 
=
  2 
 y (2)   2 3 4 2  3 

ge
 y (3)   2 2 3 4  1 

   
y (0) 17
io led

 y (1) 
=
 19 
 y (2)   22 
 y (3)   19 
ic ow
n
 y (n) = {17,19,22,19 }
bl kn

Fig. P. 3.9.2(b)
Ex. 3.9.4 : For the following sequences

{ 10 0otherwise
n2
at

x1 (n) =
Pu ch

x (n) = {
1 0n2
2 0 otherwise
Te

Compute linear convolution using circular convolution.

Soln. :
 1 0n2
Given : x1 (n) = 
 0 otherwise
Fig. P. 3.9.3(c)
 1 0n2
y(1) = 2  1 + 1  – 1 + 0  2 + 0  4 = 1 and x2 (n) =
 0 otherwise
y(2) = 0  1 + 2  – 1 + 1  2 + 0  – 4 = 0
 x1(n) = {1, 1, 1} and x2(n) = {1, 1, 1}
Here, L1 = L2 = 3
To result of linear convolution would have a
length L1 + L2 – 1

 3+3–1 = 5
We now append zeros to x1(n) and x2(n) to make their
lengths equal to 5.

 x1 (n) = { 1,1,1,0,0} and

x2 (n) = { 1,1,1,0,0}
Fig. P. 3.9.2(d)
 DSIP (MU-Sem.7-COMP) 3-50 Discrete Fourier Transform (DFT)
Step I :
The circular convolution is performed as follows.
Given :

 y(1)   1 1
y(0) 1 0 0 1 1 1
1 0 0 1 x1 (n) = {1, 2, 1, – 2}
 y(2)  =
 1 1 1 0 0
  1
 Now, X1 (k) = [W4] X1N
 y(4)   0 1   0 
y(3) 0 1 1 1 0 0

  
0 1 1 1 1 1 1 1

 1 –j –1 j  2 
2
1
 X1 (k) =
 1 –1 1 –1  1 
 –j  
=
 23  1 j –1 –2

 1    
1+2+1–2 2

=
 1 – 2j – 1 – 2j = – 4j 
 1–2+1+2   2 

ge
 y(n) = {1, 2, 3, 2, 1}  1 + 2j – 1 + 2j   4j 
This is the result that we would have got had we used
io led Similarly,
circular convolution.
Let us verify the result by performing linear X2 (k) = [W4] X2N
convolution. We use the tabular method.
ic ow

  
1 1 1 1 3

 1 –j –1 j  –2
n
 X2 (k) =
 1 –1 1 –1  1
bl kn

 1 j –1 –j  –3
at

   
Pu ch

3–2+1–3 –1

=
 3 + 2j – 1 – 3j = 
2–j

   9
Te

3+2+1+3

Fig. P. 3.9.4
 3 – 2j – 1 + 3j   2+j
Step II :
 y (n) = { 1, 2, 3, 2, 1}
Let Y (k) = X1 (k)  X2 (k)
Hence we get the same result.
= {2, – 4j, 2, 4j}  {– 1,2 – j, 9, 2 + j}
Ex. 3.9.5 : Using the DFT method, obtain the circular
convolution of the following :  Y (k) = {– 2,– 4 – 8j, 18, – 4 + 8j}
x1 (n) = [1 2 1 – 2]
Step III :
x2 (n) = [3 – 2 1 – 3]

Verify your result using the graphical method. The result of circular convolution is obtained by performing

Soln. : From the DFT property we know, IDFT of Y (k) as follows,

1
DFT
 x2(n)  X1(k)  X2(k)
x1(n) O
y (n) =
N [ W ] Y(k)
4

  
1 1 1 1 –2
We begin with computing the DFT of x1(n) and x2(n)
1  1 j –1 –j  – 4 – 8j 
= =
4
 1 –1 1 –1  18 
 1 –j –1 j  – 4 + 8j 
 DSIP (MU-Sem.7-COMP) 3-51 Discrete Fourier Transform (DFT)

  1   
– 2 – 4 – 8j + 18 – 4 + 8j 8 2
1  – 2 – 4j + 8 – 18 + 4j + 8 =  –4  –1
= =
4
 – 2 + 4 + 8j + 18 + 4 – 8j  4 24   6
 – 2 + 4j – 8 – 18 – 8 – 4j   – 36   –9

 y (n) = {2, – 1, 6, – 9}
This is the final Result we verify the result by performing direct circular convolution. We generate a circular matrix of
x1 (n)

    
y (0) 1 –2 1 2 3

 y (1) 
=
 2 1 –2 1   –2 
 y (2)   1 2 1 – 2  1 
 y (3)   –2 1 2 1   –3 

ge
   
y (0) 2

 y (1) 
=
 –1 
 y (2)   6 
io led
 y (3)   –9 

 y(n) = 2, –1, 6, – 9}
ic ow

Which is the same as obtained using the DFT.

n
3.9.2 Matrix Method
bl kn

In this method, we generate a circular matrix of x2(n) and multiply it by x1(n).

at
Pu ch
Te

Note : Since the next column of the circular matrix will be the
y(0) 4 2 1 1  1
same as the first column we stop.   = 1 4
y(1) 2 1  2
The x2 matrix is called a circular matrix. y(2) 1 1 4 2  3
y(3) 2 1 1 4  4
3.9.2(A) Solved Example on Matrix Method y(0) 15
 We get y(1) = 19
Ex. 3.9.6 : Perform circular convolution on the given two y(2) 19
sequence x1(n) = {1, 2, 3, 4}, x2(n) = {4, 1, 1, 2}. y(3) 23
Soln. : i.e., y(n) = {15, 19, 19, 23}
Soln. : We generate a circular matrix of x2(n) and multiply This is the same result as obtained using the concentric
by x1(n). circle method.
y(n) = x1(n) ⊛ x2(n) Ex. 3.9.7 : Given the two sequences of length 4 are,
x(n) = {0,1,2,3}, h(n) = { 2,1,1,2}
Find circular convolution
 DSIP (MU-Sem.7-COMP) 3-52 Discrete Fourier Transform (DFT)
Soln. : The lengths of x (n) and h (n) are equal. We plot
x (n) on the outer circle in counter clock wise direction, and
h(n) on the inner circle in the clock wise direction. We
multiply and add corresponding elements.

Fig. P. 3.9.7(d)

 y (3) = (0  2) + ( 1  1 ) + (2  1 ) + (3  2) = 9

ge
If we rotate one more time, we come back to the starting
Fig. P. 3.9.7(a)
position. Hence we stop the process.
 y (0) = (0  2) + ( 1  2 ) + (2  1 ) + (3  1) = 7
io led
 y (n) = {7, 9, 11, 9}
We now rotate the inner circle and continue to multiply and Ex. 3.9.8 : Find circular convolution for the following
ic ow

add the corresponding elements. Hence we have sequence using graphical method.
(i) x1 (n) =  (n) +  (n – 1) +  ( n – 2)
n
bl kn

x 2 (n) = 2  (n) –  (n – 1) + 2  ( n – 2)

(ii) x1 (n) =  (n) +  (n – 2) –  (n – 2) +  ( n – 3)

at
Pu ch

x2 (n) =  (n) –  (n – 2) +  ( n – 4)

Soln. :
Te

(i) x1(n) and x2(n) can be written is the form

Fig. P. 3.9.7(b) x1 (n) = {1, 1, 1 } and x2 (n) = {2, – 1, 2}

 
 y (1) = (0  1) + ( 1  2 ) + (2  2 ) + (3  1) = 9
Since both x1(n) and x2(n) are of equal lengths, we place
x1(n) on the outer circle in the counter clockwise direction
and x2(n) on the inner circular in the clockwise direction.

Fig. P. 3.9.7(c)

 y (2) = (0  1) + ( 1  1 ) + (2  2 ) + (3  2) = 11
Fig. P. 3.9.8(a)
 DSIP (MU-Sem.7-COMP) 3-53 Discrete Fourier Transform (DFT)
y(0) = (1  2) + (1  2) + (1  ( – 1)) = 3

Fig. P. 3.9.8(d)

y(0) = (1  1) + (1  1 ) + ( – 1  0 ) + (–1  1) + ( 0  0) = 1
Fig. P. 3.9.8(b)

y(1) = (1  ( – 1)) + (1  2) + (1  2) = 3

ge
io led
ic ow

Fig. P. 3.9.8(e)

y(1) = (1  0) + (1  1 ) + ( – 1  1 ) + (–1 0) + ( 0  1) = 0

n
bl kn

Fig. P. 3.9.8(c)
at

y(2) = (1  2) + (1  ( – 1)) + 1 (2) = 3

Pu ch

If we now rotate the inner circle, we come back to the

Te

starting point. Hence we stop the process.

 y (n) = { 3, 3, 3}

(ii) x1(n) and x2(n) can be written is the form Fig. P. 3.9.8(f)

x1(n) = { 1,1,– 1, – 1} y(2) = (1  1) + (1  0 ) + ( – 1  1 ) + (–1  1) + ( 0  0) = – 1

and x2(n) = {1, 0, 1, 0, 1}

We append a zero to x1 (n) to make the lengths of x1(n)

and x2(n) equal.

 x1 (n) = { 1, 1, – 1, – 1, 0} and

x2 (n) = {1, 0, 1, 0, 1}
Fig. P. 3.9.8(g)
We place x1(n) on the outer circle, and x2(n) on the
inner circle y(3) = (1  0) + (1  1 ) + ( – 1  0 ) + (–1  1) + ( 0  1) = 0
 DSIP (MU-Sem.7-COMP) 3-54 Discrete Fourier Transform (DFT)
x2(n) = {1, 1, 1, 0, 0, 0}
Note : In this case we need to append two zeros to x1(n) and
three zeros to x2(n).

We now perform circular convolution. We shall use

the matrix method for its simplicity.
y(n) = x1(n) ⊛ x2(n)

 y(1)   1 1 0 0 0 1   2 
y(0) 1 0 0 0 1 1 1
Fig. P. 3.9.8(h)

 y(4) = (1  1) + (1  0 ) + ( – 1  1 ) + (–1  0) + ( 0  1) = 0  y(2)  =  1 1 1 0 0 0   3 

 y(n) = {1 , 0, – 1, 0, 0 }
 y(3)   0 1 1 1 0 0
  1

 y(5)   0 0 0 1 1 1   0 
Note : Unlike Linear convolution, where the length of the y(4) 0 0 1 1 1 0 0
output signal is L1 + L2 – 1, the length of the output signal in

ge
circular convolution is max (L1, L2)
 y(n) = {1, 3, 6, 6, 4, 1}
3.9.3 Getting Linear Convolution from
io led We now check if this is same as what we would have
Circular Convolution got had we performed linear convolution. We shall again
use the matrix method y(n) = x1(n)* x2(n)
 We can get the linear convolution result using circular
ic ow

convolution by making a small change to the method

described in the earlier example.
n
bl kn

 The main difference between linear and circular

convolution is the length.
at


Pu ch

Given two sequences x1(n) and x2(n) of lengths L1 and  y(n) = {1, 3, 6, 6, 4, 1}
L2 respectively, the length of y(n) using linear Hence we see that the two results are the same.
convolution would be L1 + L2 – 1 while the length of
Te

Go ahead try out your own sums. This technique will

the result of circular convolution will be max(L1, L2).
always work. Given below is a simple program to
 What we do here is append zeros to both x1(n) and implement circular convolution.
x2(n) so that the lengths of both the sequences are CIRCULAR CONVOLUTION
L1 + L2 – 1. After this if we perform circular Clear all
convolution, we get the result of linear convolution. clc
 Let us take an example to prove the above statement.
x= [1 2 6 4 7 3 4];
3.9.4 Solved Examples h= [4 3 2 2];
N2= length (x);
Ex. 3.9.9 : Perform circular convolution on the following
N3= length (h);
sequences so that the result is the same as that of Linear
N = max (N2,N3);
convolution
%% Append Zeros to the shorter signal and make the two
x1(n) = {1, 2, 3, 1}, x2(n) = {1, 1, 1} equal
Soln. : Here the length of x1(n) = 4 and length of x2(n) = 3. if N2< N
Linear convolution would give us a result of length x=[x zeros(1,N–N2) ];
L1 + L2 – 1 i.e., 4 + 3 – 1= 6 else
We now append zeros to both x1(n) and x2(n) so that h=[h zeros (1,N–N3) ];
they are both of length 6. end
 x1(n) = {1, 2, 3, 1, 0, 0} b=h ; c=b ;
for n=1:1:N – 1
 DSIP (MU-Sem.7-COMP) 3-55 Discrete Fourier Transform (DFT)
b=[b(N);b(1:N–1,:) ]; Ex. 3.9.12 : Use the four point DFT and IDFT to determine
c=[c b]; the circular convolution of the given two sequences.
end x1(n) = {1, 2, 3, 1}
CIR_CONV=c*x x2(n) = {4, 3, 2, 2}

Ex . 3.9.10 : Compute circular convolution of Soln. : We use the circular convolution property
x1 (n) = { 1,1,2,2} DFT
i.e. x1(n) ⊛ x2(n)  X1(k)  X2(k)
x2 (n) = {1, 2, 3, 4}
We find the DFT’s of x1(n) and x2(n)
Soln. :
X1 = [WN]  x1
 x2 (n)
y (n) = x1 (n) O

 
X1(0)
  
1 1 1 1 1
We generate a circular matrix of x2 (n)
 X (1)
1
=
 1 – j –1 j  2 
  214 3    X (2)  –1  3 

ge
y (0) 143 2 1 1 –1 1
 –j  1 
1


 y (1) 
=
  1   X (3)
1 1 j –1
 y (2)   3 2 1 4  2 
io led  X1(k) = {7, – 2 – j, 1, – 2 + j}
 y (3)  4321   2 
Similarly

   
y (0) 15 X2 = [WN]  x2
ic ow

 y (1)   17 
 X (1) 
X2(0)
  
= 1 1 1 1 4
 y (2)   15 
   j  3 
n
 y (3)   13 
2 1 – j –1
=
   –1  2 
bl kn

X (2) 1 –1 1
 –j  2 
2

 y(n) = {15, 17, 15, 13}  X (3) 

2 1 j –1
at
Pu ch

 X2(k) = {11, 2 – j, 1, 2 + j}
Ex. 3.9.11: Compute DFT of X (n) = 1, 0  n  2 = 0
X3 = X1(k)  X2(k) = {77, – 5, 1, – 5}
Otherwise N = 4, Find | X (k) | and  X (k)
Te

Since this is in the Fourier domain, we finally calculate the

Soln. : Since N = 4, we have
IDFT.
x (n) = {1,1,1,0} 1
X3(k) = [WN]*  X3
N

 x (1) 
x3(0)
  
1 1 1 1 77
A 4 – point DFT in matrix form is given by the equation
X (k) = [W4 ] x (n)  3  =
1  1 j –1 –j  –5 

 x (2)  4
 1 –1 1 –1  1 
    j  –5 
1 1 1 1 1 3

 1 –j –1 j  1   x (3) 
3 1 –j –1
X (k) =
 1 –1 1 – 1  1   x3 = {17, 19, 22, 19}
 1 j –1 –j   0  This is the result that we would obtain if we circularly
convolved x1(n) and x2(n)
 X (k) = { 3, – j, 1, j }
 x1(n) ⊛ x2(n) = {17, 19, 22, 19}
Now, Let us check our result by performing circular
convolution of x1(n) and x2(n)
Magnitude | X (k) | = {3,1,1,1}
 x3(n)= x1(n) ⊛ x2(n)
 Imaginary
Phase  X (k) = tan –1  
 Real 

  X (k) = { 0, – 90, 0, 90}

 DSIP (MU-Sem.7-COMP) 3-56 Discrete Fourier Transform (DFT)

 x (1) 
x3(0)
  
4 2 2 3 1 Ex. 3.9.15 : Find linear and circular convolution of the

 3  =
 3 4 2 2  2  following sequences

 x (2)   2  3 
x1 [n] = [4, – 2, 2, 1]
2 3 4
 4 1 
3

 x (3) 
3 2 2 3
x2 [n] = [1, 2, 3]

 x3(n) = {17, 19, 22, 19} Soln. :

Hence we get the same result as what we got earlier. Given :

Ex. 3.9.13 : Compute circular convolution of x1(n) = {1,2,3,4} (i) Linear convolution
and x2(n) = {2,1,2,1}
x1 (n) = [4, – 2, 2, 1]
Soln. :
x2 (n) = [1, 2, 3]
x1(n) = {1,2,3,4}
Number of sample in x1 (n) is L1 = 4
x2(n) = {2,1,2,1}

ge
Number of sample in x2 (n) is L2 = 3
 x2(n)
y(n) = x1(n)O io led
OR O
y(n) = x1(n) N x2(n)
So from linear convolution number of samples will be,

L2 + L2 – 1 = 6
We generate a circular matrix of x2(n)
ic ow

x (n) = x1 (n)  x2 (n)

   1 2 1 2  
y (0) 2121 1
n
 y (1)    2  We use the tabular method,
bl kn

=
 y (2)   2 1 2 1  3 
 y (3)   1 2 1 2  4  We assume the first value as the origin,
at
Pu ch

   
y (0) 14

 y (1)   16 
Te

=
 y (2)   14 
 y (3)   16 

 y(n) = {14, 16, 14, 16}

Ex. 3.9.14 : Perform circular convolution of Fig. P. 3.9.15

x1(n) = {2, 1, 2, 1}
 y (n) = {4, 6, 10, – 1, 8, 1}
x2(n) = {4, 3, 2, 1}
(ii) Circular Convolution
Soln. : We generate a circular matrix or x2 (n)
For circular convolution, the length of both the signals
   3 4 1 2  
y (0) 4123 2

 y (1) 
=
  1  should be equal.

 y (2)   2 3 4 1  2 
 y (3)   1 2 3 4  1
We append one zero to x2 (n)

 x1(n) = {4, –2, 2, 1}

   
y (0) 16

 y (1)  =
 14  x2(n) = {1, 2, 3, 0}
 y (2)   16 
 y (3)   14  We generate a circular matrix of x2(n)

 y(n) = {16,14,16,14}  x2(n)

y1(n) = x1(n) O
 DSIP (MU-Sem.7-COMP) 3-57 Discrete Fourier Transform (DFT)

   210 3  
y (0) 103 2 4
Soln. :
 y (1) 
=
  –2  Step I : The four point DFT of x1 (n) is X1 (k) and it is given
 y (2)   3 2 1 0  2 
 y (3)   0321  1 
by,

X1 (k) = [ W4 ] x1
   
y (0) 12

 y (1)   9  
1 1 1 1

 y (2) 
=
 10   1 –j –1 j 
 y (3)   –1 
We have, [ W4 ] =
 1 –1 1 – 1
 1 j –1 –j
 y(n) = {12, 9, 10, 1}
   
1 1 1 1 1
Ex. 3.9.16 : Find circular convolution of two finite
 X1 (k) =
 1 –j –1 j  2
·

duration signals  1 –1 1 – 1  3 
 –j  1 

ge
x1 (n) = {1, – 1, – 2, 3, –1} 1 j –1

 –2–j 
x2 (n) = { 1, 2, 3} 7

 
io led
Soln. : For circular convolution, length of both signals =
 1 
should be equal. We append two zeros to x2(n).
 –2+j 
ic ow

 x1 (n) = {1, – 1, – 2, 3, –1}and

 X1 (k) = { 7 , – 2 – j , 1 , – 2 + j } …(1)
n
bl kn

x2 (n) = { 1, 2, 3, 0, 0}
Similarly, X2 (k) = [ W4 ] x2N

  3 
Now, 1 1 1 1 4
at
Pu ch

 x2 (n)
y (n) = x1 (n) O  X2 (k) =
 1 –j –1 j  
 1 –1 1 – 1 2 
 –j 2 
Te

We generate a circular matrix of x2 (n) . 1 j –1

 4 – 3j – 2 + 2j   2–j 
4+3+2+2 11
    
y (0) 1 0 0 3 2 1
y (1) 2 1 0 0 3 –1
 X2 (k) =
  
 y (2)
 =
 3 2 1 0 0
 –2
  4–3+2–2
 4 + 3j – 2 – 2j
 1
  2+j


 y (4)   0 0 3 2 1  –1 
y (3) 0 3 2 1 0 3

 X2 (k) = { 11 , 2 – j , 1 , 2 + j }

   
y (0) 8
Step II : Now according to property of circular convolution,
y (1) –2
x1 (n) O
   
* x2 (n) = X1 (k) · X2 (k) = Y (k)
y (2) = –1
 
 y (4)   –1 
y (3) –4 Y (k) = { 7 , – 2 – j , 1 , – 2 + j }
· { 11 , + 2 – j , 1 , 2 + j }
 Y (k) = { 77 , – 5 , 1 , – 5 }
 y (n) = { 8, – 2, – 1, – 4, – 1}
Step III : Let the result of x1 (n) O
* x2 (n) be sequence y(n).
Ex. 3.9.17 : Use the four point DFT and IDFT to determine
It is obtained by computing IDFT of Y (k).
the circular convolution of sequences
x1 (n) = (1, 2, 3, 1) x2 (n) = (4, 3, 2, 2) According to the definition of IDFT we have,
  1 *
y (n) = [ WN ] · XN
N
 DSIP (MU-Sem.7-COMP) 3-58 Discrete Fourier Transform (DFT)
1 * We perform element by element multiplication
= [ W4 ] · YN
4
 Y (k) = { 21, – 7 + 4 j, – 3, – 7 – 4 j}
   –5 
1 1 1 1 77
1 1 +j –1 –j    We finally compute the IDFT of y (k),
 y (n) =

4 1 –1 1 –1
  1  1 *

 1 –j –1 +j   –5  Now y (n) =
N
[W4 ] YN

     1
68 17 1 1 1  21
1 1 – j  4j
1  76   19  j –1 – 7 +
 y (n) =
y (n) = = 4 1 –1 1 – 1  – 3
   22 
4 88  1 –j –1 –j  – 7 – 4j
 76   19 
 21 –7 + 4j –2 3 – 7 – 4j 2 4
y (n) = {17‚ 19‚ 22‚ 19}
=  21 + 7 – 4j – 3 + 7 + 4j  =
1 21 – 7j + 4j + 3 + 7j + 4j 1 16
4   4 32
Ex. 3.9.18 : Find the response of FIR filter with impulse  21 + 7j – 4j2 + 3 – 7j – 4j2  32

ge
response h(n) = {1, 2, 4} to the input sequence x (n) = {1, 2}
using circular convolution.  y (n) = {1, 4, 8, 8}
Soln. :
io led 
Step I :
This is the final response of the FIR filter.
x (n) = {1, 2}, h (n) = {1, 2, 4}
ic ow

Verify.
Here L1 = 2 and L2 = 3
n
Let us, Verify the result by performing linear convolution
bl kn

Hence the length of linear convolution result will be

using tabular method
L1 + L2 = – 1 = 2 + 3 – 1 = 4
at
Pu ch

We append zero to make the length of x(n) and h (n) = 4

Te

 x (n) = {1,2,0,0} and h (n) = { 1,2,4,0}

We compute the DFT of x(n) and h (n) and use the property,
DFT
 h (n)  X (k)  H(k)
x (n) O Fig. P. 3.9.18

 1 1 1 1 1
 X (k) =  1 – 1
1 –j –1 j 2 y(n) = { 1, 4, 8, 8}
 1 – 1 0
1 j – 1 – j 0 Hence we get the same result.

 1 +2  Ex. 3.9.19 : If x(n) = {1, 2, 3, 2}, h (n) = {1, 0, 2, 0}

= 1–2 
1 – 2j
(i) Find circular convolution using time domain method.
 
 1 + 2j  (ii) Find linear convolution using circular convolution.
Soln. :
 X (k) = { 3, 1 – 2j, 3 – 3 + 2j }
(i) We will solve it using the graphical method and then
 1 1 1 1 1 test our result using the matrix method.
And H (k) =  1 – 1
1 –j –1 j 2
 1 – 1 4  x2 (n)
y (n) = x1 (n) O
1 j – 1 – j 0 Since the length's of both x1 (n) and x2 (n) are the

 H (k) = { 7, – 3 – 2j, 3, – 3 + 2j } same, we do not need to append zeros to either of the

signals.
Now Y (k) = X (k) . H (k)
 DSIP (MU-Sem.7-COMP) 3-59 Discrete Fourier Transform (DFT)
We take two concentric circles. Steps involved are as  y (0) =(1 0) + (2 1) + (3 0) + (2 2) = 6
follows :
(a) Plot x1 (n) on outer circle in counter clockwise
direction.
(b) Plot x2 (n) on inner circle in clockwise direction.

(c) Ensure the origins of the two signals are aligned.

Multiply and add corresponding elements.
(d) Move inner circle in the clockwise direction and
repeat step (c).

Fig. P. 3.9.19(c)
 y (0) =(1 2) + (2 0) + (3 1) + (2 0) = 5

ge
io led
ic ow

Fig. P. 3.9.19
n
bl kn

Hence we have,
at
Pu ch

Fig. P. 3.9.19(d)
 y (0) =(1 0) + (2 2) + (3 0) + (2 1) = 6
Te

If we now rotate the inner circle, we come back to the

starting position. Hence we stop the process.
 y(n) = { 7, 6, 5, 6 }
  x2 (n) = { 7, 6, 5, 6 }
x1 (n) O

We verify this result using the matrix method. We

form a circular matrix of x2 (n).
Fig. P. 3.9.19(a)

 y (1)   x   x (1) 
x2 (0) x2 (3) x2 (2) x2 (1) x1 (0)
 y (0) =(1 1) + (2 0) + (3 2) + (2 0) = 7 y (0)

 = (1) x2 (0) x2 (3) x2 (2)

 y (2)   x x (3)   x (2) 

2 1

(2) x2 (1) x2 (0)
 y (3)   x x (0)   x (3) 
2 2 1

2
(3) x2 (2) x2 (1) 2 1

 y (1)  0 2 6

y (0) 1 0 2 0 1 7


  =
 1 0 2   = 
 y (2)  2 0 1 0 3 5
 y (3)  0 2 0 1 2 6
 y (n) = { 7, 6, 5, 6 }
which is the same as obtained using the graphical
Fig. P. 3.9.19(b) method.
 DSIP (MU-Sem.7-COMP) 3-60 Discrete Fourier Transform (DFT)

(ii) Find linear convolution using circular convolution. Since response (filter), then the filtered output signal y (n) is
length of x1 (n) = 4 and length of x2 (n) = 4, linear given by the linear convolution operation.
convolution would give us a result of length y (n) = x(n) h (n)
L1 + L2 – 1 = 4 + 4 – 1 = 7 +

We now append zeroes to both x1 (n) and x2 (n). So

i.e. y (n) =  x(k) h (n – k)
k = –
that they are both of length 7.
 In the real world, due to high sampling rates, x(n) is
 x1 (n) = { 1, 2, 3, 2, 0, 0, 0 }
very long and usually runs into thousands of samples.
x2 (n) = { 1, 0, 2, 0, 0, 0, 0 } Because of its length, it is difficult to perform regular
 x2 (n)
y (n) = x1 (n) O linear convolution at one time.
 The way devised to deal with this issue is to break x(n)
We will now perform circular convolution using the
into smaller blocks of size N. Computation on each of
matrix method.
these blocks is then done separately. Each of these

    
y (0) 1 0 0 0 0 2 0 1
processed blocks are then, appended (connected) to

ge
y (1) 0 1 0 0 0 0 2 2

 =  
each other to obtain the final filtered output y (n).
y (2) 2 0 1 0 0 0 0 3
y (3) 0 2 0 1 0 0 0 2  There are two methods of performing this kind of

    
io led
filtering
y (4) 0 0 2 0 1 0 0 0

    
y (5) 0 0 0 2 0 1 0 0 1. Overlap - save method.
2. Overlap - add method.
ic ow

y (6) 0 0 0 0 2 0 1 0
 y (n) = {1, 2, 5, 6, 6, 4, 0} We shall discuss both these methods in detail.
n
We now verify this result by performing regular
3.10.1 Overlap - Save Method
bl kn

convolution.
y (n) = x1 (n) * x2 (n) As mentioned earlier, we break x(n) into smaller blocks
at
Pu ch

of length N. (x1 (n), x2 (n), ). It is important to note that

while the length N can be our choice, it should be longer

Te

than the length of h (n).

i.e. N > length h (n).

We use the following notations.

Where, N = Length of each block

Fig. P. 3.9.19(e)
M = Length of h (n)
 y (n) = { 1, 2, 5, 6, 6, 4, 0 }
L = Values from current x(n)
Hence, we observe that we get the same result using
linear convolution.    – 1) + L
 Each block xi (n) of length N is formed such that it
3.10 Filtering of Long Duration Signals
contains (M – 1) values from the previous block and
 Filtering of a signal is required for various reasons like the remaining L values from the current x(n).
removal of noise, boosting certain frequencies while
attenuating the rest, to name a few.  Since there is an overlap of (M – 1) values from the
previous block, it known as overlap- save method.
 Linear filtering is basically linear convolution that we
have studied earlier.  As the 1st block does not have a previous block, we
 Hence if x(n) is our signal (song, ecg, etc) which has append (M – 1) zeros to it at the beginning. A pictorial
been corrupted by noise, and h (n) is the impulse representation would help us understand this better.
 DSIP (MU-Sem.7-COMP) 3-61 Discrete Fourier Transform (DFT)

ge
io led
ic ow

Fig. 3.10.1
Once these block are created, each of these are "circularly convolved" with h (n).
n
i.e.  h (n)
y1 (n) = x1 (n) O
bl kn

 h (n)
y2 (n) = x2 (n) O
at
Pu ch

∶ ∶ etc.
The final filtered output y (n) is obtained by discarding the first (M – 1) values of each result. y1(n), y2 (n),  and
Te

appending the remaining values. A pictorial representation is shown in Fig. 3.10.2.

Fig. 3.10.2
 DSIP (MU-Sem.7-COMP) 3-62 Discrete Fourier Transform (DFT)

This y (n) is the same as what we would obtain by result. In short, we discard the first (M – 1) = 2 values of

performing linear convolution each result.

i.e.  h (n)
y(n) = x(n) O

It is important to note that we obtained y (n) by

performing circular convolution of x(n) with individual
blocks.

Let us solve an example to understand the method

explained.

Solved Example We check the result by performing linear convolution

 h (n)
y (n) = x(n) O

ge
Ex. 3.10.1 : Let x(n) = {1, 2, 3, 4, 5, 6, 7}
and h (n) = {1, 0, 2} io led
Perform convolution using overlap save method.
Soln. : We choose a block size of N. In this method, each
block contains (M – 1) data points of the previous block
ic ow

followed by L new data points. For the first block, the first
(M – 1) values are set to zero. In our case M – 1 = 2. We
n
choose N = 5 and L = 3.
bl kn

x(n) = {1, 2, 3, 4, 5, 6, 7}
at
Pu ch

3.10.2 Overlap - Add Method

Te

We use the same notations used in overlap–save

method. The overlap - add method is similar to the
overlap–save method except the way the block are created.
In this method, each block is created by taking L values
from x(n) and appending (M – 1) zeros at the end of each
block.
Now h(n) = {1, 0, 2}
The size of each block is N = L + (M – 1).
Since the block size = 5, we append 2 zeros to h(n) A pictorial representation would help us understand
this better.
 h(n) = {1, 0, 2, 0, 0}
Once these blocks are created, each of these are "Circularly
We now perform circular convolution. convolved" with h (n), just like overlap–save method.
i.e.  h (n)
y1 (n) = x1 (n) O
  h (n) = {4, 6, 1, 2, 5}
y1 (n) = x1 (n) O
 h (n)
y2 (n) = x2 (n) O
  h (n) = {12, 15, 8, 11, 14}
y2 (n) = x2 (n) O
The final filtered output y (n), is obtained by
  h (n) = {5, 6, 17, 12, 14}
y3 (n) = x3 (n) O overlapping and adding the last (M–1) values of result
y1 (n), y2 (n) , etc.
We arrange the results with an overlap of (M – 1) A pictorial representation is shown in Fig. 3.10.3.
values as shown and discard the first (M – 1) values of each
DSIP (MU-Sem.7-COMP) 3-63 Discrete Fourier Transform (DFT)

ge
io led
ic ow
n
bl kn

Fig. 3.10.3
at
Pu ch
Te

Fig. 3.10.4
 DSIP (MU-Sem.7-COMP) 3-64 Discrete Fourier Transform (DFT)

3.10.3 Solved Examples on Method Ex. 3.10.3 : Find linear convolution using overlap-add and
overlap-save method of the following sequences :
Let us solve an example to understand the method just
discussed. x[n] = {1, 2, – 1, 2, 3, – 2, – 3, – 1, 1, 1, 2, – 1};

Ex. 3.10.2 : Let x(n) = {1, 2, 3, 4, 5, 6, 7} h[n] = {1, 2, 3}

and h (n) = {1, 0, 2}
Compare the results and state the usage of each method.
Perform convolution using overlap add method.
Soln. :
Soln. : Overlap–add and overlap save methods are used to
perform convolution on long sequences. In both these Overlap save method
methods, the input is divided into blocks. Circular We will form blocks of length 4.
convolution is then performed on each of the blocks.
Here N = 4
Here, we assume that the length of input, x(n) is L
and the length of the impulse response, h (n) = M. Length of h = M = 3

ge
Overlap - add method  M–1 = 2
In this, the input is divided into blocks of data of size L
io led
and (M – 1) zeros are appended to it. This makes the size of
Now to make each block we need L = 2 data from

data blocks N = L + (M – 1). x (n). The different blocks are formed as follows :

In our example we choose a block size of N = 5. For the 1st block, we append M – 1 = 2 zeros
ic ow

Since h (n) has a length of 3, we append M – 1 = x1 (n) = {0, 0, 1, 2}

n
(3 –1) = 2 zeros to L. The length of L = 3.
bl kn

 x(n) = {1, 2, 3, 4, 5, 6, 7}
x1(n) = {1, 2, 3, 0, 0}
at
Pu ch

x2(n) = {4, 5, 6, 0, 0}
x3(n) = {7, 0, 0, 0, 0}
Te

We increase the size of h (n) by appending zeros so

that it is equal to N = 5.
Similarly other blocks can be formed,
 h (n) = {1, 0, 2, 0, 0}
We now perform circular convolution
  h (n) = {1, 2, 5, 4, 6}
y1 (n) = x1 (n) O
  h (n) = {4, 5, 14, 10, 12}
y2 (n) = x1 (n) O
  h (n) = {7, 0, 14, 0, 0}
y3 (n) = x3 (n) O

These results are placed as shown below and added.

The last two terms are added. Since we had appended
2 zeros to h (n).
y = 1 2 5 4 6
Given h (n) = {1, 2, 3}
  Add
4 5 14 10 12 We append zeros to h(n) so that it is equal to N = 4

  Add  h(n) = {1, 2, 3, 0}

7 0 14 0 0 The different output blocks are calculated as follows :
 y (n) = {1, 2, 5, 8, 11, 14, 17, 12, 14}
 h (n)
y1 (n) = x1 (n) O
We would have got the same answer had we
performed linear convolution as shown earlier.
DSIP (MU-Sem.7-COMP) 3-65 Discrete Fourier Transform (DFT)

 2 1 0 3
We create a circular of h(n) 1032
–3
 2 1 0 3   00 
1032
 y5 (n) =
  –1
     3 2 1 0   1
 1
 y1 (n) =
 3 2 1 0   12   0 3 2 1
 0 3 2 1
 – 6 –1 + 0 + 3   – 24 
–3+0+3+2

 0 + 0 + 0 + 6   76 
0+0+3+4
 = 
 =  =
 – 9 – 2 + 1 + 0   – 100 
=
 0 + 0 + 1 + 0   14   0 – 3 + 2 + 1
 0 + 0 + 2 + 2
 y5 (n) = {2, – 4, – 10, 0} ...(5)
 y1 (n) = {7, 6, 1, 4} ....(1)
y2 (n) =  h (n)
x2 (n) O  h (n)
y6 (n) = x6 (n) O

 2 1 0 3   21 
1032
 2 1 0 3
1032
   

ge
 1
 y2 (n) =    1
 3 2 1 0   – 21   y6 (n) =
   2
 0 3 2 1 3 2 1 0
 0 3 2 1  –1
io led
 2 + 2 + 0 + 6   102 
1+0–3+4

 =   2 + 1 + 0 – 3   50 
1+0+6 –2

 = 
ic ow

=
 3 + 4 – 1 + 0   66  =
 0 + 6 – 2 + 2  3 + 2 + 2 + 0   76 
 0 + 3 + 4 – 1
n

bl kn

y2 (n) = {2, 10, 6, 6} ...(2)

 h (n)
y3 (n) = x3 (n) O  y6 (n) = {5, 0, 7, 6} ...(6)
at

 2 1 0 3   –12 
1032
Pu ch

 h (n)
y7 (n) = x7 (n) O
 y3 (n) =
   
   – 23   2 1 0 3
1032
Te

3 2 1 0
 2
 0 3 2 1  y7 (n) =
  –1
 – 2 + 2 + 0 – 6  – 46 
–1+0+9 –4  3 2 1 0   0
 0
 =   0 3 2 1
=
 – 3 + 4 + 3 + 0   104  
2 +0+0+0
  22 
 0 + 6 + 6 – 2  4 –2+0+0 = 
=
 y3 (n) = {4, – 6, 4, 10} ...(3)
 6 – 2 + 0 + 0  4 
–3 
 h (n)
y4 (n) = x4 (n) O  0 – 3 + 0 + 0

 2 1 0 3   – 23 
1032

   
y7 (n) = {2, 2, 4, – 3} ...(7)
 y4 (n) =
 3 2 1 0   –– 13  The final output y (n) is obtained by discarding initial
 0 3 2 1 2 samples from each output block and then by connecting

 6 – 2 + 0 – 3   – 18 
3+0–9 –2 all output blocks one after other.

=
 =   y (n) = {1, 4, 6, 6, 4, 10, 2, – 13, – 1, 0, 0, 7,

 9 – 4 – 3 + 0   – 132  6, 4, – 3}
 0 – 6 – 6 – 1 Let us verify the result obtained by using regular linear
 y4 (n) = {– 8, 1, 2, – 13} ...(4) convolution we use the tabular method.

 h (n)
y5 (n) = x5 (n) O
 DSIP (MU-Sem.7-COMP) 3-66 Discrete Fourier Transform (DFT)

1 2 –1 2 3 –2 –3 –1 1 1 2 –1 3.12 2 D- Discrete Fourier Transform

(2-DFT)
1 1 2 –1 2 3 –2 –3 –1 1 1 2 –1

2 2 4 –2 4 6 –4 –6 –2 2 2 4 –2 1-D DFT is useful while dealing with 1-Dimensiinal

signals. Images are 2 Dimensional signals and we need to
3 3 6 –3 6 9 –6 –9 –3 3 3 6 –3
use 2-D DFT to compute their frequency components. The
 y(n) = {1, 4, 6, 6, 4, 10, 2, –13, –1, 0, 0, 7, 6, 4, –3} 2-D DFT is given by the formula,
M–1 N–1
3.11 Summary of DFT properties F(u, v) =   f(x, y) e (uxM + vyN)
–j2

x=0 y=0

1. Linearity DFT …. (3.12.1)

ax1(n) + b x2(n)  a X1(k) + b X2(k)
We use the separability property while computing the
2. Periodicity X (k) = X (k + N) 2-D DFT.

ge
3. Circular 3.12.1 The Separability Property
DFT
time shift x (n – m)N io led e –j2km/N X(k) This property states that a 2-D DFT can be separated
into two 1-D DFTs.
4. Circular We know (assume a square image)
DFT
x (n) e –j2km/N X(k – m )
–j2 
frequency N N–1 N–1 ux vy
ic ow

shift F(u, v) =   f(x, y) e N + N ...( 3.12.2)

x=0 y=0
n
5. Time
DFT
bl kn

reversal x ((– n)) N  x((– k ))N This can be split up as,
N–1 ux N – 1 –j2vy
 
at

–j2 N
Pu ch

F(u, v) = e N f(x, y) e
6. Complex
DFT x=0 y=0
conjugate x * ((n)) N  x((– k )) N
N–1 vy
Te


–j2
Let F(x, v) = f(x, y) e N
DFT
x (n)  X(N – K) y=0
N–1 ux

–j2
7. Parsevals  F(u, v) = f(x, v) e N
N –1 N –1
theorem 1 y=0
 | x (n) | 2 =  | X (k) | 2
N
n=0 n=0 As stated, the principle advantage of the separability
property is that the 2-D DFT can be obtained in two steps
8. Circular DFT by successive application of 1-D DFT.
x1(n) * x2 (n)  X1(k) · X2 (k)
convolution F(u, v) can be obtained by applying 1-D DFT along the
rows and then along the columns. One could also take the
1-D DFT first along the columns and then along the rows.

Fig. 3.12.1
 DSIP (MU-Sem.7-COMP) 3-67 Discrete Fourier Transform (DFT)
3.12.2 Solved Examples on 2- DFT

Ex. 3.12.1 : Find the DFT of the image :

0 1 2 1

1 2 3 2

2 3 4 3

1 2 3 2
Soln. : We have studied the 1-D DFT matrix. We shall use the DFT along the rows and then along the columns.
Length of each row is N = 4  we need a 4 H 4 DFT matrix.

1 1 1 1  0  4
1 – j – 1 j    = –2
1
DFT of 1
st
row
1 – 1 1 – 1  2  0
1 j – 1 – j  1 –2

ge
1 1 1 1  1  8
1 – j – 1 j    = –2
2 nd
io led DFT of 2 row
1 – 1 1 – 1  3  0
1 j – 1 – j  2 –2
1 1 1 1  2  12 
ic ow

1 – j – 1 j    =  –2 
3
DFT of 3
rd
row
1 – 1 1 – 1  4  0 
1 j – 1 – j  3  –2 
n
bl kn

1 1 1 1  1  8
1 – j – 1 j    = –2
2
DFT of 4
th
row
1 – 1 1 – 1  3  0
at
Pu ch

1 j – 1 – j  2 –2

8 
4 –2 0 –2
Te

Hence we have an intermediate stage 12 

–2 0 –2
–2 0 –2
8 –2 0 2 
Now using the 1-D DFT along the columns of this intermediate image we get

1 1 1 1   4  32 
1 – j – 1 j    =  –8 
8
DFT of 1
st
column
1 – 1 1 – 1   12   0 
1 j – 1 – j   8  –8 
1 1 1 1   –2  –8
1 – j – 1 j   – 2  =  0 DFT of 2
nd
column
1 – 1 1 – 1   –2   0
1 j – 1 – j   –2   0
1 1 1 1  0 0
1 – j – 1 j    = 0
0
DFT of 3
rd
column
1 – 1 1 – 1  0 0
1 j – 1 – j  0 0
1 1 1 1   –2  –8
1 – j – 1 j   – 2  =  0 DFT of 4
th
column
1 – i 1 – i   –2   0
1 j – 1 – j   –2   0
 DSIP (MU-Sem.7-COMP) 3-68 Discrete Fourier Transform (DFT)

32 –8 0 –8

–8 0 0 0
 The final DFT of the entire image is
0 0 0 0

–8 0 0 0

Another way of computing the 2D-DFT is by using the

formula Summary
F = T f T’ ...( 1 )
where T is the N x N DFT matrix. This chapter deals with understanding the DFT. The
The DFT matrix is both symmetric as well as separable importance of the DFT was discussed. Examples were
and hence can be obtained using the relationship solved using the conventional method as well as the matrix

ge
method. Computation of Inverse DFT was explained.
F = TfT ...( 2 )
Drawing the Magnitude as well as the Phase spectrum was
We shall solve the same example by using this formula carried out. We also explained how to identify the
io led
Ex. 3.12.2 : Find the DFT of the given image.
frequency components from the x- axis of the DFT
1 2 1
spectrum. Properties of the DFT were listed and examples
 10
 2 3 2 were solved using each of the properties. The chapter also
ic ow

2 3 4 3 introduced the 2-D DFT.

1 2 3 2
n
Soln. : Since f is a 4  4 matrix, T which is the DFT matrix
Review Questions
bl kn

is given below
1 1 1 1
at

Q. 1 State and prove differentiation property of Fourier

T = 1 –1 1 –1
Pu ch

1 –j –1 j
  transform.
1 j –1 –j Q. 2 State and prove circular convolution property of
Te

1 1 1 10 1 2 1 Fourier transform.

F = 1 –1 1 –12 3 4 3
1 –j –1 j 1 2 3 2 Q. 3 State and prove properties of Discrete Time Fourier
   Transform (DTFT).
1 j –1 –j1 2 3 2
Q. 4 Explain the convolution property of Fourier transform
1 1 1 1 of a DT signal.
1 –j –1 j
1 –1 1 –1 Q. 5 State and prove following properties of Fourier
1 j –1 –j transform :

 4 8 12 8   1 1 1 1  (i) Periodicity (ii) Linearity

=  0
–2 –2 –2 –21 –j –1 j
Q. 6 Explain Overlap-save and Overlap-add algorithms.
 0 0 01 –1 1 –1
–2 –2 –2 –2 1 j –1 –j Q. 7 Explain Circular convolution.
 32 – 8 0 – 8  Q. 8 Prove that linear convolution is obtained using circular
F =  0
–8 0 0 0
convolution property of DFT.
 0 0 0
–8 0 0 0 Q. 9 With the help of neat flow diagram explain DIF FFT
algorithm consider N = 8.
This is the same as what we had got using the
conventional method. Q. 10 Explain the relationship between DTFT and DFT.



 4 Fast Fourier Transform

Syllabus :
Need of FFT, Radix-2 DIT-FFT algorithm, DIT-FFT Flow graph for N = 4 and 8, Inverse FFT algorithm.

4.1 Introduction

ge
We studied the Discrete Fourier Transform (DFT) in the
Chapter 3. io led
The DFT is used to identify the frequency components
present in a discrete time signal. The DFT converts a time
domain signal into a frequency domain signal. We have also written the number of multiplication and
ic ow

A N–point DFT is given by the equation addition required to calculate each X(k)
N–1 X(0) = 1·x(0) + 1· x(1) + 1· x(2) + 1· x(3)  4 multi. & 3 add
n

nk
X(k) = x(n) WN ; k =0, 1, , N–1 (4.1.1)  X(1) = 1·x(0) + (–j)·x(1) + (–1)·x(2) + j·x(3)  4 multi. & 3 add
bl kn

n=0
 X(2) = 1·x(0) + (–1)x(1) + (1)·x(2) + –(1)x(3)  4 multi. & 3 add
In the matrix form it is given by the equation
 X(3) = 1·x(0) + j x · (1) + (–1)·x(2) + –(j)·x(3)  4 multi. & 3 add
at
Pu ch

X = [W]x (4.1.2)
 16multi. & 12add

4.2 Computational Complexity of DFT X(k) = {4, –2, 0, –2}

Te

As is shown each X(k) requires 4 multiplications and 3

Now, the question that we need to ask ourselves is additions.
How many multiplications and additions are required to Hence a 4-point DFT requires a total of
compute a N-point DFT ? 16 multiplications and 12 additions. i.e.(4)2 multiplications
We will solve an example to answer this question. and 43additions.
We can generalize this by stating the A N-point DFT
Solved Example requires N2 multiplications and N (N – 1) additions.

Ex. 4.2.1 : Compute the DFT of the sequence Therefore computing the DFT is computationally very
expensive.
x(n) = {0, 1, 2, 1}.
It has been mentioned in earlier chapters that because
Soln. : Since the length of the signal is 4,
of high sampling rates in the real world, the length of x(n) is
 N = 4  4-Point DFT very large and runs into thousands of samples.
We generate a 44 DFT matrix and use Computing the DFT of such signals would require a
Equation (4.1.2). machine which has a lot of computation power. It would
also take a lot of time.
To cut a long story short, the DFT is computationally
very demanding and we require something that would
reduce the number of computations.
This is where the Fast Fourier Transform comes in.
 DSIP (MU-Sem. 7-COMP) 4-2 Fast Fourier Transform

Fast Fourier Transform (FFT)  The two commonly used algorithms are
 In 1965, Cooley and Tukey proposed a method of 1. Decimation in Time FFT (DIT-FFT)
computing the DFT which required only N log2 N 2. Decimation in Frequency FFT (DIF-FFT)
calculations.  We shall derive and discuss each one in detail but
 This method came to be known as the Fast Fourier before doing that let us try to visualize a N-point DFT
Transform (F.F.T)
formula.
 The FFT refers to a class of algorithms for efficiently
computing the DFT. We know a N-point DFT is given by the equation
N-1
 Hence FFT is not a new transform but a set of
 x(n) WN ;
nk
algorithms used for computing the DFT efficiently. X(k) = k =0, 1, , N– 1
n=0
 The results obtained from the DFT and the FFT are the
same.  Here x(n) is the input and X(k) the output. This
 The computational efficiency of the FFT becomes clear equation can be pictorially represented as shown in
when the values of N2 and N log2 N are compared for Fig. 4.2.1

ge
several values of N.
Number of Number of
Number of
io led
computations computations
Samples (FFT) N log2 N
(DFT) N2
ic ow

32 1024 160
128 16,384 896
n
2048 4,194,304 22,528
bl kn

 We notice that as N increases, the FFT becomes more Fig. 4.2.1

efficient.
at


Pu ch

The DFT block computes the DFT of the input x(n) and
 As mentioned earlier, FFT is an algorithm. The
gives us the output X(k).
commonly used technique is called the RADIX-2 FFT

Te

algorithm. Both, DIF-FFT and DIT-FFT, are based on the basic

principle of breaking this N-Point DFT into smaller DFT
 We have studied the Twiddle factor in the previous
Blocks.
chapter. There are two important properties of the
Twiddle factor. 4.3 Decimation In Time FFT (DIT-FFT)
N k
Wk + 2 = –WN Symmetric Property
N  To decimate means to kill. Decimation in time means
k+N k
WN = WN Periodic property killing (re-arranging) the time signal.

 The FFT exploits these two properties to reduce the  Here we exploit the symmetric and periodic property
number of complex multiplications. of the twiddle factor WN
k + N/2 k
 The basic principle of the FFT algorithm is to break WN = – WN  Symmetric
(decompose) a N-point DFT into smaller DFTS. As k+N k
WN = WN  Periodic
mentioned earlier, FFT is an algorithm and is known as
Radix-2 which simply means that the number of  These two important properties reduce the number of
output -points is expressed an a power of 2 i.e. N=2 ,m computations.
where m is an integer. A N-point-DFT is given by the equation,
 Hence when the FFT algorithm breaks a DFT into N–1


nk
smaller DFTS’ the output of the smaller DFTS’ are X(k) = x(n) WN k = 0, 1, 2, … N – 1
always powers of 2. n=0
 DSIP (MU-Sem. 7-COMP) 4-3 Fast Fourier Transform
N
 In decimation-in-time technique, we split the input 2 –1
nk
x(n) into odd and even parts. Let
 x(2n) WN/2 = F1 (k)
n=0

nk

nk
X(k) = x(n) WN + x(n) WN N
n  even n  odd 2 –1
N

nk
N N and x(2n + 1) WN/2 = F2 (k) ; k = 0, 1, …
2 –1 2 –1 2
n=0
 
2nk (2n + 1)k
X(k) = x(2n)WN + x(2n + 1)WN k
n=0 n=0  X(k)= F1 (k) + WN F2 (k) … (4.3.2)
N N
2 –1 2 –1 N
This equation gives us the first 2 values
 x(2n) WN +  x(2n + 1)W2nk
2nk
X(k) =
N
n=0 n=0  Since F 1 (k) and F2 (k) are periodic with period N/2 we
N N
2 –1 2 –1 can write
 

ge
2nk k
X(k) = x(2n) WN + WN x(2n + 1)W2nk F1 (k + N/2) = F1 (k) and
N
n=0 n=0
F2 (k + N/2) = F2 (k)
2
Now, WN = WN/2
io led k + N/2 k
also WN = – WN
N N
2 –1 2 –1  Equation (4.3.2) can be written as,
 
nk nk
X(k) = x(2n) WN/2 + x(2n + 1) WN/2 N
ic ow

k
 X(k) = F1 (k) + WN F2 (k) ; k = 0,1, … –1
n=0 n=0 2

 X k +
N N
n
k
= F (k) – WN F2 (k); k = 0, 1, … – 1
N/2 point DFT N/2 point DFT  2 1 2
bl kn

…(4.3.1) … (4.3.3)
at

 2
We observe that F1(k) requires (N/2) complex multiplications and the same applies to F2 (k). Furthermore, there are
Pu ch

N/2 additional complex multiplications required to compute WNk F2(k). Hence the computation of X(k) requires
complex multiplications which is almost half the computations required by the conventional method.
Te

2
N2 N
2 
N N
+ = +
2 2 2 2
Let us take an example where N = 8,
 Equation (4.3.3) can be shown using a signal flow graph.
Since in DIT - FFT, we have split the input x(n) into even and odd parts, we have

Fig. 4.3.1
 DSIP (MU-Sem. 7-COMP) 4-4 Fast Fourier Transform
 As stated earlier, F1(k) and F2 (k) are N/2 point DFT’s (4 point DFT’s if N = 8). These can be further split up into odd and
even parts. We rewrite Equation 4.3.1.
N N
2 –1 2 –1

 
nk k nk
i.e. X(k) = x(2n) WN/2 + WN x(2n + 1) WN/2
n=0 n=0

F1(k) F2 (k)

 Splitting up F1 (k) and F2 (k) into odd and even parts, we get,
N N
2 –1 2 –1

 
nk k nk
X(k) = x(2n) WN/2 + WN x(2n + 1) WN/2

ge
n=0 n=0

io led
2 parts of N/4 each 2 parts of N/4 each

(A) (B)
ic ow

N N
N
4 –1 4 –1
2 –1
 x(4n) WN/2 
nk 2nk (2n + 1) k
n
(A) F1 (k) =
 x(2n) WN/2 = + x(4n + 2) WN/2
bl kn

n=0 n=0
n=0
N N
u –1 u –1
at
Pu ch

 
2nk 2nk k
= x(un) WN/L + x(4n + 2) WN/2 .WN/2
n=0 n=0
Te

= x(un) WN/2
2nk
+ WN/2
k
x(un + 2) W2nk
N/2

2 1
Now WN/2 = WN/4
N N
N
4 –1 4 –1
2 –1
 
nk nk k nk
 F1 (k) =
 x(2n) WN/2 =
n=0
x(4n) WN/4 + WN/2
n=0
x(4n + 2) WN/4
n=0

N/4 point DFT N/4 point DFT

N N
4 –1 4 –1

 
nk k nk
 F1 (k) = x(4n) WN/4 + WN/2 x(4n + 2) WN/4
n=0 n=0
N
4 –1


nk
Let x(4n) WN/4 = G1(k)
n=0
N
4 –1


nk
and x(4n + 2) WN/4 = G2 (k)
n=0
 DSIP (MU-Sem. 7-COMP) 4-5 Fast Fourier Transform
k
F1(k) = G1(k) + WN/2 G2 (k)
N
This equation will give us the first 4 values

In a similar manner we can write for (B).

k
(B) F2 (k) = H1 (k) + WN/2 H2 (k)
Now, G1 (k), G2 (k), H1 (k) and H2 (k) are all N/4 point DFT’s and are periodic with period N/4. Using the symmetric and
periodic properties of the twiddle factor we get


k
F1(k) = G1(k) + WN/2 G2 (k)

 k = 0, 1, 2, …, 4 – 1
k N
 F1(k + N/4) = G1(k) – WN/2 G2 (k) … (4.3.4)

k
F2 (k) = H1 (k) + WN/2 H2 (k) 

ge
k
 F2 (k + N/4) = H1 (k) – WN/2 H2 (k)
io led
For N = 8, k will have values 0, 1.
We have split F1 (k) and F2 (k) which were N/2 point DFT’s into two N/4 point DFT’s each.
We draw the signal flow graph using Equation (4.3.4) and the earlier Fig. 4.3.1. Let N = 8
ic ow
n
bl kn
at
Pu ch
Te

Fig. 4.3.2
A 2-Point DFT is drawn as,

Fig. 4.3.3
Hence the final signal flow graph is shown in Fig. 4.3.4.
 DSIP (MU-Sem. 7-COMP) 4-6 Fast Fourier Transform

ge
io led Fig. 4.3.4 : 8-point DIT-FFT Butterfly diagram
This diagram is also called an 8-point butterfly diagram.
The DIT-FFT refers to the method in which the input samples are broken into odd and even parts. This time decimation
leads to the scrambled order of the input.
ic ow

The first decimation gives x(0), x(2), x(4), x(6), x(1)....

The second decimation gives x(0), x(4), x(2), x(6)....
n
bl kn

4.4 Bit Reversal Format

at
Pu ch

The shuffling of the input data is known as bit-reversal. This is because the order of the input data can be obtained by
reversing the bits of the binary representation of the normal input data index order
Te

x(4) = 100  BR  001  x(1)

Original Binary New order
Bit reversed Decimal
Samples Representation of input
x(0) –– 0 000 000 0 –– x(0)
x(1) –– 1 001 100 4 –– x(4)
x(2) –– 2 010 010 2 –– x(2)
x(3) –– 3 011 110 6 –– x(6)
x(4) –– 4 100 001 1 –– x(1)
x(5) –– 5 101 101 5 –– x(5)
x(6) –– 6 110 011 3 –– x(3)
x(7) –– 7 111 111 7 –– x(7)
Hence the input that is fed to the algorithm is in the bit reversed format.

4.4.1 Solved Examples on Bit Reversal Format

Ex. 4.4.1 : An 8-point sequence x(n) is given by , {0, 1, 2, 3, 2, 1, 5, 2, 1}

Find the DFT of the sequence using DIT-FFT.
 DSIP (MU-Sem. 7-COMP) 4-7 Fast Fourier Transform

Soln. :Since N = 8, we draw a 8-point DIT-FFT Butterfly diagram remember, the input to the network is in the bit reversed
order.

ge
io led
Fig. P.4.4.1
ic ow

x(n) = 1 , 2 , 3 , 2 , 1 , 5 , 2 , 1 } Stage 2
n
x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7) 0
F1(0) = G1 (0) + W4G2 (0) = 2 + 1 · (5) = 7
bl kn

Here W40 = 1 1
F1(1) = G1 (1) + W4G2 (1) = 0 + (– j)(1) = – j
-j2
 
at
Pu ch

1 4
W4 = e = cos – j sin =–j 0
2 2 F1(2) = G1 (0) –W4G2 (0) = 2 – 1 · (5) = – 3
Similarly, 1
F1(3) = G1 (1) –W4G2 (1) = 0 – (– j)(1) = +j
Te

0
W8 = 1 0
–j2·
F2(0) = H1 (0) + W4H2(0) = 7 + 1 · (3) =10
1 8  
W8 = e = cos – j sin = 0.707 –j0.707 1
F2(1) = H1 (1) + W4H2(1) = –3 + (– j)(1) =– 3–j
4 4
–j2·
2 8   0
F2(2) = H1 (0) –W4H2(0) =7 – 1 · (3) =4
W8 = e = cos – j sin = –j
2 2
1
–j2· F2(3) = H1 (1) –W4H2(1)
3 8 3 3
W8 = e = cos – j sin = –0.707 –j0.707
4 4 = –3 – (– j)(1) =– 3 + j
We write down the equations at each stage and
Stage 3
calculate the outputs.
0
Stage 1 X(0) = F1(0) + W8F2(0) = 7 + 1 · (10) =17
G1 (0) = x(0) + x(4) = 1 + 1 = 2 1
X(1) = F1(1) + W8F2(1)
G1 (1) = x(0)–x(4) = 1 – 1 = 0
= – j + (0.707 – j 0.707)(– 3 – j)
G2 (0) = x(2) + x(6) = 3 + 2=5
= – 2.83 + j0.41
G2 (1) = x(2) – x(6) = 3 – 2 = 1 2
X(2) = F1(2) + W8F2(2)
H1 (0) = x(1) + x(5) = 2 + 5 = 7
= – 3 + (– j) (4) =– 3 – j4
H1 (1) = x(1) – x(5) = 2 – 5 = – 3 3
X(3)= F1(3) + W8F2(3)
H2 (0) = x(3) + x(7) = 2 + 1 = 3
= j + (– 0.707 – j0.707) (– 3 + j)
H2 (1) = x(3) – x(7) = 2 – 1 = 1 = 2.83 + j2.41
 DSIP (MU-Sem. 7-COMP) 4-8 Fast Fourier Transform
0
X(4) = F1(0) – W8F2(0) =7 – 1 · (10) = – 3
1
X(5) = F1(1) – W8F2(1) = – j–(0.707 – j0.707) (– 3 – j) = 2.83 – j2.41
2
X(6) = F1(2) – W8F2(2) = – 3 – (– j) (4) =– 3+ j4
3
X(7) = F1(3) – W8F2 (3) = j – (– 0.707 – j0.707) (– 3 + j) = – 2.83 – j0.41

 X(k) = {17, – 2.83 + j0.41, – 3–j4,2.83 + j2.41, – 3, 2.83 – j2.41, – 3 + j4, – 2.83 – j0.41}
Ex. 4.4.2 Find s-point DIT of the sequence x(n) = {1, 2, 3, 4, 4, 3, 2, 1} using DIT Radix 2 FFT
Soln. : Since N = 8, we draw a 8-point DIT-FFT diagram
x(n) = { 1 , 2 , 3 , 4 , 4 , 3 , 2 , 1 }
x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)

ge
io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 4.4.2

Here W04 = 1 Stage 1

–j2 G1 (0) = x(0) + x(4) = 1 + 4 = 5
1 4
·1  
W4 = e = cos – j sin = –j
2 2 G1 (1) = x(0) – x(4) = 1 – 4 = – 3
Similarly, G2 (0) = x(2) + x(6) = 3 + 2 = 5
0
W8 = 1 G2 (1) = x(2) – x(6) = 3 – 2 = 1
–j2
1 4
·1   H1 (0) = x(1) + x(5) = 2 + 3 = 5
W8 = e = cos – j sin = 0.707 – j0.707
4 4 H1 (1) = x(1) – x(5) = 2 – 3 = – 1
–j2
2 4
·2   H2 (0) = x(3) + x(7) = 4 + 1 = 5
W8 = e = cos – j sin = – j
2 2
–j2 H2 (1) = x(3) – x(7) = 4 – 1 = 3
3 4
·3 3 3
W8 = e = cos – j sin = – 0.707 – j0.707 Stage 2
4 4
0
We write down the equation at each stage and F1(0) = G1 (0) + W4 G2 (0) = 5 + 1 (5) = 10
calculate the outputs. 1
F1(1) = G1 (1) + W4 G2 (1) = –3 + (– j) (1) = – 3 – j
0
F1(2) = G1 (0) – W4 G2 (0) = 5 – 1 (5) = 0
 DSIP (MU-Sem. 7-COMP) 4-9 Fast Fourier Transform
1 3
F1(3) = G1 (1) – W4 G2 (1) = – 3 – (– j) (1) = – 3 + j X(7) = F1 (3) – W8 F2 (3)
0
F2(0) = H1 (0) + W4 H2(0) = 5 + 1 (5) = 10 = (– 3 – j) – (– 0.707 – j0.707) (– 1 + j3)

F2(1) = H1 (1) +
1
W4 H2(1) = – 1 + (– j) (3) = – 3 – j3 = – 5.828 + j2.414
0
F2(2) = H1 (0) – W4 H2(0) = 5 – 1 (5) = 0
 X(k) = { 20, – 5.828 – j2.414, 0, – 0.172 – j0.414, 0,
1
F2(3) = H1 (1) – W4 H2(1) = – 1 – (– j) (3) = – 1 + j3 – 0.172 + j0.414, 0, – 5.828 + j 2.414
Stage 3
0
X(0) = F1 (0) + W8 F2 (0) = 10+ (1) (10) = 20 Ex. 4.4.3 : Given x(n)=2n. Find X(k) using DIT-FFT
1
X(1) = F1(1) + W8 F2 (1) algorithms. Assume x(n) is of length 8.
Soln. : Here N=8we vary n from 0to 7.
= (– 3 – j) + (0.707 – j 0.707) (– 1 – j3)
  x(n) = 2n, 0  n <7
= – 5.825 – j 2.414
  x(0) = 20=1

ge
2
X(2) = F1(2) + W8 F2 (2)
  x(1) = 21 =2
= 0 + (– j) (0) = 0   x(2) = 22 =4
io led
3
X(3) = F1(3) + W8 F2 + (3)   x(3) = 23 =8
= (– 3 + j) + (– 0.707 – j0.707) (– 1 + j3)   x(4) = 24 =16
ic ow

= – 0.172 – j 0.414   x(5) = 25=32

0
X(4) = F1 (0) – W8 F2 (0) =10 – (1) (10) = 0   x(6) = 26=64
n
  x(7) = 27=128
bl kn

1
X(5) = F1(1) – W8 F2 (1)
  x(n) = {1, 2, 4, 8, 16, 32, 64, 128}
= (– 3 – j) – (0.707 – j0.707) (– 1 – j3)
at
Pu ch

Since N= 8, we draw a 8 point DIT-FFT. Remember the

= – 0.172 + j0.414
input is the bit-reversed format.
2
X(6) = F1 (2) – W8 F2 (2) = 0 – (–j) (0) = 0
Te

Fig. P.4.4.3
0
Here W4 = 1
–j2
1 4
·1  
W4 = e = cos – j sin = – j
2 2
 DSIP (MU-Sem. 7-COMP) 4-10 Fast Fourier Transform
0
Similarly, W8 = 1 = (– 15 + j60) + (0.707 – j0.707)(– 30 + j120)
–j2
1 8
·1   = 48.64 + j166.07
W8 = e = cos – j sin
2 2
X(2) = F1(2) + W82 F2(2)
= 0.707 – j 0.707
–j2 = – 51 + (– j) (– 102) =– 51 + j102
2 8
·2  
W8 = e = cos – j sin = – j X(3) = F1(3) + W83 F2(3)
2 2
–j2
3 8
·3 3 3 = (– 15 – j60) + (– 0.707 – j0.707)(– 30 – j120)
W8 = e = cos – j sin
4 4
= – 78.64 + j46.07
= – 0.707 – j 0.707
X(4) = F1(0) – W08 F2(0) = 85 – 1 · (170) = – 85
We write down the equations at each stage
Stage 1 X(5) = F1(1) – W18 F2(1)
G1 (0) = x(0) + x(4) = 1 + 16 = 17 = (– 15 + j60) – (0.707 – j0.707) (– 30 + j 120)

ge
G1 (1) = x(0) – x(4) = 1–16 = – 15
= – 78.64 – j46.07
G2 (0) = x(2) + x(6) = 4 + 64 = 68
io led X(6) = F1(2) – W28 F2(2)
G2 (1) = x(2) – x(6) = 4 – 64 = – 60
= – 51 – (– j)(– 120)=–51 – j102
H1 (0) = x(1) + x(5) = 2 + 32 = 34
X(7) = F1(3) – W38 F2(3)
H1 (1) = x(1) – x(5) = 2 – 32 = – 30
ic ow

H2 (0) = x(3) + x(7) = 8 + 128 = 136 = (– 15 – j60) – (– 0.707 – j0.707) (– 30 – j120)

n
H2 (1) = x(3)–x(7) = 8–128 = – 120 = 48.6 – j166.07
bl kn

Stage 2 X(k) = {255,48.64 + j166.047,– 51 + j102,

– 78.64 + j46.07, – 85, – 78.64 – j46.07,
at

0
F1 (0) = G1(0) + W4 G2 (0) =17 + 1 ·(68) = 85
Pu ch

– 51 – j102,48.6 – j166.07}
F1 (1) = G1(1) + W14 G2 (1)
Ex.4.4.4 : Compute the DFT of x(n)= {1,0,2,0,3,0,4,0} using
Te

= – 15 + (– j) (– 60) = – 15 + j60
0
DIT-FFT.
F1 (2) = G1(0) – W4 G2 (0)
Soln. : Since N = 8, we drawn a 8-point DIT-FFT Butterfly
= 17 – 1(68) = –51 diagram. The input is in the bit reversed order.
F1 (3) = G1(1) – W14 G2 (1) x(n) = {1, 0, 2, 0, 3, 0, 4, 0}
Here W40 = 1
= – 15 – (– j)(– 60) =–15 – j 60
–j2 · 1
F2 (0) = H1(0) + W04 H2 (0)  
W14 = e u
= cos – j sin = – j
2 2
= 34 + 1 · (136) = 170 Similarly,
1
F2(1) = H1(1) + W4 H2(1) W08 = 1
–j2 · 1
= – 30 + (– j)(– 120) =– 30 + j120  
W18 = e 8
= cos – j sin = 0.707 – j0.707
0 4 4
F2(2) = H1(0) – W4 H2(0) =34 – 1 · (136) = – 102 –j2 · 2
 
F2(3) = H1(1) –
1
W4 H2(1) W28 = e 8
= cos – j sin = – j
2 2
–j2 · 3
= – 30 – (– j)(– 120) = – 30 – j120 3 3
W38 = e 8
= cos – j sin = – 0.707 – j0.707
4 4
Stage 3
We write down the equation at each stage and calculate
X(0) = F1(0) + W80 F2(0) = 85 + 1(170) = 255
the outputs
X(1) = F1(1) + W81 F2(1)
 DSIP (MU-Sem. 7-COMP) 4-11 Fast Fourier Transform

ge
io led
Fig. P. 4.4.4

Stage 1 F2 (1) = H1(1) + W14H2 (1) = 0 + (– j) (0) = 0

ic ow

G1 (0) = x(0) + x(4)=1 + 3 = 4 F2 (2) = H1(0) – W04H2 (0) = 0 – 1 ·(0) = 0

n
bl kn

G1 (1) = x(0) – x(4)=1 – 3 = – 2 F2 (3) = H1(1) – W14H2 (1) = 0 – (– j) (0) = 0

G2 (0) = x(2) + x(6)= 2 + 4 = 6 Stage 3

at
Pu ch

G2 (1) = x(2) – x(6) =2 – 4 = – 2 X(0) = F1(0) + W08F2 (0) = 10 + 1· (0) = 10

Te

H1 (0) = x(1) + x(5) = 0+0 = 0 X(1) = F1(1) + W18F2(1)

H1 (1) = x(1) – x(5) = 0 – 0 = 0 = (– 2 + 2 j) + (0.707 – j0.707) · 0

= –2+j2
H2 (0) = x(3) + x(7) = 0 + 0 = 0
X(2) = F1(2) + W28F2 (2) = – 2 + (– j) (0) = – 2
H2 (1) = x(3) – x(7) = 0 – 0 = 0
X(3) = F1(3) + W38F2(3)
Stage 2
= (– 2 – 2j) + (– 0.707 – j 0.707) · 0
F1 (0) = G1(0) + W04G2(0) = 4 + 1 ·(6) = 10
= – 2– j 2
F1 (1) = G1(1) + W14G2 (1) = – 2 + (– j) (– 2)
X(4) = F1(0) – W08F2 (0) = 10 – 1 · (0) = 10

= – 2 + j2 X(5) = F1(1) – W18 F2 (1)

F1 (2) = G1(0) – W04G2 (0) = 4 – 1 · (6) = – 2 = (– 2+2 j) – (0.707 – j 0.707) · 0

F1 (3) = G1(1) – W14G2 (1) = – 2 – (– j) (– 2) = – 2 + j2

X(6) = F1(2) – W28F2 (2) = – 2 – (– j) (0)
= – 2 – j2

F2 (0) = H1(0) + W04H2 (0) = 0 + 1 · (0) = 0 = –2

 DSIP (MU-Sem. 7-COMP) 4-12 Fast Fourier Transform

X(7) = F1(3) – W38 F2 (3) x(3) = 0.2

x(4) = 0
= (– 2 – 2 j) – (– 0.707 – j0.707)·0 x(5) = – 0.2

= –2–j2 x(6) = – 1
x(7) = – 1.2
 X(k) = {10, – 2 + j 2, – 2, – 2 – j2, 10,
x(n) = {0, 1.2, 1, 0.2, 0, – 0.2, – 1, – 1.2}
– 2 + j 2, – 2,–2 – j 2} x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)

Since N = 8, we draw a 8 – point DIT-FFT Butterfly

Ex. 4.4.5 : Sample the given continuous time signal
diagram. The input is in bit reversed order.
x(t) = sin (2  1000t)+ 0.5 sin (2 1000 t) at 8000 sample
Here, W40 = 1
sec. Find out the Eight point DFT using the DIT-FFT
algorithm. –j2 · 1
4  
Soln. :x(t) = sin (2 t) + 0.5 sin (2 2000 t) W14 = e = cos – j sin = – j
2 2

ge
n
To obtain a sampled version of x(t), we replace t by . Similarly,
Fs
W08 = 1
io led
Here Fs = 8000 samples /sec
–j2 · 1
 
W18 = e 8
x(n) = sin 
21000n 22000n = cos – j sin = 0.707 – j 0.707
 + 0.5 sin  F  4 4
 F   
ic ow

s s –j2 · 2
 
W28 = e 8
x(n) = sin 
2n 22000n = cos – j sin = – j
2 2
 8000 + 0.5 sin  8000 
n
–j2 · 3
3 3
bl kn

n n
x(n) = sin  + 0.5 sin  
W38 = e 8
= cos – j sin
(1) 4 4
4 2
at

= – 0.707 – j 0.707
Pu ch

Since we require a 8 point DFT, we vary n from 0 to 7 in

Equation (1), We write down the equations at each stage and calculate
 
x(0) = sin  · 0 + 0.5 Sin  – 0 = 0 the outputs.
Te

4  2 
Similarly, x(1) = 1.2
x(2) = 1

Fig. P.4.4.5
 DSIP (MU-Sem. 7-COMP) 4-13 Fast Fourier Transform

Stage 1 X(4) = F1(0) – W08F2(0) = 0 – 1(0) = 0

G1(0) = x(0) + x(4) = 0 + 0 = 0 X(5) = F1(1) – W18F2(1)

G1(1) = x(0) – x(4) = 0 – 0 = 0 = – j2 – (0.707 – j0.707)(1.4 – j1.4) = – j0.02

X(6) = F1(2) – W28F2(2) = 0 – (– j) (2) = j2
G2(0) = x(2) + x(6) = 1 + (– 1) = 0
X(7) = F1(3) – W38F2(3)
G2(1) = x(2) – x(6) = 1 – (– 1) = 2
= j2 – (– 0.707 – j0.707) (1.4 + j1.4)
H1(0) = x(1) + x(5) = 1.2 + (– 0.2) = 1 = j0.398

H1(1) = x(1) – x(5) = 1.2 – (– 0.2) = 1.4  X(k) = {0, – j3.98, – 2j, j0.02, 0 – j0.02, j2, j0.398}

Ex. 4.4.6 : An eight point sequence x(n)is given by

H2(0) = x(3) + x(7) = 0.2 + (– 1.2) = – 1
x1(n)= {1, 2, 3, 4, 5, 6, 7, 8}

ge
H2(1) = x(3)–x(7) = 0.2 – (– 1.2) = 1.4 (i) Find DFT of x1(n) using any of the FFT technique
(ii) Let x2(n) = {5, 6, 7, 8, 1, 2, 3, 4}
Stage 2
io led
Using appropriate DFT property and answer of earlier
F1(0) = G1 (0) + W40G2 (0) = 0 + 1 · (0) = 0 part determine X2(k)

(iii) Again we DFT property and Find

ic ow

F1(1) = G1 (1) + W41G2 (1) = 0 + (– j) (2) = – j2

x3(n)=x1(n) + x2(n)
F1(2) = G1 (0)–W04G2 (0) = 0 – 1 · (0) = 0
n
Soln. :
bl kn

(i) Computing the DFT of X1(n):

F1(3) = G1 (1) – W14G2 (1) = 0 – (– j) (2) = + j2
We use the DIT-FFT. Since N = 8, we draw the 8-point
at
Pu ch

F2(0) = H1 (0) + W40H2 (0) = 1 + 1(– 1) = 0 DIT-FFT Butterfly diagram. The input is in the bit reversed
order.
F2(1) = H1 (1) + W41H2 (1)
Te

x1(n) = { 1 2 3 4 5 6 7 8 }
x(0)x(1)x(2)x(3)x(4)x(5)x(6)x(7)
= 1.4 + (– j) (1.4) = 1.4 – j1.4
Here, W04 =1
F2(2) = H1 (0) – W04H2 (0) –j2
 
W14 = e 4
= cos – j sin = – j
2 2
= 1 – 1(– 1) = 2
Similarly,
F2(3) = H1 (1) – W14H2 (1) W08 = 1
–j2· 1
= 1.4 – (– j) (1.4) = 1.4 + j1.4  
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4
Stage 3 –j2· 2
 
W28 = e 8
= cos – j sin = – j
X(0) = F1(0) + W08F2 (0) = 0 + 1 · (0) = 0 2 2
–j2· 3
X(1) = F1(1) + W18F2(1) 3 3
W38 = e 8
= cos – j sin
4 4
= – j2 + (0.707 – j0.707) (1.4– j1.4) = – j3.98 = – 0.707 – j 0.707
X(2) = F1(2) + W28F2(2) = 0 + (– j) (2) = – j 2 We write down the equations at each stage and calculate
the outputs.
X(3) = F1(3) + W38F2(3)

= j2 + (– 0.707 – j0.707) (1.4 + j1.4)

= j0.02
 DSIP (MU-Sem. 7-COMP) 4-14 Fast Fourier Transform

ge
io led Fig. P. 4.4.6

Stage 1 Stage 3

G1(0) = x1(0) + x1(4) = 1 + 5 = 6 X1(0) = F1(0) + W08F2(0) = 16 + 1 · (20) = 36

ic ow

G1(1) = x1(0) – x1(4) = 1 – 5 = – 4 X1(1) = F1(1) + W18F2(1)

n
G2(0) = x1(2) + x1(6) = 3 + 7 = 10 = (– 4 + j4) + (0.707 – j0.707)(–4 + j12)
bl kn

X1(2) = F1(2) + W28F2 (2) = – 4 + (– j)(–4) = – 4 + j4

G2(1) = x1(2) – x1(6) = 3 – 7 = – 4
at
Pu ch

X1(3) = F1(3) + W38F2(3)

H1(0) = x1(1) + x1(5) = 2 + 6 = 8
= (– 4 + j 4) + (– 0.707 – j0.707) (– 4 + j12)
Te

H1(1) = x1(1) – x1(5) = 2 – 6 = – 4

X1(4) = F1(0) – W80F2(0) = 16 – 1 · (20) = – 4
H2(0) = x1(3) + x1 (7) = 4 + 8 = 12
X1(5) = F1(1) – W81F2(1)
H2(1) = x1(3) – x1(7) = 4 – 8 = – 12
= (– 4 + j 4) – (0.707 – j 0.707) (– 4 + j 12)
Stage 2
X1(6) = F1(2) – W82F2(2) = – 4 – (– j) (– 4) = – 4 – j4
F1(0) = G1(0) + W40G2(0) = 6 + 1 · (10) = 16
X1(7) = F1(3) – W83F2(3)
F1(1) = G1(1) + W41G2(1) = – 4 + (– j) (– 4) = – 4 + j 4
= (– 4 – j12) – (– 0.707 – j0.707) (– 4 – j12)
F1(2) = G1(0) – W04G2(0) = 6 – 1 · (10) = – 4
X(k) = {36, – 4 + j9.65, – 4+j4, – 4 + j1.65, – 4,
F1(3) = G1(1) – W14G2(1) = – 4 – (– j) (– 4) = – 4 –j 4 – 4 – j1.65, – 4 – j4, – 4 – j 9.656}

F2(0) = H1(0) + W40H2(0) = 8 + 1 · (12) = 20 (ii) Compute the DFT of x2(n) using appropriate DFT
property.
F2(1) = H1(1) + W41H2(1) = – 4 + (– j)(– 12) = – 4 + j12
Here, x1 (n) = { 1, 2, 3, 4, 5, 6, 7, 8 }
F2(2) = H1(0) – W04H2(0) = 8 – 1 · (12) = – 4 and x2 (n) = { 5, 6, 7, 8, 1, 2, 3, 4 }

F2(3) = H1(1) – W14H2(1) = – 4 – (– j) (– 12) = – 4 – j12 Imagine x1(n) and x2(n) to lie on 2 circles.
 DSIP (MU-Sem. 7-COMP) 4-15 Fast Fourier Transform

Fig. P. 4.4.6(a)

If we rotate the x2(n) circle in the clockwise direction = 4 + j9.65

by 4 steps, we would get the x1(n).circle.  X2(k) = {36, 4 – j9.65, – 4 + j4, 4 – j1.65, – 4,
Hence, x2 (n) is a delayed version of x1(n). 4 + j1.65, – 4 – j4, 4 + j 9.65}
 x2 (n) = x1((n – 4)) (iii) Compute X3 (k) if x3(n)=x1(n) + x2(n).

ge
From the time shifting property of the DFT we know, We use the linearity property which is as follows then
DFT DFT
If x(n) X(k) If x (n) X (k)
io led
N
DFT
DFT
– j2km ax1(n) + b x2(n)  a X1(k) + b X2(k)

Then x((n – m))N e X(k)
N Since,x3(n) = x1(n) + x2(n)
ic ow

i.e.X2(k) = DFT {x2(n)}=DFT { x1 ((n – 4))}

  X3(k) = DFT {x3(n)} = DFT {x1(n) + x2(n)}
n
– j2km
  X3(k) = X1(k) + X2(k)
bl kn

= e X1(k)
  X3 (0) = X1(0) + X2(0) = 36 + 36 = 72
Since the delay is of 4 units, m = 4, N = 8.
  X3(1) = X1(1) + X2(1) = (– 4 + j9.65) + (– 4 – j 9.65)
at


Pu ch

We can write
– j2·k = 0

X2(k) = e X1 (k)   X3(2) = X1(2) + X2(2) = (– 4 + j4) + (– 4 – j 4)
Te

 X2(k) = X1 (k) e –jk = – 8 + j8

We have calculated X1(k) in the previous part of the   X3(3) = X1(3) + X2(3) = (– 4 + j 1.65) + (4 – j1.65) = 0
example. X3(4) = X1(4) + X2(4) = (– 4) + (– 4) = – 8
We vary k from 0 to 7 X3(5) = X1(5) + X2(5) = (– 4 – j 1.65) + (4 + j 1.65)= 0
 X2(0) = e–j·0 X1(0) = (1) · X1(0) =(1).36 = 36 X3(6) = X1(6) + X2(6) = (– 4 – j4) + (– 4 – j4) = – 8 – j8
 X2(1) = e –j·1
X1(1) = (– 1) · X1(1) X3(7) = X1(7) + X2(7) = (– 4 – j9.65) + (4 + j9.65) = 0
= (– 1)(– 4 + j9.65) =4 – j9.65  X3(k) = {72,0, – 8 + j8, 0, – 8, 0, – 8 – j8, 0}
 X2(2) = e –j·2
X1(2) = (1) X1(2) =(1)(– 4 + j4)
Ex. 4.4.7 : Compute the DFT of the sequence
= – 4 + j4 X(0) ={1, 2, 3, 4}.Use DIT-FFT algorithm.
 X2(3) = e –j·3
X1(3) = (– 1) X1(3) Soln. : Since the length of the sequence is 4, N = 4.We use
= (– 1) (– 4 + j1.65) =4 – j1.65 the 4 - point DIT-FFT Butterfly diagram. The input is in the
 X2(4) = e–j·4 X1(4) = (1) X1(4) =(1)(– 4)= – 4 bit reversed order,
j2
  X2(5) = e–j·5 X1(5) = (1) X1(5) = (– 1) (– 4 – j 1.65) ·
W04 = e 4
=1
= 4 + j1.65 -j
·1  
 X2(6) = e–j·6 X1(6) = (1) X1(6) = (1) (– 4 –j4) W14 = e 4
= cos – j sin = – j
2 2
= –4–j4
 X2(7) = e–j·7 X1(7) = (– 1) X1(7) = (– 1) (– 4 – j9.65)
 DSIP (MU-Sem. 7-COMP) 4-16 Fast Fourier Transform

Fig. P. 4.4.7

–j2
Stage 1 1 4
·1  
W4 = e = cos – j sin = – j
G1(0) = x(0) + x(2) = 1 + 3 = 4 2 2

ge
G1(1) = x(0) – x(2) = 1 – 3 = – 2
G1(2) = x(1) + x(3) = 2 + 4 = 6
io led
G1(3) = x(1) – x(3) = 2 – 4 = – 2

Stage 2
ic ow

X(0) = G1(0) + W04G1(2) = 4 + 1(6) = 10

X(1) = G1(1) + W14G1(3) = – 2 + (– j) (– 2)
n
bl kn

= – 2 + j2
X(2) = G1(0) – W04G1(2) = 4 – 1 (6) = – 2 Fig. P. 4.4.8
at
Pu ch

X(3) = G1(1) – W14G1(3) = – 2 – (– j) (– 2) Stage 1

= – 2 – j2 G1 (0) = x(0) + x(2) = 1 + 1 = 2
Te

G1 (1) = x(0) – x(2) = 1 – 1 = 0

X(k) = {10, – 2 + j2, – 2, – 2 – j2 }
All along we have observed that in DIT-FFT, the input gets G1 (2) = x(1) + x(3) = – 1 – 1 = – 2
reordered and the output is in order. G1 (3) = x(1) – x(3) = – 1 – (–1) = 0

Ex. 4.4.8 : Compute 4-point DFT of the sequence given by Stage 2

x(n) = (– 1)n using DIT-FFT algorithms. 0
X (0) = G1 (0) + W4 G1 (2) = 2 + (1) (–2) = 0
Soln. Since N = 4, we vary n from 0 to 3 1
X (1) = G1 (1) + W4 G1 (3) = 0 + (–j) (0) = 0
  x(n) = (– 1) n
0
  x(0) = (– 1)0 = 1 X (2) = G1 (0) + W4 G1 (2) = 2 – (1) (– 2) = 4
1
  x(1) = (– 1)1 = – 1 X (3) = G1 (0) + W4 G1 (3) = 2 – (–j) (0) = 0
  x(2) = (– 1)2 = 1
 X(k) = {0, 0, 4, 0}
  x(3) = (– 1)3 = 1
Ex. 4.4.9 : Compute 4-point DFT of the sequence
  x(n) = {1, 2, 4, 8, 16, 32, 64, 128}
x(n) = {1, 2, 3, 1} using DIT – FFt Radix–2
x(n) = { 1 , – 1 , 1 , – 1 }
algorithms.
x(0) x(1) x(2) x(3) Soln. :
The twiddle factors required are
–j2 Here, N = 4
0 ·0
4
W4 = e =1 x(n) = { 1 , 2 , 3 , 1 }
x(0) x(1) x(2) x(3)
 DSIP (MU-Sem. 7-COMP) 4-17 Fast Fourier Transform

The twiddle factors required are

–j2
0 ·0
4
W4 = e =1
–j2
1 4
·1  
W4 = e = cos – j sin = – j
2 2

Fig. P. 4.4.10

Stage 1

G1(0) = x(0) + x(2) = 0 + 2 = 2

ge
G1(1) = x(0) – x(2) = 0 – 2 = – 2

Fig. P. 4.4.9
io led G1(2) = x(2) + x(3) = 1 + 3 = 4
G1(3) = x(1) – x(3) = 1 – 3 = – 2
Stage 1
G1(0) = x(0) + x(2) = 1 + 3 = 4 Stage 2
0
ic ow

G1(1) = x(0) – x(2) = 1 – 3 = – 2 X (0) = G1 (0) + W4 G1 (2) = 2 + 1 (4) = 6

G1(2) = x(1) + x(3) = 2 + 1 = 3 1
X (1) = G1 (1) + W4 G1 (3)
n
G1(3) = x(1) – x(3) = 2 – 1 = 1
bl kn

= – 2 + (– j) (– 2) = – 2 + j2
Stage 2 0
X (2) = G1 (0) + W4 G1 (2) = 2 – 1 (4) = – 2
at

W04
Pu ch

X(0) = G1(0) + G1(2) = 4 + (1) (3) = 7 1

X (3) = G1 (1) + W4 G1 (3) = – 2 – (–j) (– 2) = – 2 – j2
X(1) = G1(1) + W14 G1(3) = – 2 + (– j) (1) = – 2 – j
X(k) ={6, – 2 + j2, – 2, – 2 – j2}
Te

X(2) = G1(0) – W04 G1(2) = 4 – (1) (3) = 1

X(3) = G1(1) – W14 G1(3) = – 2 – (– j) (1) = – 2 + j 4.5 Decimation in Frequency FFT
(DIT-FFT)
 X(k) = {7, –2 –j, 1, – 2 + j}
Ex. 4.4.10 : Given x(n) = [0, 1, 2, 3].  DIF-FFT is another important FFT-algorithm.
Compute X(k) using DIT – FFT. Decimation in frequency implies retaining the order of

Soln. : the input sequence but getting an output which is

shuffled up.
Here, N = 4
 In DIF-FFT, we split up the input sequence into the first
x(n) = { 0 , 1 , 2 , 3 }
N/2 data points and the second N/2 data points
x(0) x(1) x(2) x(3)
(instead of odd and even as was the case in DIT-FFT).
The twiddle factors required are N–1


–j2 nk
·0 X(k) = x(n) WN
W04 = e 4
= 1 n=0
–j2
·1   N
W14 = e 4
= cos – j sin = – j 2 –1 N–1
2 2
 
nk nk
X(k) = x(n) WN + x(n) WN
We draw a 4-points DIT – FFT Butterfly
n=0 N
n = 2
 DSIP (MU-Sem. 7-COMP) 4-18 Fast Fourier Transform
N N
In the second sum, put n = n + 2 –1
2
 x(n) – x n + 2  WN
N (2k + 1)n
N N X(2k + 1) =
2 –1 2 –1 n=0
+  x n + WN
N

nk (n + N/2) k
X(k) = x(n) WN
 2 ( ··· (– 1)2k + 1 = – 1)
n=0 n=0 2 1
N N Now WN = WN/2
2 –1 2 –1
N
 x n + 2 WN WN WN
N

nk nk nk kN/2
= x(n) WN + 2 –1

 x(n) + x n + 2  WN/2 ;

N nk
n=0 n=0 X(2k) =
N N n=0
2 –1 2 –1
N
x n +
N  nk
 
nk kN/2
= x(n) WN + WN W k = 0, 1, … – 1…(4.5.1)
 2 N 2
n=0 n=0 N
Now 2 –1

 x(n) – x n + 2  WN  WN/2

 N n nK
kN/2 X(2k + 1) =

ge
WN = (– 1)k
n=0
N N N
–1 2 –1 ; k = 0,1,2, … – 1…(4.5.2)
2
2
io led
x n +
N  nk

nk
= x (n) WN + (– 1)k W
  2 N
Let g1(n) = x(n) + x n + N
n=0
n=0   2 
N x(n) – x n + N W n
ic ow

and g2(n) =
2 –1   2  N
 x(n) + (– 1)k x n + 2  WN
N nk
 X(k) = N
n
n=0 2 –1
bl kn


nk
We now split X(k) into odd and even parts,  X(2k) = g1(n) WN/2 …(4.5.3)
n=0
N
at

2 –1 N
Pu ch

2 –1
(2k) =  x(n) + x n + N W2nk ( ··· (– 1)2k = + 1 )
  2  N and X (2k + 1) = 
nk
g2 (n) WN/2 …(4.5.4)
n=0
Te

n=0

We now draw the signal flow graph using

Equations (4.5.1), (4.5.2), (4.5.3) and (4.5.4) using N = 8.

Fig. 4.5.1
 DSIP (MU-Sem. 7-COMP) 4-19 Fast Fourier Transform
As is evident, the order of the input is retained but the N
4 –1
output is shuffled.
 g1(n) + (– 1)k g1 n + 2  WN/2
N nk
Equations (4.5.3) and (4.5.4) represent two N/2 point X(2k) =
DFT’s. Since N = 8, these two equations represent two n=0
4-point DFT’s. Splitting X(2k) into odd and even parts we get,
Each of these can be further split up into two - 2 point N
DFT’s. We will do just that, 4 –1

 g1(n) + g1 n + 4  WN/4 ;

N nk
N X(4k) =
2 –1 n=0

nk
X(2k) = g1(n) WN/2 N
k = 0, 1, … – 1 …(4.5.5)
n=0 4
Splitting X(2k) we get, N
N N 4 –1

 g1(n) – g1 n + 4  WN/2 WN/4 ;

4 –1 4 –1 N n nk
X(4k + 2) =
g1 n + WN/2
N k (n + N/4)
 
nk
X(2k) = g1(n) WN/2 +
 4
n=0
n=0 n=0 N

ge
k = 0, 1, … – 1 …(4.5.6)
N N 4
4 –1 4 –1
g1 n +
N nk We also split X(2K + 1) into two N/4 point DFTs. Using
 
nk kN/4
= g1(n) WN/2
io led
+ WN/2
 W
4  N/2
n=0 n=0
Equations (4.5.5) and (4.5.6), we draw the signal flow
kN/4 graph.
Now, WN/2 = (– 1)k
ic ow

N N
4 –1 4 –1
g1 n +
N nk
 
nk
g1(n) WN/2 + (– 1)k
n
X(2k) = W
 4  N/2
bl kn

n=0 n=0
at
Pu ch
Te

Fig. 4.5.2

A 2-point DFT can be represented as shown in Fig. 4.5.3.

Fig. 4.5.3
 DSIP (MU-Sem. 7-COMP) 4-20 Fast Fourier Transform
The final 8-point DIF-FFT signal flow graph is shown in Fig. 4.5.4.

ge
io led
Fig. 4.5.4 : 8-point DIF-FFT Butterfly diagram

In the DIF-FFT, the output is in the bit reversed format while the input is unaltered.
ic ow

The number of computations required by the DIT-FFT and the DIF-FFT are the same.
As stated earlier, FFT refers to a class of algorithms for efficiently computing the DFT. Hence FFT is not an
n
approximation of the DFT. It is the DFT with a reduced number of computations.
bl kn

4.5.1 Solved Examples on DIT-FFT

at
Pu ch

Ex. 4.5.1 : A 8-Point sequence x(n) is given by {1, 2, 3, 2, 1, 5, 2, 1} find the DFT of the sequence using DIF-FFT.
Soln. : x(n)={1, 2,3, 2, 1, 5, 2, 1}
Te

Since N = 8 we draw a 8-point DIF-FFT Butterfly diagram. Remember the input is ordered but the output obtained is
in bit reversed order.

Fig. P. 4.5.1
 DSIP (MU-Sem. 7-COMP) 4-21 Fast Fourier Transform
–j2
g2(1) = [x(1) – x(5)] W18
Here W04 = e 4
= cos 0 – j sin 0 = 1
–j2 = [2 – 5] (0.707 – j0.707) = – 2.12 + j2.12
 
W14 = e 4
= cos – j sin = – j g2(2) = [x(2) – x(6)] W28= [3 – 2] (– j) = – j
2 2
Similarly, g2(3) = [x(3) – x(7)] W38
–j2
·0 = [2 – 1] (– 0.707 – j0.707) = – 0.707 – j0.707
W08 = e 8
= cos 0 – j sin 0 = 1
Stage 2
–j2
·1   f1(0) = g1 (0) + g1 (2) = 2 + 5 = 7
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4
–j2 f1 (1) = g1 (1) + g1 (3) = 7 + 3 = 10
·2  
W28 = e 8
= cos – j sin = – j f2(0) = [g1 (0) – g1 (2)] W04 = [2 – 5] (1) = – 3
2 2
–j2
·3 3 3 f2(1) = [g1 (1) – g1 (3)] W14= [7 – 3] (– j) = – 4 j
W38 = e 8
= cos – j sin = – 0707 – j0.707
4 4
f3(0) = g2 (0) + g2 (2) = 0 + (– j) = – j

ge
Stage 1
f3(1) = g2 (1) + g2 (3)
g1(0) = x(0) + x(4) = 1 + 1 = 2
= (– 2.12 + j2.12) + (– 0.707 – j0.707)
g1(1) = x(1) + x(5) = 2 + 5 = 7
io led
= – 2.827 + j1.413
g1(2) = x(2) + x(6) = 3 + 2 = 5
f4(0) = [g2 (0) – g2 (2)]W04= 0 – (– j)
g1(3) = x(3) + x(7) = 2 + 1 = 3
ic ow

= 0 – (– j) (1) = + j
g2(0) = [x(0) – x(4)] W08= [1 – 1] (1) = 0
f4(1) = [g2(1) – g2 (3)]W14
n
bl kn

= [(– 2.12 + j2.12) – (– 0.707 – j0.707)] (– j)

= 2.827 + j1.413
at
Pu ch

Stage 3
X(0) = f1(0) + f1(1) = 7 + 10 = 17
Te

X(4) = f1(0) – f1(1) = 7 – 10 = – 3

X(2) = f2(0) + f2(1) = – 3 + (– 4j) = – 3 – 4j
X(6) = f2(0) – f2(1) = – 3 – (– 4j) = – 3 + 4j
X(1) = f3(0) + f3(1) = (– j) + (– 2.827 + j1.413) = – 2.827 + j0.413
X(5) = f3(0) – f3(1) = (– j) – (– 2.827 + j 1.413) = 2.827 – j 2.413
X(3) = f4(0) + f4(1) = j + (2.827 + j1.413)

= 2.827 + 2.413
X(7) = f4(0) – f4(1) = j – (2827 + j1413)

= 2.827 – j0.413
We now rearrange the output in the proper order
 X(k) = {17, – 2.827 + j 0.413, – 3 – 4j, 2.827
+ j2.413, – 3, 2.827 – j2.413, – 3 + 4j, – 2.827 – j0.413}
Ex. 4.5.2 : Find 8-point DFT using Radix – 2 – DIF FFT algorithms for a given sequences
x(n) = {–1, 0, 2, 0, – 4, 0, 2, 0}
Soln. :
x(n) = { – 1, 0 , 2 , 0 , – 4, 0 , 2 , 0 }
x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)
 DSIP (MU-Sem. 7-COMP) 4-22 Fast Fourier Transform
We draw a 8-point DIF-FFT Butterfly diagram,

ge
io led
ic ow

Fig. P. 4.5.2

Here, g2(3) = [x(3) – x(7)] W38 = [0 – 0] (– 0.707 – j0.707) = 0

n
–j2
bl kn

·0 Stage 2
W04 = e 4
= cos 0 – j sin 0 = 1
f1(0) = g1 (0) + g1 (2) = – 5 + 4 = – 1
–j2
 
at

·1
Pu ch

W14 = e 4
= cos – j sin = – j f1 (1) = g1 (1) + g1 (3) = 0 + 0 = 0
2 2
Similarly, f2(0) = [g1 (0) – g1 (2)] W04 = [– 5 – 4] (1) = – 9
Te

–j2
·0 f2(1) = [g1 (1) – g1 (3)] W14 = [0 – 0] (– j) = 0
W08 = e 8
= cos 0 – j sin 0 = 1
f3(0) = g2 (0) + g2 (2) = 3 + 0 = 3
–j2
·1  
W18 = e 8
= cos – j sin = 0.707 – j0.707 f3(1) = g2 (1) + g2 (3) = 0 + 0 = 0
4 4
–j2 f4(0) = [g2 (0) – g2 (2)] W04 = [3 – 0] (1) = 3
·2  
W28 = e 8
= cos – j sin = – j
2 2 f4(1) = [g2(1) – g2 (3)] W41 = [0 – 0] (– j) = 0
–j2
·3 3 3
W38 = e 8
= cos – j sin = – 0.707 – j 0.707 Stage 3
4 4
X(0) = f1(0) + f1(1) = – 1 + 0 = – 1
Stage 1
X(4) = f1(0) – f1(1) = – 1 – 0 = – 1
g1(0) = x(0) + x(4) = – 1 + (– 4) = – 5
X(2) = f2(0) + f2(1) = – 9 + 0 = – 9
g1(1) = x(1) + x(5) = 0 + 0 = 0
X(6) = f2(0) – f2(1) = – 9 – 0 = – 9
g1(2) = x(2) + x(6) = 2 + 2 = 4
X(1) = f3(0) + f3(1) = 3 + 0 = 3
g1(3) = x(3) + x(7) = 0 + 0 = 0
X(5) = f3(0) – f3(1) = 3 – 0 = 3
g2(0) = [x(0) – x(4)] W08 = [– 1 – (– 4)] (1) = 3
X(3) = f4(0) + f4(1) = 3 + 0 = 3
g2(1) = [x(1) – x(5)] W18
X(7) = f4(0) – f4(1) = 3 – 0 = 3
= [0 – 0] (0.707 – j0.707) = 0
We now arrange the output in the proper order
g2(2) = [x(2) – x(6)] W28 = [2 – 2] (– j) = 0
 X(k) = { – 1, 3, – 9, 3, – 1, 3, – 9, 3 }
 DSIP (MU-Sem. 7-COMP) 4-23 Fast Fourier Transform

Ex. 4.5.3 : Given x(n) = n + 1 and N = 8, find DFT X(k) using DIF-FFT algorithm.
Soln. :
Given : x(n) = n + 1 ; 0  n  7 (˙.˙ N = 8)
x(n) = {1, 2, 3, 4, 5, 6, 7, 8}
We draw a 8 – point DIF-FFT Butterfly diagram. In DIF-FFT, the input is ordered but the output obtained is in bit
reversed order.
–j2
·0
Here W04 = e 4
= cos 0 – j sin 0 = 1
–j2
·1  
W14 = e 4
= cos – j sin = – j
2 2
–j2
·0
Similarly, W08 = e 8
= cos 0 – j sin 0 = 1
–j2
 

ge
·1
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4
–j2
·2  
W28 = e 8
io led = cos
2
– j sin = – j
2
–j2
·3 3 3
W38 = e 8
= cos – j sin = – 0.707 – j 0.707
4 4
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 4.5.3

Stage 1: g2(1) = [x(1) – x(5)] W81= [2 – 6] (0.707 – j0.707)

g1(0) = x(0) + x(4) = 1 + 5 = 6
= – 2.83 + j2.83
g1(1) = x(1) + x(5) = 2 + 6 = 8 g2(0) = [x(2) – x(6)] W82= [3 – 7] (– j) = + 4j
g1(2) = x(2) + x(6) = 3 + 7 = 10 g2(3) = [x(3) – x(7)] W83= [4 – 8] (– 0.707 – j0.707)
g1(3) = x(3) + x(7) = 4 + 8 = 12
= 2.83 + j2.83
g2(0) = [x(0) – x(4)] W80= [1 – 5] (1) = – 4
 DSIP (MU-Sem. 7-COMP) 4-24 Fast Fourier Transform

Stage 2 : X(4) = f1(0) – f1(1) = 16 – 20 = – 4

f1(0) = g1(0) + g1(2) = 6 + 10 = 16 X(2) = f2(0) + f2(1) = – 4 + 4 j
f1(1) = g1(1) + g1(3) = 8 + 12 = 20 X(6) = f2(0) – f2(1) = – 4 – 4j
f2(0) = [g1(0) – g1(2)] W04= [6 – 10] (1) = – 4 X(1) = f3(0) + f3(1) = (– 4 + 4j) + (j5.66)

f2(1) = [g1(1) – g1(3)] W14 = [8 – 12] (– j) = + 4 j = – 4 + j9.66

f3(0) = g2(0) + g2(2) = – 4 + 4 j X(5) = f3(0) – f3(1) = (– 4 + 4j) – (j5.66)

f3(1) = g2(1) + g2(3) = – 4 – j1.66

X(3) = f4(0) + f4(1) = (– 4 – 4j) + (j5.66)
= (– 2.83 + j2.83) + (2.83 + j2.83)
= j5.66 = – 4 + j1.66
X(7) = f4(0) – f4(1) = (– 4 – 4j) – (j5.66)
f4(0) = [g2(0) – g2(2)]W04 = [ – 4 – (+ 4j)]
= – 4 – j9.66
= – 4 – 4j

ge
We now rearrange the output in the proper order
f4(1) = [g2(1) – g2(3)]W14
 X(k) = { 36, – 4 + j9.66, – 4 + 4j, – 4 + 1.66j, – 4,
= (– 2.83 + j2.83) (2.83 + j2.83) (– j) = j5.66
io led – 4 – j1.66, – 4 – 4 j, – 4 – j9.66}
Stage 3
X(0) = f1(0) + f1(1) = 16 + 20 = 36
ic ow

Ex. 4.5.4 : Compute DFT of sequence x(n) = {1, 2, 2, 2, 1, 0, 0, 0} using DIF-FFT algorithm. Sketch its magnitude
spectrum.
n
bl kn
at
Pu ch
Te

Fig. P. 4.5.4

Soln. :
x(n) = {1, 2, 2, 2, 1, 0, 0, 0}
We draw a 8-point DIF-FFT Butterfly diagram in DIF-FFT, the input is ordered but the output obtained in bit reversed
order.
 DSIP (MU-Sem. 7-COMP) 4-25 Fast Fourier Transform

Here, f4(1) = [g2(1) – g2(3)]W14

–j2
·0
W04 = e 4
= cos 0 – j sin 0 = 1 = (1.414 – j1.414) – (– 1.414 – j1.414) (– j)
–j2 = – j2.83
·1  
W14 = e 4
= cos – j sin = – j
2 2 Stage 3
Similarly, X(0) = f1(0) + f1(1) = 4 + 4 = 8
–j2
·0 X(4) = f1(0) – f1(1) = 4 – 4 = 0
W08 = e 8
= cos 0 – j sin 0 = 1
–j2
X(2) = f2(0) + f2(1) = 0 + 0 = 0
·1  
W18 = e 4
= cos – j sin = 0.707 – j 0.707 X(6) = f2(0) – f2(1) = 0 – 0 = 0
4 4
–j2 X(1) = f3(0) + f3(1) = – j2 + (– j2.83) = – j4.83
·2  
W28 = e 4
= cos – j sin = – j
2 2 X(5) = f3(0) – f3(1) = – j2 – (– j2.83) = + j0.83
–j2
·3 3 3 X(3) = f4(0) + f4(1) = j2 + (– j2.83) = – j0.83
W38 = e 4

ge
= cos – j sin = – 0.707 – j0.707
4 4
X(7) = f4(0) – f4(1) = j2 – (– j2.83) = j4.83
Stage 1
io led We now rearrange the output in the proper order.
g1(0) = x(0) + x(4) = 1 + 1 = 2
X(k) = {8, – j4.83, 0, – j0.83, 0, + j0.83, 0, j4.83}
g1(1) = x(1) + x(5) = 2 + 0 = 2
To sketch the magnitude spectrum. We compute the
ic ow

g1(2) = x(2) + x(6) = 2 + 0 = 2 magnitude of the Fourier coefficient X(k)

g1(3) = x(3) + x(7) = 2 + 0 = 2  X(k) ={8, 4.83, 0, 0.83, 0, 0.83, 0, 4.83}

n
bl kn

g2(0) = [x(0) – x(4)] W80 = [1 – 1] (1) = 0

g2(1) = [x(1) – x(5)] W81 = [2 – 0] (0.707 – j0.707)

at
Pu ch

= 1.414 – j1.414
g2(2) = [x(2) – x(6)] W82 = [2 – 0] (– j) = – 2j
Te

g2 (3) = [x(3) – x(7)] W83 = [2 – 0] (– 0.707 – j0.707)

= 1.414 – j1.414
Stage 2 Fig. P. 4.5.4(a)
f1(0) = g1(0) + g1(2) = 2 + 2 = 4 We observer that the magnitude spectrum is symmetric
f1(1) = g1(1) + g1(3) = 2 + 2 = 4 
about the point, in this case 4.
2
f2(0) = [g1(0) – g1(2)] W04 = [2 – 2] (1) = 0
Ex. 4.5.5 : N = 8 point DFT using Radix FFT for the given
f2(1) = [g1(1) – g1(3)] W14 = [2 – 2] (– j) = 0 sequence x(n) = {1, 1, 1, 1, 0, 0, 0, 0}

f3(0) = g2(0) + g2(2) = 0 + (– j2) = – j2 Soln. :

f3(1) = g2(1) + g2(3) x(n) = { 1 , 1 , 1 , 1 , 0 , 0 , 0 , 0 }
x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)
= (1.414 – j1.414) + (– 1.414 – j1.414)
We draw a 8-point DIF-FFT Butterfly diagram
= – j2.83
f4(0) = [g2(0) – g2(2)]W04 = 0 – (– j2) = + j2
 DSIP (MU-Sem. 7-COMP) 4-26 Fast Fourier Transform

ge
io led
Fig. P. 4.5.5
ic ow

Here, g2(2) = [x(2) – x(6)] W28 = [1 – 0] (– j) = – j

n
bl kn

–j2
·0 g2(3) = [x(3) – x(7)] W38
W04 = e 4
= cos 0 – j sin 0 = 1
= [1 – 0] (– 0.707 – j0.707)
at

–j2
·1  
Pu ch

W14 = e 4
= cos – j sin = – j = – 0.707 – j0.707
2 2
Stage 2
Similarly,
Te

–j2 f1(0) = g1 (0) + g1 (2) = 1 + 1 = 2

·0
W08 = e 8
= cos 0 – j sin 0 = 1
f1 (1) = g1 (1) + g1 (3) = 1 + 1 = 2
–j2
·1   f2(0) = [g1 (0) – g1 (2)] W04 = [1 – 1] (1) = 0
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4
–j2 f2(1) = [g1 (1) – g1 (3)] W14 = [1 – 1] (– j) = 0
·2  
W28 = e 8
= cos – j sin = – j
2 2 f3(0) = g2 (0) + g2 (2) = 1 + (– j) = 1– j
–j2
·3 3 3 f3(1) = g2 (1) + g2 (3)
W38 = e 8
= cos – j sin = – 0.707 – j 0.707
4 4
= (0.707 – j0.707) + (– 0.707 – j0.707)
Stage 1
g1(0) = x(0) + x(4) = 1 + 0 = 1 = – j1.414
g1(1) = x(1) + x(5) = 1 + 0 = 1 f4(0) = [g2 (0) – g2 (2)]W04 = [1 – (–j)] (1) = 1 + j
g1(2) = x(2) + x(6) = 1 + 0 = 1 f4(1) = [g2(1) – g2 (3)] W41
g1(3) = x(3) + x(7) = 1 + 0 = 1
= [(0.707 – j0.707) – (– 0.707 – j0.707)] (– j)
g2(0) = [x(0) – x(4)] W08 = [1 – 0] (0) = 1
= – j1.414
g2(1) = [x(1) – x(5)] W18
Stage 3
= [1 – 0] (0.707 – j0.707) X(0) = f1(0) + f1(1) = 2 + 2 = 4
= 0.707 – j0.707 X(4) = f1(0) – f1(1) = 2 – 2 = 0
 DSIP (MU-Sem. 7-COMP) 4-27 Fast Fourier Transform
X(2) = f2(0) + f2(1) = 0 + 0 = 0 Soln. :We draw a 8-point DIF-FFT Butterfly diagram. In
X(6) = f2(0) – f2(1) = 0 – 0 = 0 DIF-FFT, the input is ordered but the output obtained is in
bit reversed order.
X(1) = f3(0) + f3(1) = (1 – j) + (– j1.414) = 1 – j2.414 –j2
·0
X(5) = f3(0) – f3(1) = (1 – j) – (– j0.414) = 1 + j0.414 Here W04 = e 4
= cos 0 – j sin 0 = 1
–j2
X(3) = f4(0) + f4(1) = (1 + j) + (– j0.414) = 1 – j0.414 ·1  
W14 = e 4
= cos – j sin = – j
2 2
X(7) = f4(0) – f4(1) = (1 + j) – (1 – j1.414) = 1 + j2.414
Similarly,
We now rearrange the output in the proper order –j2
·0
W08 = e 8
= cos 0 – j sin 0 = 1
 X(k) = {4 , 1 – j2.414, 0, 1 – j0.414, 0, 1
–j2
+ j0.414, 0, 1 + j2.414} ·1  
W18 = e 8
= cos – j sin = –0.707 – j 0.707
4 4
Ex. 4.5.6 : Using DIF - FFT Find a –j2
x(n) = {1, 2, 1, 2, 4, 2, 1, 2}
·2  
W28 8

ge
= e = cos – j sin = – j
2 2
If x1(n) = x(– n), find x1(k) using earlier solution and not
–j2
·3 3 3
otherwise. W38 = e 8
= cos – j sin = – 0.707 – j0.707
io led 4 4
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 4.5.6
Stage 1
g1(0) = x(0) + x(4) = 1 + 4 = 5
g1(1) = x(1) + x(5) = 2 + 2 = 4
g1(2) = x(2) + x(6) = 1 + 1 = 2
g1(3) = x(3) + x(7) = 2 + 2 = 4
g2 (0) = [x(0) – x(4)] W08= [1 – 4] (1) = – 3
g2 (1) = [x(1) – x(5)] W18= [2 – 2] (0.707 – j0.707) = 0
g2 (2) = [x(2) – x(6)] W28= [1 – 1] (– j) = 0
g2 (3) = [x(3) – x(7)] W38= [2 – 2] (– 0.707 – j0.707) = 0
 DSIP (MU-Sem. 7-COMP) 4-28 Fast Fourier Transform

Stage 2 We draw a 4 – point DIF–FFT butterfly diagram. In

f1(0) = g1(0) + g1(2) = 5 + 2 = 7 DIF–FFT, the input is ordered but the output obtained is in
f1(1) = g1(1) + g1(3) = 4 + 4 = 8 the bit reversed order.
–j2
·0
f2(0) = [g1(0) – g1(2)] W04 = [5 – 2] (1) = 3 Here W40 = e 4
= cos 0 – j sin 0 = 0
–j2
f2(1) = [g1(1) – g1(3)] W14 = [4 – 4] (– j) = 0 ·1  
W14 = e 4
= cos – j sin = – j
2 2
f3(0) = g2(0) + g2(2) = – [ – 3 + 0] = – 3
f3(1) = g2(1) + g2(3) = [0 + 0] = 0

f4(0) = [g2(0) – g2(2)]W04 = [ – 3 – 0] (1) = – 3

f4(1) = [g2(1) – g2(3)]W14 = [0 – 0] (– j) =0

Stage 3

ge
X(0) = f1(0) + f1(1) = 7 + 8 = 15
X(4) = f1(0) – f1(1) = 7 – 8 = – 1
io led
X(2) = f2(0) + f2(1) = 3 + 0 = 3
Fig. P. 4.5.7
X(6) = f2(0) – f2(1) = 3 – 0 = 3
Stage 1
ic ow

X(1) = f3(0) + f3(1) = – 3 + 0 = – 3 g1(0) = x(0) + x(2) = 2 + 4 = 6

X(5) = f3(0) – f3(1) = – 3 – 0 = 3 g1(1) = x(1) + x(3) = 3 + 5 = 8
n
g1(2) = [x(0) – x(2)] W40= [2 – 4] (1) = – 2
bl kn

X(3) = f4(0) + f4(1) = – 3 + 0 = – 3

X(7) = f4(0) – f4(1) = – 3 – 0 = – 3 g1(3) = [x(1) – x(3)] W41= [3 – 5] (– j) = + 2j
at
Pu ch

Stage 2
We now rearrange the output in the proper order.
X(0) = g1(0) + g1(1) = 6 + 8 = 14
X(k) = {15, – 3, 3, – 3, – 1, – 3, – 3, 3, – 3}
X(2) = g1(0) – g1(1) = 6 – 8 = – 2
Te

We now need to find X1(k).

X(1) = g1(2) + g1(3) = – 2 + (+ 2j) = – 2 + 2j
Given x1(n) = x(– n)
X(3) = g1(2) – g1(3) = – 2 – (+ 2j) = – 2 – 2j
According to the time reversal properly of the DFT
We rearrange the output in the proper order
we have
 X(k) = {14, – 2 + 2j, – 2, – 2 – 2j}
x1((–n))N 
DFT
X((– k))n n
Ex. 4.5.8 : Compute the DFT of x(n) = cos 2 using
 x1(n) = x((– n))N X((– k))N
DFT
DIF-FFT. Assume N = 4.
= X1(– k) n
Soln. : Since N = 4, we have x(n) = cos ;0n3
Hence X1(k) is obtained by circularly folding X1(k) 2

X1(k) = {15, – 3, 3, –3, – 1, – 3, 3, – 3} x(0) = cos (0) = 1


X1(k) is the same as X(k). x(1) = cos  = 0
 2
Ex. 4.5.7 : Compute the DFT of
x(2) = cos () = – 1

x(3) = cos  = 0
3
x(n) = 2(n) + 3(n – 1) + 4(n – 2) + 5(n – 3)
2
Use DIF-FFT
 x(n) = {1, 0, – 1, 0}
Soln. : Here (n) is a impulse functions while (n – 1),
(n – 2) and (n – 3) are impulse shifted to the right.
  x(n) = {2, 3, 4, 5}
 DSIP (MU-Sem. 7-COMP) 4-29 Fast Fourier Transform

We draw a 4 - point DIF-FFT Butterfly diagram X(1) = g1(2) + g1(3) = 2 + 0 = 2

X(2) = g1(2) – g1(3) = 2 – 0 = 2

We rearrange the output in the proper order

 X(k) = {0, 2, 0, 2}
Ex. 4.5.9 : If x(n) = {1 + 2j, 3 + 4j, 5 + 6j, 7 + 8j}. Find DFT
X(k) using DIF-FFT.
Soln.: Since N = 4, we draw a 4-point DIF-FFT Bufferfly
diagram in DIF-FFT, the input is ordered but the output is in
Fig. P. 4.5.8 bit reversed order.
–j2 –j2
·0
Here W40 = e 4
=1 0 ·0
Here W4 = e 4 = 1
-j2
 
W14 = e 4

ge
= cos – j sin = – j –j2
2 2 1 ·1  
W4 = e 4 = cos – j sin = – j
Stage 1
io led 2 2
g1(0) = x(0) + x(2) = 1 – 1 = 0 Stage 1
g1(1) = x(1) + x(3) = 0 + 0 = 0 g1(0) = x(0) + x(2) = (1 + 2j) + (5 + 6j) = 6 + j8
ic ow

g1(2) = [x(0) – x(2)] W04 = [1 – (– 1)] (1) = 2 g1(1) = x(1) + x(3) = (3 + j4) + (7 + j8) = 10 + j12
g2(2) = [x(0) – x(2)] W40= [(1 + 2j) – (5 + 6j)] (1)
g1(3) = [x1(1) – x(3)] W14 = [0 – 0] (– j) = 0
n
bl kn

= – 4 – j4
Stage 2
X(0) = g1(0) + g1(1) = 0 + 0 = 0 g1(3) = [x(1) – x(3)] W81
at
Pu ch

X(2) = g1(0) – g1(1) = 0 – 0 = 0 = [(3 + j4) – (7 + j8)] (– j) = – 4 + j4

Te

Fig. P. 4.5.9

Stage 2
X(0) = g1(0) + g1(1) = (6 + j8) + (10 + j12) = 16 + j20
X(2) = g1(0) – g1(1) – (6 + j8) – (10 – j12) = – 4 – j4
X(1) = g1(2) + g1(3) = (– 4 – j4) + (– 4 + j4) = – 8
X(3) = g1(2) – g1(3) = (– 4 – j4) + (– 4 + j4) = – j8

We rearrange the output in the proper order

 X(k) = {16 + j20, – 8, – 4 – j4, – j8}
 DSIP (MU-Sem. 7-COMP) 4-30 Fast Fourier Transform

 Three such stages are required to compute 8–point

4.6 Computational Complexity DFT. In general, for ‘N’ point DFT, log2N stages are
Comparison between DFT and FFT
required.

 We now compare the number of computations Complex multiplications

required for DFT and FFT. N
 At each stage there are 2 butterflies and the total
 It was mentioned earlier that FFT is merely a faster
way of computing the DFT. We will now show how number of stages are log2N. For each butterfly, one
much is the reduction in computation when a FFT complex multiplication is required.
algorithm is used. N
 ence total complex multiplications = 2 log2 N…(4.6.3)
1. Computations required for DFT algorithm
Complex additions
We have shown at the beginning of the chapter that a
N-point DFT required N2 multiplication and N(N – 1) N

ge
 Total number of stages are log2 N and at each stage, 2
additions.
butterflies are required. For each butterfly, ‘2’
2. Computations required using FFT algorithm
io led complex additions are required.
 We begin by calculating the computation complexity 
N
Total complex additions = 2  log2 N
2
required for one butterfly. Consider the general
ic ow

structure of butterfly as shown in Fig. 4.6.1.   Total complex additions = N log2 N

 Table 4.6.1 shows comparison of direct DFT

n
bl kn

computation and computation using FFT algorithms.

Table 4.6.1
at
Pu ch

Direct Computation Using FFT

Number
Te

Fig. 4.6.1 : General structure of butterfly of Complex Complex Complex Complex

points Multiplication Additions Multiplication Addition
 Here ‘a’ and ‘b’ are inputs and A and B are outputs of
(N 2 – N) (N log2 N)
butterfly. The outputs are given by, ‘N’
N2  N log2 N 
2 
A = a + WNb …(4.6.1)
and B = a – WNb  …(4.6.2) 4 16 12 2 4
 To calculate any output (A or B), we need to multiply
each input by the twiddle factor WN. So one complex 8 64 56 12 24

multiplication is required for one butterfly.

16 256 240 32 64
 From Equations (4.6.1) and (4.6.2) we note that to
calculate output ‘A’, one complex addition is required, 32 1024 992 40 80
while to calculate output ‘B’ one complex subtraction
is required. But the computational complexity of 64 4096 4032 96 192
addition and subtraction is same. So we can say that
for one butterfly two complex additions are required. 128 16,384 16,256 224 448

 For a 8 point DFT, 4 butterflies are required at each The Table 4.6.1 shows that, by the use of FFT
'N' algorithms the number of complex multiplications and
stage. So for ‘N’ point DFT, at each stage, 2
complex additions are reduced. `So there is tremendous
butterflies are required. improvement in the speed
 DSIP (MU-Sem. 7-COMP) 4-31 Fast Fourier Transform
4.6.1 Solved Examples on Computation of
Inverse DFT using FFT Algorithms

Ex. 4.6.1 : How many computations are required to compute

16 point DFT using DFT and FFT algorithms

Soln. :
Here, N = 16
Using DFT
Fig. P. 4.6.2
Number of multiplications = N2 = 162 = 256
Number of addition = N(N – 1) = 16  15 = 240 Ex. 4.6.3 : Find the I-DFT using DIT-FFT of the given
Using FFT example.
N X(k) = {10, – 2 + 2j, – 2, – 2 – 2j }
Number of Multiplication = logN = 32
2 2
Soln. :

ge
Number of Addition = N logN
2
= 64

Ex. 4.6.2 : How to find inverse one dimensional DFT using

io led
forward DIT FFT flow graph. Derive necessary formula.

Soln. : The 1-D-DFT formula is

ic ow

N–1
X(k) =  x(n) Wnk
N ...(1)
n
k=0
bl kn

Fig. P. 4.6.3
The 1-D-IDFT formula is
N–1 y(0) = X(0) + X(2) = 8
at

1

Pu ch

WN* nk
y(1) = X(0) – X(2) = 12
x(n) = X(k) ...(2)
N
k=0
y(2) = X(1) + X(3) = – 4
Te

The DIT-FFT algorithm is computed using

Equation (1). We now try to make Equation (2) appear like y(3) = X(1) – X(3) = – 4j
1
Equation (1) x(0) = [Y(0) + Y(2)  (1) ] = 1
4
We take the complex conjugate on both sides of
1
Equation (2). x(1) = [Y(1) + Y(3)  (– j) ] = 2
4
N–1
1
 x(n) =
1
 X(k) W** nk x(2) = [Y(0) + Y(2) (– 1) ] = 3
N N 4
k=0
1
N–1 x(3) = [Y(1) + Y(3)  (j) ] = 4
1 4
 x(n) =
N  X(k) Wnk
N ...(3)
 x(n) = {1, 2, 3, 4}
k=0
We finally take the complex conjugate
Equation (3) is identical to Equation (1) except for the
1 x*(n) = {1, 2, 3, 4}
complex sign and the term.
N
We have thus found out the inverse DFT using the
Hence in DIT-FFT, if we replace the input x(n) by X(k) DIT-FFT.
1
and divide it by , we get x(n). Taking the conjugate of Ex. 4.6.4 : Given :
N
this gives us x(n). Steps involved in computing DIT-IFFT. X(k) = {36, – 4 + j9.656, – 4 + j4, – 4 + j1.656, – 4, – 4
– j1.656, – 4 – j4, – 4 – j9.656}
1. Replace x(n) by X(k) in the DIT FFT.
Find x(n) using any IFFT algorithm.
2. On obtaining x(n), take the conjugate.
 DSIP (MU-Sem. 7-COMP) 4-32 Fast Fourier Transform
Soln. : Let us use the DIF-FFT algorithm to compute the Stage 1
inverse FFT. g1(0) = X*(0) + X*(4) = 36 + (– 4) = 32
Here, N = 8 g1(1) = X*(1) + X*(5)
We draw a 8 point DIF-FFT butterfly diagram and = (– 4 – j9.656) + (– 4 + j1.651)
make the required for computation of the inverse FFT. In
= – 8 – j8.005
required DIF-FFT, the input is in ordered but the output is in
g1(2) = X*(2) + X*(6) = (– 4 – j4) + (– 4 + j4)
bit reversed order.
= –8
Here,
–j2 g1(3) = X*(3) + X*(7)
·0
W04 = e 4
=cos 0 – j sin 0 = 1 = (– 4 – j1.656) + (– 4 + j9.656)]
–j2
·1   = – 8 + j8
W14 = e 4
=cos – j sin = – j
2 2
g2(0) = X*[(0) – X*(4)] W08 = 36 – (– 4) (1) = 40
Similarly,

ge
–j2
g2(1) = X*[(1) – X*(5)] W18
·0
W08 = e 8
= cos 0 – j sin 0 = 1
io led = [(– 4 – j9.656) – (– 4 + j 1.651)] (0.707 – j0.707)
–j2
·1   = –8–j8
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4 g2(2) = [ X*[(2) – X*(6)] W82
–j2
·2  
ic ow

W28 = e 8
=cos – j sin = – j = [(– 4 – j4) – (– 4 + j 4)] (– j)] = – 8
2 2
–j2
g2(3) = X*[(3) – X*(7)] W38
·3 3 3
n
W38 = e 8
=cos – j sin = [(– 4 –j1.656) – (– 4 + j9.656)] (– 0.707 – 0.707)
4 4
bl kn

= – 0.707 – j 0.707 = – 8 + j8
at

Note: The input given to the network is the complex

Pu ch

conjugate of the input and the output obtained needs

to be divided by N.
Te

Fig. P. 4.6.4
 DSIP (MU-Sem. 7-COMP) 4-33 Fast Fourier Transform

Stage 2 x*(6) = [f2(0) – f2(1)]/8 = [40 – (– 16)]/8 = 7

f1(0) = g1(0) + g1(2) = 32 + (– 8) = 24 x*(1) = [f3(0) + f3(1)]/8 = [32 + (– 16)]/8 = 2

f1(1) = g1(1) + g1(3) = (– 8 – j8) + (– 8 + j8) x*(5) = [f3(0) – f3(1)]/8 = [32 – (– 16)]/8 =6

= – 16 x*(3) = [f4(0) + f4(1)]/8 = [48 + (– 16)]/8 = 4

f2(0) = [g1(0) + g1(2)] W04= [32 – (– 8)](1) = 40 x*(7) = [f4(0) – f4(1)]/8 = [48 – (– 16)]/8 = 8

f2(1) = [g1(1) – g1(3)] W14 We rearrange the output in the proper order. Since x*(n)
values are all real
= [(– 8 –j8) – (– 8 + j8) ](– j) = – 16
x(n) = x* (n)
f3(0) = g2(0) + g2(2) = 40 + (– 8) = 32
x(n) = {1, 2, 3, 4, 5, 6, 7, 8}
f3(1) = g2(1) + g2(3) = (– 8 – j8) + (– 8 + j8)
Ex. 4.6.5 : Let x(n) be 8-point sequence. Its
= – 16 corresponding DFT X(k) is,

ge
f4(0) = [g2(0) – g2(2)]W04 = [40 – (– 8)] (1) = 48 X(k) = { 0.5, 2 + j, 3 + j2, j, 3, - j, 3 – j2, 2 – j }
f4(1) = [g2(1) – g2(3)]W14io led Soln. :

= [(– 8 – j8) – (– 8 + 8)](– j) = – 16 We will use the DIT – FFT algorithms since N = 8, we use
Stage 3 the DIT – FFT algorithms we give X*(k) as the input and
divide the output obtained by N. We also take the complex
ic ow

x*(0) = [f1(0) + f1(1)]/ 8 = [24 + (– 16)]/8 = 1

x*(4) = [f1(0) – f1(1)]/8 = [24 – (– 16)] / 8 = 5 conjugate of the result obtained.
n
x(n) = { 0.5, 2 + j ,3 +j2, j , 3 , – j , 3 – j2, 2 – j}
x*(2) = [f2(0) + f2(1)]/8 = [40 + (– 16)]/ 8 = 3
bl kn

x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)

at
Pu ch
Te

Fig. P. 4.6.5

Here, W04 = 1
–j2
·1  
W14 = e 4
= cos – j sin = – j
2 2
Similarly, W08 = 1
–j2
·1  
W18 = e 8
= cos – j sin = 0.707 – j0.707
4 4
 DSIP (MU-Sem. 7-COMP) 4-34 Fast Fourier Transform
–j2
·2   F2(3) = H1(1) – W14 H2 (1)
W28 = e 8
= cos – j sin = – j
2 2
–j2 = (2 – j2) – (– j) (– 2 – j2)
·3 3 3
W38 = e 8
= cos – j sin = – 0.707 – j 0.707 = 4 – j4
4 4
We write down the equation at each stage and Stage 3
calculate the outputs x*(0) = [F1(0) + W08 F2(0)] /8 = [9.5 + (1) 4] /8 = 1.69
Stage 1
x*(1) = [F1(1) + W18 F2(1)] /8
G1(0) = X*(0) + X*(4)
= [– 6.5 + (0.707 – j0.707 (0)] /8 = – 0.81
= 0.5 + 3 = 3.5 x*(2) = [F1(2) + W28 F2(2)] /8
G1(1) = X*(0) – X*(4)
= [– 2.5 + (– j) (0)] /8 = – 0.31
= 0.5 – 3 = – 2.5 x*(3) = [F1(3) + W38 F2(3)] /8

ge
G2(0) = X*(2) + X*(6)
= [1.5 + (– 0.707 – j0.707) (4 – j4) ]/8 = – 0.52
= (3 – j2) + (3 + j2) = 6
x*(4) = [F1(0) – W08 F2(0)] /8 = [9.5 – (1) 4] /8 = 0.69
io led
G2(1) = X*(2) – X*(6)
x*(5) = [F1(1) – W18 F2(1)] /8
= (3 – j2) – (3 + j2) – j4
= [– 6.5 – (0.707 – j0.707) (0)]/8 = – 0.81
ic ow

H1(0) = X*(1) + X*(5)

x*(6) = [F1(2) – W28 F2(2)] /8
n
= (2 – j) + (j) = 2
= [– 2.5 – (– j) (0)] /8 = – 0.31
bl kn

H1(1) = X*(1) – X*(5) = (2 – j) – (j) = 2 – j2

x*(7) = [F1(3) – W38 F2(3)] /8
H2(0) = X*(3) + X*(7) = – j + (2 + j) = 2
at
Pu ch

= [1.5 – (– 0.707 – j0.707) (4 – j4)] /8 = 0.89

H2(1) = X*(3) – X*(7) = – j – (2 + j) = – 2 – j2
We finally take the complex conjugate of the result.
Te

Stage 2
F1(0) = G1(0) + W04 G2(0)  x(n) ={1.69, – 0.81,– 0.31, – 0.52, 0.69,

= 3.5 + (1) (6) = 9.5 – 0.81, – 0.31, 0.89 }

F1(1) = G1(1) + W14 G2 (1) Ex. 4.6.6 : Given

X(k) = {2, – 6j, 2 – 8j, 6j, 2 – 6j, 2 + 8j, 6j, 0} Find x(n) using
= – 2.5 + (– j) (– j4) = – 6.5
any IFFT algorithm.
F1(2) = G1(0) – W04 G2(0)
Soln. :We will use the DIT-FFT algorithm. Since N = 8, we
= 3.5 – (1) (6) = – 2.5 use a 8-point DIT-FFT Butterfly diagram. In DIT-FFT, the
F1(3) = G1(1) – W14 G2 (1) input is in bit reversed order. We give X*(k) as the input
and divide the output - obtained by N. We also take
= 2.5 – (– j) (– j4) = 1.5 complex conjugate of the result obtained.
F2(0) = H1(0) + W04 H2(0) X(k) = { 2, – j6, 2 – j8, j6, 2 – j6, 2 + j8, j6, 0}

= 2 + (1) (2) = 4  {X(0), X(1), X(2), X(3), X(4), X(5), X(6), X(7)}

F2(1) = H1(1) + W14 H2 (1) Here

W04 = 1
= (2 – j2) + (– j) (– 2 – j2) = 0 –j2
·1  
F2(2) = H1(0) – W04 H2 (0) W14 = e 4
= cos – j sin = – j
2 2
= 2 – (1) (2) = 0
 DSIP (MU-Sem. 7-COMP) 4-35 Fast Fourier Transform

ge Fig. P. 4.6.6
io led
Similarly, F1(1) = G1(1) + W14 G2 (1)
W08 = 1
ic ow

= – j6 + [(– j) (2 + j14)] = 14 – j8
–j2
·1  
W18 = e 8
= cos – j sin F1(2) = G1(0) – W04G2(0)
n
4 4
bl kn

= 0.707 – j 0.707 = (4 + j6) – [(1) (2 + j2)] = 2 + j4

–j2
·2   F1(3) = G1(1) – W14 G2 (1)
W28 = e
at

8
= cos – j sin = – j
Pu ch

2 2
–j2
·3 3 3 = – j6 – [(– j) (2 + j14)] = – 14 – j4
W38 = e 8
= cos – j sin
Te

4 4
F2(0) = H1(0) + W04H2(0)
= –0.707 – j 0.707
We write down the equation at each stage and = (2 – j2) + [(1) (– j6)] = 2 – j8
calculate the outputs. F2(1) = H1(1) + W14 H2 (1)
Stage 1
G1(0) = X*(0) + X*(4) = 4 + j6 = (– 2 + j14) + [(– j)(– j6)]

G1(1) = X*(0) – X*(4) = – j6 = – 8 + j14

G2(0) = X*(2) + X*(6) = 2 + j2 F2(2) = H1(0) – W04H2 (0)

G2(1) = X*(2) – X*(6) = 2 + j14 = (– 2 – j2) – [(1) (– j6)]
H1(0) = X*(1) + X*(5) = 2 – j2
= 2 + j14
H1(1) = X*(1) – X*(5) = – 2 + j14
F2(3) = H1(1) – W14H2 (1) = (– 2 + j14) – [(–j) (– j6)]
H2(0) = X*(3) + X*(7) = – j6
H2(1) = X*(3) – X*(7) = – j6 = 4 + j4

Stage 2 Stage 3

F1(0) = G1(0) + W04G2(0) = (4 + j6) + (1) (2 + j2) x*(0) = [F1(0) + W08F2(0)] /N

= 6 + j8 = [(6 + j8)] + [(1) (2 – j8)]/8 = 1

 DSIP (MU-Sem. 7-COMP) 4-36 Fast Fourier Transform
x*(1) = [F1(1) + W18F2(1)] /N = [(6 + j8) – [(1) (2 – j8)]/8 = 0.5 + j2

= [(14 – j8) + [(0.707 – j0.707) (– 8 + j4)]/ 8 x*(5) = [F1(1) – W18F2(1)] /N

= 2.28 + j0.95 = [(14 – j8) – [(0.707 – j0.707)(– 8 + j4)]/8

x*(2) = [F1(2) + W28F2(2)] /N = 1.22 – j2.94

= [(2 + j4) + (– j) (2 + j4)]/8 = 0.75 + j0.25 x*(6) = [F1(2) – W28F2(2)] /N

x*(3) = [F1(3) + W38F2(3)] /N = [(– 14 – j4) = [(2 + j4) – [(– j) (2 + j4)]]/8 = – 0.25 + j0.75

+ [(– 0.707 – j0.707) (4 + j14)]/8 x*(7) = [F1(3) – W38F2(3)] /N

= – 0.86 – j2.09 = [(– 14 – j4) – [(– 0.707 – j0.707) (4 + j14)]/8

x*(4) = [F1(0) – W08F2(0)] /N = – 2.63 + j1.09

ge
 x*(n) = {1, 2.28 + j0.95, 0.75 + j0.25, – 0.86 – j2.09, 0.5 + j2, 1.22 – j2.94, – 0.25 + j0.75, – 2.63 + j1.09}

What we have obtained is x(n)

io led
 x(n) = {1, 2.28 – j0.95, 0.75 – j0.25, – 0.866 + j2.09, 0.5 – j2, 1.22 + j2.94, – 0.25 – j0.75, – 2.63 – j1.09}
ic ow

Ex. 4.6.7 : Find IDFT of X(k) = {10, – 2 + 2j, – 2, – 2 – 2j} using DIT-FFT algorithm.
Soln. :
n
bl kn

X(k) = {10, – 2 + 2j, – 2, – 2 – 2j}

{X(0) , X(1) , X(2) , X(3)}
at
Pu ch

Since N = 4, we use a 4 – point DIT-FFT butterfly diagram.

In DIT-FFT, the input is in bit reversed order. We give X*(k) as input and divide the output obtained by N. We also take
the complex conjugate of the result obtained.
Te

Fig. P. 4.6.7

Here, W40 = 1
W14 = –j
Stage 1
G1(0) = X*(0) + X*(2) = 10 + (– 2) = 8
G1(1) = X*(1) – X*(2) = 10 – (– 2) = 12
G1(2) = X*(1) + X*(3) = (– 2 – j2) + (– 2 + j2) = – 4
G1(3) = X*(1) – X*(3) = (– 2 – j2) – (– 2 + j2) = – j4
 DSIP (MU-Sem. 7-COMP) 4-37 Fast Fourier Transform

Stage 2 –j2
4 ·0
x*(0) = [G1(0) + W04G1(2)] /4 Here, W40 = e = cos 0 – j sin 0 = 1
–j2
= [8 + (1) (– 4)]/4 = 1 4 ·1  
W14 = e = cos – j sin = – j
2 2
x*(1) = [G1(1) + W14G1(3)] /4
Stage 1
= [12 + (– j) (– j4)]/4 = 2
G1(0) = x(0) + x(2) = 1 + 1 = 2
x*(2) = [G1(0) – W04G1(2)] /4 = [8 – (1) (– 4)]/4 = 3
G1(1) = x(0) – x(2) = 1 – 1 = 0
x*(3) = [G1(1) – W14G1(3)] /4 G1(2) = x(1) + x(3) = 2 + 4 = 6
= [12 – (– j) (– j4)]/4 = 4 G1(3) = x(1) – x(3) = 2 – 4 = – 2
 x*(n) = {1, 2, 3, 4}
Stage 2
 x(n) = {1, 2, 3, 4}
X(0) = G1(0) + W04G1(2) = 2 + (1) (6) = 8
4.6.2 Filtering using FFT Algorithms X(1) = G1(1) + W14G1(3) = 0 + (– j) (– 2) = j2

ge
Ex. 4.6.8 : Determine circular convolution of X(2) = G1(0) – W04G1(2) = 2 – (1) (6) = – 4
x(n) = {1, 2, 1, 4} and h(n) = {1, 2, 3, 2} using Radix-2 - FFT.
io led X(3) = G1(1) – W14G1(3) = 0 – (– j) (– 2) = – j2
Soln. : From the circular convolution property of the DFT,
 X(k) = {8, j2, – 4, – j2}
it is known than,
ic ow


DFT Step 2 : In a similar manner we compute
 h(n)
y(n) = x(n) O X(k) H(k) = Y(k)

IDFT the DFT of h(n)
n
Hence to determine circular convolution using DFT, we
bl kn

follow the given steps

1. Compute DFT of x(n) to obtain X(k)
at
Pu ch

2. Compute DFT of h(n) to obtain H(k)

3. Obtain Y(k) = X(k) · H (k)
Te

4. Compute IDFT of Y(k) to obtain y(n)

We will use the DIT-FFT to compute the DFT of x(n)
and h(n). Since both are of length 4, we will use a 4-point
Fig. P. 4.6.8(a)
DIF-FFT butterfly diagram. In DIF-FFT the output is
ordered and input is in bit reversed form. Stage 1
G1(0) = h(0) + h(2) = 1 + 3 = 4
Step 1 : Compute DFT of x(n)
G1(1) = h(0) – h(2) = 1 – 3 = – 2
G1(2) = h(1) + h(3) = 2 + 2 = 4
G1(3) = h(1) – h(3) = 2 – 2 = 0

Stage 2
H1(0) = G1(0) + W04G1(2) = 4 + (1) (4) = 8
H1(1) = G1(1) + W14G1(3) = – 2 + (– j) (0)

= –2
H1(2) = G1(0) – W04G1(2) = 4 – (1) (4) = 0
H1(3) = G1(1) – W14G1(3) = – 2 – (– j) (0)

= –2
Fig. P. 4.6.8
 H(k) = {8, – 2, 0, – 2}
 DSIP (MU-Sem. 7-COMP) 4-38 Fast Fourier Transform
Step 3 : y(n) = x(n)  h(0)
Y(k) = X(k) · H(k) We generate a circular matrix of x(n)
Y(0) = X(0) · H(0) = (8) · (8) = 64
Y(1) = X(1) · H(1) = (j2) · (– 2) = – j4
Y(2) = X(2) · H(2) = (– 4) · (0) = 0
Y(3) = X(3) · H(3) = (– j2) (–2) = j4
 Y(k) = {64, – j4, 0, j4}
Step 4 :  y(n) = {16, 18, 16, 14}

The final step is to obtain y(n). This is done by Hence it is seen the results obtained are the same.
computing the IDFT of Y(k). We will use the DIT-FFT Ex. 4.6.9 : By means of FFT-IFFT technique, compute the
butterfly diagram with the required modification. linear convolution of x(n) = {2, 1, 2, 1},
h(n) = {1, 2, 3, 4}

ge
Soln. : We are supposed to compute linear convolution
using FFT-IFFT.
io led
We know from the circular convolution property that,

DFT
 h(n)
y(n) = x(n) O X(k) · H(k) = y(k)

ic ow

IDFT

We also know that to obtain linear convolution from

n
circular convolution, the lengths of each of signals should
bl kn

be N1 + N2 – 1.
Fig. P. 4.6.8(b)
In this case length x(n) and h(n) = 4.
at
Pu ch

Stage 1
G1(0) = y*(0) + y*(2) = 64 + 0 = 64 If x(n) and h(n) are made equal to (4 + 4 – 1) = 7,
then circular convolution and linear convolution will give us
Te

G1(1) = y*(0) – y*(2) = 64 – 0 = 64

the same result. Then we can use the above mentioned
G1(2) = y*(1) + y*(3) = j4 + (– j4) = 0
property to obtain linear convolution using FFT-IFFT. We
G1(3) = y*(1) – y*(3) = j4 – (– j4) = j8 increase the size of x(n) and h(n) by Zero padding.
Stage 2  x(n) = {2,1, 2, 1, 0, 0, 017

y(0) =[G1(0) + W04G1(2)]/4 = [64 + (1)(0)]/4 = 16 h(n) = {1, 2, 3, 4, 0, 0, 0}17

y(1) = [G1(1) + W14G1(3)]/4 In order to use the radix-2 algorithm, the lengths of
the signal must be powers of 2. We therefore pad one more
= [64 + (– j) (j8)]/4 = 18
zero to both x(n) and h(n) and make their lengths equal to 8.
y(2) = [G1(0) – W04G1(2)/4 = [64 – (1) (0)]/4 = 16
Once the final result is obtained, we shall delete the
y(3) = [G1(1) – W14G1(3)]/4 last value.
x(n) = {2, 1, 2, 1, 0, 0, 0}18
= [64 – (– j) (j8)/4 = 14
h(n) = {1, 2, 3, 4, 0, 0, 0, 0}18
y(n) = {16, 18, 16, 14}
Since y*(n) has all real values Steps involved are:

y(n) = y*(n) 1. Compute DFT of x(n) to obtain X(k)

 y(n) = {16, 18, 16, 14} 2. Compute DFT of h(n) obtain H(k)

We would have obtained the same result had we 3. Obtain Y(k) = X(k) · H(k)
performed time domain circular convolution. Let us check 4. Compute IDFT of y(k) to obtain y(n)
the result.
 DSIP (MU-Sem. 7-COMP) 4-39 Fast Fourier Transform

We shall use the DIT-FFT algorithm to compute the DFT of x(n) and h(n).
Step 1 :Computing X(k) from x(n)
X(n) = {2, 1, 2, 1, 0, 0, 0, 0} x(0) x(1) x(2) x(3) x(4) x(5) x(6) x(7)
Here W04 = 1
–j2
·1  
W14 = e 4
= cos – j sin = – j
2 2
Similarly, W08 = 1
–j2
·1  
W18 = e 8
= cos – j sin = 0.707 – j 0.707
4 4
–j2
·2  
W28 = e 8
= cos – j sin = – j
2 2
–j2
·3 3 3
W38 8

ge
= e = cos – j sin = – 0.707 – j 0.707
4 4
We write down the equation at each stage and calculate the outputs.
io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 4.6.9

Stage 1 G1(0) = x(0) + x(4) = 2 + 0 = 2 F1(1) = G1(1) + W14 G2(1) = 2 +(– j) (2) = 2 – j2
G1(1) = x(0) – x(4) = 2 – 0 = 2 F1(2) = G1(0) – W04 G2(0) = 2 – (1) (2) = 0
G2(0) = x(2) + x(6) = 2 + 0 = 2 F1(3) = G1(1) – W14 G2(1) = 2 – (– j) (2) = 2 + j2
G2(1) = x(2) – x(6) = 2 – 0 = 2
F2(0) = H1(0) + W04 H2(0) = 1 + (1) (1) = 2
H1(0) = x(1) + x(5) = 1 + 0 = 1
F2(1) = H1(1) + W14 H2(1) = 1 + (– j) (1) = 1 – j1
H1(1) = x(1) – x(5) = 1 – 0 = 1
F2(2) = H1(0) – W04 H2(0) = 1 – (1) (1) = 0
H2(0) = x(3) + x(7) = 1 + 0 = 1
F2(3) = H1(1) – W14 H2(1) = 1 – (– j) (1) = 1 + j1
H2(1) = x(3) – x(7) = 1 – 0 = 1
Stage 3
Stage 2
X(0) = F1(0) + W80F2(0) = 4 + (1) (2) = 6
F1(0) = G1(0) + W04 G2(0) = 2 +(1) (2) = 4
X(1) = F1(1) + W81F2(1)
 DSIP (MU-Sem. 7-COMP) 4-40 Fast Fourier Transform

= (2 – j2) + [(0.0707 – j0.707) (1 – j1)]  X(k) = {6, 2 – j 3.41, 0, 2 + j0.58, 2,

= 2 – j3.41 2 – j0.58, 0, 2 + j3.41}

X(2) = F1(2) + W82F2(2) = 0 + (– j) (0) = 0 Step 2 :Computing H (k) from h (n)

X(3) = F1(3) + W83F2(3) h(n) = {1, 2, 3, 4, 0, 0, 0, 0}

{h(0), h(1), h(2), h(3), h(4), h(5), h(6), h(7)}
= (2 + j2) + [(– 0.707 – j0.707 (1 + j1)]
We write down the equation at each stage and calculate
= 2 + j0.58 the outputs
X(4) = F1(0) – W08F2(0) = 4 – (1) (2) = 2 Stage 1
X(5) = F1(1) – W18F2(1) G1(0) = h(0) + h(4) = 1 + 0 = 1
G1(1) = h(0) – h(4) = 1 – 0 = 1
= (2 – j2) – [(0.707 – j0.707) (1 – j1)]
G2(0) = h(2) + h(6) = 3 + 0 = 3
= 2 – j0.58

ge
G2(1) = h(2) – h(6) = 3 – 0 = 3
X(6) = F1(2) – W28F2(2) = 0 – (j) (0) = 0
io led H1(0) = h(1) + h(5) = 2 + 0 = 2
X(7) = F1(3) – W38F2(3) H1(1) = h(1) – h(5) = 2 – 0 = 2

= (2 + j2) – [(– 0.707 – j0.707) (1 + j1)] H2(0) = h(3) + h(7) = 4 + 0 = 4

ic ow

H2(1) = h(3) – h(7) = 4 – 0 = 4

= 2 + j3.41
n
bl kn
at
Pu ch
Te

Fig. P.4.6.9(a)

Stage 2 F1 (0) = G1(0) + W04 G2(0) = 1 + (1)(3) = 4

F1 (1) = G1(1) + W14 G2(1) = 1 + (– j)(3) = 1 – j3
F1 (2) = G1(0) – W04 G2(0) = 1 – (1) (3) = – 2
F1 (3) = G1(1) – W14 G2(1) = 1 – (– j)(3) = 1 + j3
F2 (0) = H1(0) + W04H2(0) = 2 + (1) (4) = 6
F2 (1) = H1(1) + W14H2(1) = 2 + (– j) (4) = 2 – j4
F2 (2) = H1(0) – W04H2(0) = 2 – (1) (4) = – 2
F2 (3) = H1(1) – W14H2(1) = 2 – (– j) (– 4) = 2 + j4
 DSIP (MU-Sem. 7-COMP) 4-41 Fast Fourier Transform

Stage 3  y(n) = {2, 5, 10, 16, 12, 11, 4}

H(0) = F1(0) + W80F2(0) = 4 + (1) (6) = 10 This is the same result that would be obtained of we
H(1) = F1(1) + W81F2(1) perform linear convolutions in the time domain. Let us
check the result,
= (1 – j3) + [(0.707 – j0.707) (2 – j4)]
y(n) = x(n) * h(n)
= – 0.41 – j7.24
H(2) = F1(2) + W82F2(2) = – 2 + [(– j) (– 2)] = – 2 + j2 We use the matrix method

H(3) = F1(3) + W83F2(3)

= (1 + j3) + [(– 0.707 – j0.707) 2 + j4]
= 2.41 – j1.24
H(4) = F1(0) – W08F2(0) = 4 – (1) (6) = – 2
H(5) = F1(1) – W18F2(1)

ge
= (1 – j3) – [(0.707 – j0.707) (2 – j4)]
= 2.41 + j1.24
io led Fig. P.4.6.9(b)
H(6) = F1(2) – W28F2(2) = – 2 – [(– j) (– 2)]
y(n) = {2, 5, 10, 16, 12, 11, 4}
= – 2 – j2
Hence it is seen that the results obtained one the same.
ic ow

H(7) = F1(3) – W38F2(3)

= (1 + j3) – [(– 0.707 – j0.707) (2 + j4)] Summary
n
= – 0.41 + j7.24
bl kn

H(k) = {10, – 0.41 – j7.24, – 2 + j2, In this chapter we study the algorithms proposed by
Cooley and Tukey. This method came to be known as the
at

2.41 – j1.24, – 2, 2.41

Pu ch

Fast Fourier Transform (F.F.T). The FFT requires only N log2

+ j1.24, – 2 – j2, – 0.41 + j7.24} N calculations as opposed to N2 used by the DFT. Two
Step3: Obtain Y (k)= X(k) · H(k) important algorithms viz. Decimation in Time and
Te

Y(0) = X(0) · H(0) = 60 decimation in Frequency for Radix 2 were derived. We also
learnt how to use the butterfly diagram to compute the
Y(1) = X(1) · H(1) = – 25.55 – j13.07
Inverse DFT. Many examples were solved to understand
Y(2) = X(2) · H(2) = 0 how to compute the DFT using the FFT. Filtering using the
Y(3) = X(3) · H(3) = 5.55 – j1.07 FFT algorithms were also carried out.
Y(4) = X(4) · H(4) = – 4
Y(5) = X(5) · H(5) = 5.55 + j1.07 Review Questions
Y(6) = X(6) · H(6) = 0
Q. 1 Draw basic butterfly structure for DIT- FFT.
Y(7) = X(7) · H(7) = – 25.55 + j13.07
Q. 2 Draw total signal flow graph for DIT-FFT.
  Y(k) = {60, – 25.55 – j13.07, 0,
Q. 3 Obtain computational complexity of FFT
5.55 – j1.07, – 4, 5.55 + j 1.07, 0, algorithm.
– 25.55 + j13.07} Q. 4 Explain in place computation of FFT.
Step 4: Compute IDFT of y(k) to obtain y(n) we use the Q. 5 Derive the first stage of DIF-FFT algorithm.
DIT–FFT algorithm to compute the IDFT. The final output Draw the basic butterfly structure for the same.
obtained after computing the IDFT is, Q. 6 Compare DIT-FFT and DIF-FFT algorithms.
y(n) = {2, 5, 10, 16, 12, 11, 4, 0} Q. 7 Write a short note on bit reversal in FFT.
Since an extra zero was added to x(n) and h(n) to
increase their lengths to 8, we remove the last zero of y(n).

 5 Introduction to Image Processing

Syllabus
Introduction to Digital Image, Digital Image Processing System, Image File Formats: BMP, TIFF and JPEG.

5.1 Introduction  Early human beings strove to record their world by

carving crude diagrams on stone. All the drawings that
Human beings are primarily visual creatures who we see in old caves is just that; storing images seen,
depend on their eyes to gather information around them. trying to analyse them and discussing it with others in

ge
Of the five senses that human beings have, sight is what we the tribe. Refer Fig. 5.2.1.
depend upon the most. Not many animals depend on their
 This art developed through the ages by way of
visual systems; the way human beings do.
materials and skill. By the mid-nineteenth century,
io led
Bats use high frequency sound waves. They emit
photography was well established. Image processing
sound waves which reflect back when they encounter some
that we study starts from this era.
obstruction. Cats have poor vision but an excellent sense of
ic ow

smell. Snakes locate prey by heat emission and fish have

organs that sense electrical fields.
n
5.2 What Do We Mean by Image
bl kn

Processing ?
at
Pu ch

What happens when we look at an object ?

 The eye records the scene and sends signals to the
brain. These signals get processed in the brain and
Te

some meaningful information is obtained. Let us take a Fig. 5.2.1

simple example : when we see fire, we immediately
 Though it was stated earlier that the human eye-brain
identify it as something hot. Two things have
mechanism represents the ultimate imaging system,
happened here.
image processing as a subject involves processing
(1) The scene has been recorded by the eye.
images obtained by a camera. With the advent of
(2) The brain processed this scene and gave out a
computers, image processing as a subject grew rapidly.
warning signal.
This is image processing !!!  Images from a camera are fed into a computer where
 We start processing images from the day we are born. algorithms are written to process these images. Here,
Hence image processing is an integral part of us and the camera replaces the human eye and the computer
we continue to process images till the day we die. So does the processing.
even if this subject seems to be new, we have been  Hence image processing as an engineering subject is
subconsciously doing this, all these years. The human basically manipulation of images by a computer.
eye-brain mechanism represents the ultimate imaging
system.
 Apart from our vision, we have another important trait
that is common to all human beings. We like to store
information, analyse it, discuss it with others and try to
better it. This trait of ours is responsible for the rapid
development of the human race.
Fig. 5.2.2
 DSIP (MU-Sem. 7-COMP) 5-2 Introduction to Image Processing

5.3 Images  Let us see why images are 2-dimensional functions.

 Let us now define what we mean by an image. The Consider the image shown in Fig. 5.3.3.
world around us is 3-dimensional, while images
obtained through a camera are 2-dimensional. Hence
an image can be defined as a 2-dimensional
representation of a 3-dimensional world. Consider the
image shown in Fig. 5.3.1.

Fig. 5.3.3

Fig. 5.3.1

ge
 In moving from the 3-dimensional world to the
2-dimensional image, we loose one dimension. Depth
information is lost. As can be seen from the Fig. 5.3.1,
io led
the handle of the tea cup is missing.
 All family pictures, photographs on identity cards etc.
are 2-dimensional. If this statement is not clear, let us
ic ow

take a simple example.

Example of a 1-dimensional and a 2-dimensional signal
n
 Consider a voltage signal shown in Fig. 5.3.2. We are all Fig. 5.3.4
bl kn

familiar with a signal of this kind. Here the voltage is  In this case, to locate the dot shown, we need to know
varying with respect to time. This is a typical its position in two directions (x and y) Fig. 5.3.4. Hence
at
Pu ch

1-dimensional signal. If we want to locate a dot on the all images that we see are 2-dimensional functions. A
wave, all we need to know is its corresponding time. typical image is represented as shown in Fig. 5.3.4.
Here (x1, y1) are the spatial coordinates and f is the
Te

grey level (colour in the case of colour image) at that

point. Hence grey level f varies with respect to the x
and y coordinates.

Fig. 5.3.2

Fig. 5.3.5
 DSIP (MU-Sem. 7-COMP) 5-3 Introduction to Image Processing

5.4 The Electromagnetic Spectrum

 The apparatus shown in Fig. 5.2.2 will work only if light is incident on the object. What we call light is actually a very
small section of the electromagnetic energy spectrum. The entire spectrum is shown in Fig. 5.4.1.

ge
io led Fig. 5.4.1

 Electromagnetic energy, as the name suggests, exists Solved Example

in the simultaneous form of electricity and magnetism.
ic ow

These two forms of the same energy are transmitted Ex. 5.4.1 : Calculate the frequency of oscillation of green
light.
n
together as electromagnetic radiation. One cannot
Soln. :
bl kn

exist without the other. A flow of electric current

always produces magnetism, and magnetism is used to It has been known that green light has a wavelength of
at

produce electricity. Electromagnetic radiation is approximately 500 nm

Pu ch

–9
propagated outwards from its source at a velocity of (500  10 m)
8
300,00,0000 meters per second (3  10 m/sec). Its frequency of oscillations can be calculated using
Te

 Although our natural source of electromagnetic Equation (5.4.1).

radiation is the sun, there are also a number of man- c
 =

made sources which, among many others, include
3  10 m/sec
8
c
tungsten filament lamps, gas discharge lamps and  = =
 500  10 m
–9
lasers. Light is a band of electromagnetic radiation
 = 6  10 Hz
14
mediated by the human eye and is limited to a
spectrum extending from 380 nm to 760 nm. i.e., the frequency of green light is
600,000,000,000,000 cycles/sec !
 Most of the images that we encounter in our day to
Hence it is more convenient to discuss
day life are taken from cameras which are sensitive to
electromagnetic radiation in terms of wavelengths (nm)
this range of the electromagnetic spectrum (380 - 760
rather than frequencies (Hz).
nm). We must not forget though, that there are
cameras which are capable of detecting infrared, 5.5 Units of Intensity
ultraviolet light, X-rays and radio waves too.
 The electromagnetic spectrum can be expressed in  We know that for an object to be seen, some amount
of light has to fall on it.
terms of wavelength and frequency. The wavelength
() and the frequency () are related by the expression  The unit of luminous intensity (I) is candela (cd) (SI
th
c unit) which by definition, is (1/60) of a square
 = …(5.4.1) centimetre of the surface of a black body radiator at

8 the absolute temperature of 2045 K.
Here c is the speed of light = 3  10 m/sec
 DSIP (MU-Sem. 7-COMP) 5-4 Introduction to Image Processing

Note : A black body radiator is one that absorbs all the  If we want an infra-red image, we use cameras which
radiation incident upon it and has no reflecting power. The
are sensitive to infra-red radiations. For normal images
(family pictures etc.) we use cameras which are
radiation from a heated black body source is called black
sensitive to the visual spectrum. In this book we shall
body radiation and is always quoted in Kelvin.
discuss cameras (sensors) which are sensitive only to
The unit is called candela because initially it was the visual range.
defined as a standard candle burning a specific amount of
Quantum detectors
wax per hour. There is another important unit apart from
 Quantum detection is the most important mechanism
candela that we encounter. This unit is called lux.
of image sensing and acquisition. It relies upon the
5.6 Basic Elements/Steps of an Image energy of absorbed photons being used to promote
Processing System electrons from their stable state to a higher state
above an energy threshold.
MU - Dec. 2018
 Whenever this occurs, the properties of that material

ge
Q. Explain Fundamental steps in Image processing. get altered in some measurable way. Planck/Einstein
(Dec. 2018, 5 Marks) came up with a relationship between the wavelength
of the incident photon and the energy that it carries.
io led
Digital image processing is basically modification of
images on a computer. The basic components of an image hc
e = …(5.6.1)
processing system are as follows : 
ic ow

(1) Image Acquisition. Where, e = Energy carried by a photon

h = Planck’s constant, 6.626  10
–34
(2) Image Storage. Js
n
c = Speed of light, 3  10 m/sec
8
(3) Image Processing.
bl kn

   = Wavelength of the incident radiation.

(4) Display.
 On collision, the photon transfers all or none of this
at

(5) Transmission (if required).

Pu ch

quantum of energy to the electron.

We shall discuss each one in detail.  The Equation (5.6.1) is very important as it tells us that
Te

the maximum wavelength to which the quantum

detector will respond is determined by the energy
threshold and hence by the material selection.
 Of the several modes of operation of quantum
detectors, the most important ones are
(a) Photoconductive
(b) Photovoltaic.

(a) Photoconductive : The resistance of photoconductive

Fig. 5.6.1 materials drop in the presence of light due to the
generation of free charge carriers. An external bias is
(1) Image Acquisition
applied across the material to measure this change.
 Image acquisition is the first step in any image This principle of photoconductivity is used in Vidicon
processing system. The general aim of image imaging tubes.
acquisition is to transform an optical image (real world
Vidicon imaging tubes
data) into an array of numerical data which could be
later manipulated on a computer.  Vidicon is a vacuum tube used for image acquisition.
 Image acquisition is achieved by suitable cameras. We Fig. 5.6.2 is the schematic diagram of the vidicon tube
use different cameras for different applications. If we image sensor. The light image is focused on a
need an X-ray image, we use a camera (film) that is transparent signal plate coated on the inside with a
sensitive to X-rays. photoconductive target material.
 DSIP (MU-Sem. 7-COMP) 5-5 Introduction to Image Processing

 The front face of the signal plate is coated with SnO2 (tin oxide) which forms the conducting layer. This plate is biased
to a positive potential of about 50 V. The target coating is maintained at approximately 0 V by a mesh grid which is
placed behind it.

ge
io led
Fig. 5.6.2

 Due to this the signal plate and target behave as a  The biggest drawback of the vidicon tube is that it uses
ic ow

continuous array of capacitors containing a the concept of deflection coils to position the electron
photoconductive dielectric. beam. These deflection coils make the vidicon bulky
n
 When there is no light incident on the target, its and are responsible for major power drain. Since the
bl kn

resistance is very high due to which the charge across coils are magnetic, the vidicon also becomes sensitive
the capacitor is maintained. However when light to external fields.
at
Pu ch

strikes the target, the dielectric material becomes

(b) Photovoltaic
leaky and the target side voltage rises.
 Photovoltaic devices consist of semiconductor
 The target is now scanned by a low velocity electron
Te

junctions. They are solid state arrays composed of

beam. When this beam strikes the discharged
capacitor, current flows to restore the target side to its discrete silicon imaging elements known as Photosites.
original voltage of 0 V.  Photovoltaic devices give a voltage output signal that
 This current also flows in the external bias circuit, is proportional to the intensity of the incident light. No
which produces a voltage drop in the load resistance. external bias is required as was in the case of
 In order to produce the output necessary for photoconductive devices.
broadcast video applications, the electron beam is  The technology used in solid-state imaging sensors is
scanned across the surface of the target by based principally on charge-coupled devices,
electromagnetic deflection coils. (Deflection coils were commonly known as Charged Coupled Devices CCDs.
encountered when you studied CRO, (cathode ray Hence the imaging sensors are called CCD sensors.
oscilloscopes), in your lower semesters).
 The solid state array (CCD) can be arranged in two
 The original vidicons use antimony-trisulphide, which
different configurations.
is a photoconductive material, as their target coatings.
(i) Line array CCD.
Other forms of vidicon are made using more complex
target materials. (ii) Area array CCD.

 The silicon target vidicon achieves improved sensitivity

and a spectral response similar to that of the human (i) Line Arrays : The line array represents the simplest
eye. form of CCD imager and has been employed since the
early 1970s. Line arrays consist of a one-dimensional
array of photo sites.
 DSIP (MU-Sem. 7-COMP) 5-6 Introduction to Image Processing

Fig. 5.6.3

 A single line of CCD pixels are clocked out into the

parallel output register as shown in Fig. 5.6.3. The
amplifier outputs a voltage signal proportional to the
contents of the row of photosites.
Fig. 5.6.5
 One thing to note is that line array CCD scans only one  Area arrays or matrix arrays consist of a two-
line (hence is one-dimensional). In order to produce a dimensional array of photosites. They make it possible

ge
two-dimensional image, the line array CCD imager has to investigate static real world scenes without any
to be used as a scanning device by moving this array mechanical scanning. Thus much more information
io led
over the object by some mechanical activity. can be deduced from a single real time glance than
would be possible with line arrays.
 Area arrays can be seen in all the digital cameras that
ic ow

we use for video imaging. The area arrays are more

versatile than the line arrays, but there is a price to be
n
paid for this. Area arrays are higher on cost and
bl kn

complexity.
 Area sensors come in different ranges. i.e. 256  256,
at
Pu ch

490  380, 640  480, 780  575. CCD arrays are

typically packaged as TV cameras.
Te

 A significant advantage of solid state array sensors is

that they can be shuttered at very high speeds
(1/10,000 secs). This makes them ideal for applications
in which freezing motion is required.

(2) Image Storage

Fig. 5.6.4  All video signals are essentially in analog form i.e.
electrical signals convey luminance and colour with
 This technique is used in flat bed scanners (the
continuously variable voltage. The cameras are
scanners that you come across in your laboratory or in interfaced to a computer where the processing
a cyber cafe). A line array CCD can have anything from algorithms are written.
a few elements to upto 6000 or more.
 This is done by a frame grabber card. Usually a frame
(ii) Area Arrays : The problem with line arrays is that it grabber card is a Printed Circuit Board (PCB) fitted to
scans only one line. To get a two-dimensional image, the host computer with its analog entrance port
we need to mechanically move the array over the matching the impedance of the incoming video signal.
entire image.  The A/D converter translates the video signals into
digital values and a digital image is constructed. Frame
grabber cards usually have a A/D card with resolution
of 8 - 12-bits (256 to 4096 gray levels). Hence a frame
grabber card is an interface between the camera and
the computer.
 DSIP (MU-Sem. 7-COMP) 5-7 Introduction to Image Processing

 Frame grabber card has a block of memory, separate from the computers own memory, large enough to hold any
image. This is known as the frame buffer memory. Image data, usually in the form of bytes (8-bits), is written into the
frame grabber memory under the computer control, usually by DMA transfers.
 The contents of the memory are continuously read out at video rate (30 frames/sec), passed through the D/A
converter and displayed on the monitor.
 The output has a colour map or a colour Look Up Table (LUT) to permit pseudo colouring. The table consists of 8-bit
values for each of the red, green and blue guns of the monitor.
 The commercially available frame grabber cards have additional features such as ability to zoom regions of the image
and pan the images.
 The frame buffer memory can be upto 5 MB. Images being 2-dimensional functions, occupy a lot of space and hence
the storage space on the host computer has to be large. Let us try to find out as to how much space is actually taken
by images.

ge
io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. 5.6.6 : Simplified diagram of a frame grabber card

 Each image is stored as a matrix, where every value of

the matrix represents the grey level at that point.
Suppose the image (matrix) is of size A  B and
suppose we require C bits to represent each element
of the matrix. Memory space required to store this
image is approximately equal to A  B  C bits.
 Suppose we have D such images then total number
of bits  A  B  C  D ...(5.6.2)

Fig. 5.6.7(b)

 Thus the formula  ABCD gives us a fairly good idea of

the amount of space required.

Fig. 5.6.7 (a)

 DSIP (MU-Sem. 7-COMP) 5-8 Introduction to Image Processing

Table 5.6.1 important role of merging the image processing

hardware with the host computer.
A B C D Number of bits
6  One thing that we should remember is, image
Low quality 64 64 8 6  0.2  10 bits
video phone processing is characterized by specific solutions.
6 Hence there is no one way to process images. A
Digital 720 576 25 8  83  10 bits
technique that works well in one area can be totally
broadcast
TV useless in some other applications.

HDTV 1920 1150 50 8

6
 883  10 bits  The image processing software that is used in this
book is MATLAB.
 With the advent of the PCI bus, the host computers
memory could now be used as the frame buffer (4) Display
memory. This is because the PCI bus facilitates fast
 A display device produces and shows a visual form of
communication. numerical values stored in a computer as an image
 Hence the image could now directly be stored on the array. Principal display devices are printers, TV

ge
computer RAM. Due to this, the frame buffer memory monitors and CRTs.
has become more or less obsolete. Only the very high
 Any erasable raster graphics display can be used as a
io led
end frame grabber cards still have a small frame
display unit with an image processing system.
grabber memory embedded on them.
However monochrome and colour TV monitors are the
 Images being two-dimensional functions, occupy a lot
principal display devices used in modern image
ic ow

of space and hence it is advisable to have a computer

processing systems.
with a sizeable hard disk and also a good RAM.
 These raster devices convert image data into a video
n
Processed images could be stored on magnetic
frame. One major problem is, it must refresh the
bl kn

floppies or CDs. Archival data are stored on magnetic

screen at a rate of about 30 frames per second to
tapes.
avoid flicker.
at
Pu ch

(3) Image Processing

 Since some computers are unable to transfer data at
 Systems ranging from microcomputers to general
such high speeds, it is a good idea to buy a high end
purpose large computers are used in image
Te

frame grabber card which has frame buffer memory

processing. Dedicated image processing systems
on it.
connected to host computers are very popular now-a-
days.  If you want to build a image processing the system a
 Processing of digital images involve procedures that home, you should have a Video Graphics Card (VGA)
attached to a computer and a monitor that supports
are usually expressed in algorithmic form due to which
most image processing functions are implemented in that VGA card. One could then use a simple camera
that comes along with the system.
software.
 The only reason for specialized image processing (5) Transmission
hardware is the need for speed in some applications or
 There are a lot of applications where we need to
to overcome some fundamental computer limitations.
transmit images. The stages in the transmission of an
 The trend though is to merge general purpose small
image over a channel or network are shown in
computers with image processing hardware. As stated
Fig. 5.6.8.
in the earlier section, frame grabber cards play the

Fig. 5.6.8
 DSIP (MU-Sem. 7-COMP) 5-9 Introduction to Image Processing

 The image sequence from the camera is coded into as 5.7 Image Types
concise a representation as possible for transmission
over the channel. Images are 2-dimensional functions, as shown in
Fig. 5.7.1.
 Most transmission in broadcast television are still in
analog form and analog coding methods are used to
make it as efficient as possible. NTSE, PAL and SECAM
are the 3 major coding systems used in various parts of
the world (USA uses NTSE, while India uses PAL).
 Digital image coding is a high activity area. In image
processing; it is concerned with efficient transmission
of images over digital communication channels. A
Fig. 5.7.1
variety of indigenous ideas have been developed in
recent years. Images can be classified as follows :


ge
Coding is influenced by the type of the channel used to
carry the image signals. Several different types of
transmission channels are encountered in practice
io led
including cables, terrestrial radio, satellites, and optical
fibres. After decoding at the receiver, the image
sequence is
ic ow

 Gamma : In the image transmission chain, there are

many elements which exhibit a non-linear behaviour.
n
If x is the input and y is the output, then
bl kn


y = cx
Fig. 5.7.2 : Image Types
where, c = Constant
at
Pu ch

(1) Monochrome Image : In this, each pixel is stored as a

 = Gamma of the device
single bit (0 or 1). Here, 0 represents black while 1
 The camera, the display device and eye all have non-
represents white. It is a black and white image in the
Te

unity gammas (all are non-linear). Hence to make sure

strictest sense. These images are also called bit
that the perceived grey scale in the displayed image is
mapped images. In such images, we have only black
correct, it is necessary to insert an additional
compensating, non-linear device called the gamma and white pixels and no other shades of grey. Refer
corrector. Fig. 5.7.3.
 We have discussed a lot of things so far and it is
advisable at this stage to have a branch diagram of
these topics.

Fig. 5.7.3
(2) Grey Scale Image : Here each pixel is usually stored as
a byte (8-bits). Due to this, each pixel can have values
ranging from 0 (black) to 255(white). Grey scale
images, as the name suggests have black, white and
various shades of grey present in the image. Refer
Fig. 5.7.4.
Fig. 5.6.9
 DSIP (MU-Sem. 7-COMP) 5-10 Introduction to Image Processing

new(x,y) = (a(x,y)+b(x,y)+c(x,y))/3;
new1(x,y) = 3*a(x,y)+0.59*b(x,y)+0.11*c(x,y);
end
end
figure(1)
imshow(uint 8(im))
figure(2)
imshow(uint 8(new))
figure(3)
Fig. 5.7.4
imshow(uint8(new1))
(3) Colour Image (24-bit)
Matlab has an inbuilt command for conversion
 Colour images are based on the fact that a variety of rgb2gray.

ge
colours can be generated by mixing the three primary
(4) Half Toning
colours viz, Red, Green and Blue, in proper
 It is obvious that a grey scale image definitely looks
io led
proportions. In colour images, each pixel is composed
of RGB values and each of these colours require 8-bits better than the monochrome image as it utilizes more
(one byte) for its representation. grey levels. But there is a problem in hand.

ic ow

 Hence each pixel is represented by 24-bits [R(8-bits), Most of the printers that we use (inkjet, lasers, dot
G(8-bits), B(8-bits)]. matrix) are all bi-level devices. i.e., they have only a
n
A 24-bit colour image supports 16, 777, 216 different black cartridge and can only produce two levels (black
bl kn

combination of colours. on a white back-ground). In fact, most of the printing

jobs are done using bi-level devices.
 Colour images can be easily converted to grey scale
at

 You have all read newspapers at some point of time

Pu ch

images using the equation

(hopefully). The images do look like grey level images.
X = 0.30 R + 0.59 G + 0.11 B ...(5.7.1)
But if you look closely, all the images generated are
Te

 An easier formula that could achieve similar results is

basically using black colour. Refer Fig. 5.7.5.
R+G+B
X = ...(5.7.2)
3
MATLAB code for converting a colour image to a grey
scale image is as follows :
%% Converting a colour image to a grey level image %%
clear
clc
im = imread(‘lily.tif’);
[row col byt]=size(im);
a = im(:,:,1); % Red plane
b = im(:,:,2); % Green plane
Fig. 5.7.5
c = im(:,:,3); % Blue plane
 Even the images that you see in most of the books
a = double(a);
(including this one) are generated using black colour
b = double(b); on a white background. In spite of this we do get an
c = double(c); illusion of seeing grey levels. The technique to achieve
for x = 1:1:row an illusion of grey levels from only black and white
for y = 1:1:col levels is called half-toning.
 DSIP (MU-Sem. 7-COMP) 5-11 Introduction to Image Processing

 The dots that are placed in the 3  3 matrix example

can be in any order. But we need to follow two rules.
(1) Dots should be arranged in such a manner so that they
don’t form straight lines. Lines are very obvious to the
viewer and hence should be avoided.
Example, suppose in an image we have 4 consecutive
grey levels of value 3, and if we use a code as shown in
Fig. 5.7.6 Fig. 5.7.8, the half-toned image would look like
 The human eye integrates the scene that it sees. Fig. 5.7.9.
Consider a simple example. Consider two squares of
say 0.03  0.03 sq.inch.
 One of these squares contains a lot of black dots while
the other square contains fewer black dots. When we

ge
look at these squares from a distance, the two squares
give us a perception of 2 different grey levels. Fig. 5.7.8

io led
This integration property of the eye is the basis for half
toning. In this, we take a matrix of a fixed size and
depending on the grey level required, we fill the matrix
ic ow

with black pixels.

Let us take an example.
n
 Consider a 3  3 matrix. This matrix can generate an Fig. 5.7.9
bl kn

illusion of 10 different grey levels when viewed from a It can be seen that it does not look like a grey level. It
distance. looks like a straight black line. Hence lines should be
at
Pu ch

avoided while defining the matrices.

(2) If a pixel in a particular matrix is black for grey level i, it
Te

should be black for all further levels j > i. This reduces

false contouring.
Half-toning gives excellent results and we do perceive
a grey level image just by using black pixels on a white
background.
The logic to implement a half-toned image from a grey
level image is as follows :
(i) Define the size of the half-toned matrices based
on the number of grey levels the original image
Fig. 5.7.7 has.

 Here, the first block represents white, the last block (ii) Generate the matrices starting from all white
represents black and all the other blocks represent pixels to all black pixels.
intermediate values of grey. So in this case, while (iii) Read the original image. For every grey level value
printing, if we encounter grey level 0, we plot 9 pixels read, plot the corresponding matrix.
which are all white. If we encounter grey level 1, we Remember, the physical size of the half-toned image
plot 9 pixels of which only one pixel is black and so on. will always be bigger than the original image as for
 As is evident a 3  3 matrix will generate 10 different every single grey level value, we output an entire
grey levels. matrix.
 DSIP (MU-Sem. 7-COMP) 5-12 Introduction to Image Processing

5.8 Image File Formats  JPEG (Joint Photographic Expert Group) is the name of
the committee that developed an image format which
MU : May 2016, Dec. 2016
used compression algorithms.
Q. Explain in detail any two types of image file formats.  JPEG images are compressed images which occupy
(May 2016, 8 Marks) very little space. It is based on the Discrete Cosine
Q. What are various file formats ? Explain each in brief. Transform-DCT (explained in chapter of Image
(Dec. 2016, 8 Marks) Transforms). JPEG images use lossy compression
algorithms which result in some loss of original data
 Images obtained from the camera are stored on the
and hence the quality of JPEG images is not as good as
host computer using different formats. A file format is
BMP images.
a structure which defines how information is stored in
the file and how that information is displayed on the  Although these formats differ in technical details, they
monitor. share structural similarities.

 There are numerous image file formats which are  The Fig. 5.8.1 shows the typical organisation of

ge
available. Of these only a few of them can be used information encoded in an image file.
universally. By universally it means, they can be The image file consists of two parts;
understood by different operating systems.
io led (1) Header
 Some of the image formats that can be used on either (2) Image data
Macintosh or PC platforms are;
ic ow

BMP (Bit Mapped JPEG (Joint Photographic

Graphic Image) Expert Group)
n
bl kn

TIFF (Tagged Image File EPS (Encapsulated Post

Format) Script)
at
Pu ch

GIF (Graphic Interchange PICT (Macintosh Supported)

Format)
Te

 TIFF, which is one of the most well known formats was

developed in 1986 and in its many versions is a
standard image format for a bit-mapped graphics
image. The TIFF format is also a data compression
technique for monochrome as well as colour images.
Images seen on the Internet sites are normally
TIFF/GIF images simply because they occupy less
space. GIF uses a form of Huffman coding. Fig. 5.8.1
 BMP images developed by Microsoft can save both  The header file gives us the information about the kind
monochrome as well as colour images. All the wall of image. The header file, begins with a binary code or
papers that your computer has are BMP images. ASCII string which identifies the format being used.
 Similarly when we work in Paint Brush, we can save  The width and height of the image are given in number
our work as a BMP image. The quality of BMP files is of pixels. Common image types include binary images,
very good but they occupy a lot of memory space. PICT 8-bit grey scale images, 8-bit colour and 24-bit colour
is a general purpose format supported by Macintosh images.
and used for storing bit-mapped images.
 The image data format specifies the order in which
 EPS is file format of the post script page description pixel values are stored in the image data section. A
language and is device independent. This simply commonly used order is left to right and top to
means that images can readily be transferred from one bottom. Image data format also specifies if the RGB
application to another. However EPS images can only values in the image are interlaced.
be printed on post script compatible printers.
 DSIP (MU-Sem. 7-COMP) 5-13 Introduction to Image Processing

 By interlaced we mean that the RGB values of each  From this value, we could find out the number of bits
pixel stay together consecutively followed by the three allocated for each pixel. For example, if the maximum
colour components for the next pixel i.e., allowable pixel value is 255 then we know that each
8
R G B R G B R G B …… bit is represented by 8 bits (2 = 256).
st nd  Fig. 5.8.2 shows a 8  8 grey scale image and its
1 Pixel 2 Pixel ...........
representation as a ASCII PGM file
 If the RGB values are not interlaced, the values of one
primary colour for all pixels appear first, then the
values of the next primary colour followed by the
values of the third primary colour. Thus the image data
is in the form.
R R R R........ G G G G.......... B B B B.......
 The compression type in the header indicates whether
the image data is compressed using algorithms such as

ge
Run length encoding. Apart from these, there are a
few more formats which we come across while dealing
Fig. 5.8.2
with images.
io led
PGM (Portable Grey Map) PGM is used to represent
grey level images.
ic ow

PBM (Portable Bit Map) PBM is used to represent

binary images
n
bl kn

PPM (Portable Pixel Map) PPM is used to represent

RGB colour images
at
Pu ch

These formats are distinguished by 2 character

signatures as follows :
Te

Signature Image type Storage type

P1 Binary (PBM) ASCII

 Just by observing the header, we know that it is a PGM
P2 Grey scale (PGM) ASCII file. P2 indicates that it is a grey level image stored is
P3 RGB (PPM) ASCII ASCII. The 8 8 255 below the comment line indicates
that the size of the image is 8  8 and can have
P4 Binary (PBM) Raw bytes
256 grey levels i.e., each value is represented by 8-bits.
P5 Grey scale (PGM) Raw bytes
5.9 Image Processing, Computer
P6 RGB (PPM) Raw bytes Graphics and Computer Vision
Most often the storage type is ASCII.
 Before we end, one issue needs to be sorted out. A
 The header of a PGM, PBM or PPM file begins with the
very common question that people ask is what is the
appropriate signature followed by a blank line. There is
difference between image processing, computer
a comment line after the signature which starts with
graphics and computer vision ?
the # character. Next in the header are the width and
height of the image as ASCII decimal values.  Image processing deals with manipulation of images. A
input image is modified into a new image. Computer
 PBM files have no further header information. PGM
graphics deals with creation of images. In computer
and PPM files contain a third integer value, again in
graphics, models (2D or 3D) are created using
ASCII decimal form, representing the maximum
mathematical functions (Descriptors).
allowable pixel value.
 DSIP (MU-Sem. 7-COMP) 5-14 Introduction to Image Processing

 Computer vision which is also known as Machine chapters. This chapter forms the fundamental base
vision deals with the analysis of image content. required to understand image processing. Image acquisition
Computer vision is used to automate a process. forms the first step in any image processing system. In this
chapter, a basic block diagram of the image processing
Table 5.9.1 system is introduced. Each block in the system is explained
in detail. Important devices such as Vidicon cameras, CCD
Output Image Description
cameras and the frame grabber card are explained using
Input (Mathematical function) simple illustrations. Importance of each is stated.
Applications of line array CCD’s and area array CCD’s are
Image I.P. C.V.
presented. Basic concepts of image formats are also
Description C. G. – explained.
(Mathematical function)
Review Questions
From the Table 5.9.1,
Image processing  Input (Image) – Output (Image) Q. 1 What do we mean by image processing ?

ge
Computer graphics  Input (Description) – Output (Image) Q. 2 Why do we say that an image is a 2D-function ?

Computer vision  Input (Image) – Output (Description) Q. 3 Calculate the frequency of oscillation of red light.
io led
The material provided in this chapter is primarily basic Q. 4 Distinguish between image processing and graphics.
information which would be required in subsequent Q. 5 List the basic elements of an image processing
discussions. Our study of the human visual system, though system.
ic ow

not exhaustive, provides a basic idea of the capabilities of Q. 6 Explain working of a Vidicon camera.
the eye in perceiving pictorial information.
n
Q. 7 Explain line array and area array CCD cameras.
bl kn

Summary Q. 8 Write a short note on frame grabber card.

Q. 9 Explain the concept of half toning.
at
Pu ch

In this chapter, preliminary concepts of digital image Q. 10 Explain image formats.

processing are presented. Difference between one-
Q. 11 Advantages of CCD over Vidicon.
dimensional and two-dimensional signals is explained.
Te

Topics such as electromagnetic spectrum are discussed Q. 12 Explain the basics of quantum detector.
with examples. Q. 13 Explain gamma of a camera.
The concepts explained here will be found useful in Q. 14 How many grey levels will a half toned image have ?
understanding image processing algorithms in subsequent Explain in detail.

 6 Sampling and Quantization

Syllabus :
Sampling and Quantization. Representation of Digital Image

6.2 Sampling
6.1 Introduction
 We know that an image is basically a 2-dimensional MU - May 2012, May 2013, Dec. 2013.

ge
representation of the 3-dimensional world. We have Q. Write short note on : Sampling and Quantization.
also studied that images can be acquired using a
(May 2012, May 2013, Dec. 2013, 5 Marks)
io led
Vidicon or a CCD camera or using scanners.
 The basic requirement for image processing is that  Consider a band-limited signal. Although real world
images obtained be in the digital format. For example, signals are rarely band limited, we can produce one by
ic ow

one cannot work with or process photographs on passing the signal first through a low-pass filter known
identity cards unless he/she scans it using a scanner. as an anti-aliasing filter.
n
 The scanner digitizes the photograph and stores it on
Note : A signal is said to be band limited, if its Fourier
bl kn

the hard disk of the computer. Once this is done, one

can use image-processing techniques to modify the transform (FT), F(u), is zero outside a bounded region.
image as per requirement. In a Vidicon too, the output
at


Pu ch

(To understand sampling, we first need to understand

which is in analog form needs to be digitized in order
convolution involving impulse functions). If f(t) is a
to work with the images.
step signal shown in Fig. 6.2.1 and g(t) is a train of
Te

 To cut a long story short, to perform image processing, impulses (g(t) is also known as a COMB function), then
we need to have the images on the computer. This
the convolution of the two is given by f(t) * g(t). Let us
will only be possible when we digitize the analog
see what happens when we convolve f(t) with g(t).
pictures.
Proof : We know that the convolution of f(t) and g(t) is
 Now that we have understood the importance of
digitization, let us see what this term actually means.  
 The process of digitization involves two steps : y(t) =  f() g(t – ) d =  g() f(t – ) d
– –

y(t) =  [(t – T) + (t) +  (t + T)] f(t – )d
–
 
y(t) =  (t – T) f(t – ) d +  (t) f(t – ) d
– –
Fig. 6.1.1 : Steps of process of digitization 
In other words, Digitization = Sampling + Quantization. +  (t + T) f(t – ) d
 We have had exposure to these terms in the lower –
semesters in subjects like Principles of Communication 
Engineering, Signals and Systems and Signal Processing but,  (t – T) f(t – ) d = (t – T) * f(t)
which dealt with 1-dimensional signals. Let us take a –
brief look at these concepts and move ahead to the [This is impulse at t = –T convolved with f(t)]
2-dimensional domain.
 DSIP (MU-Sem. 7-COMP) 6-2 Sampling and Quantization

  As is seen in Fig. 6.2.2, if we use Fig. 6.2.2(c) as the

and,  (t + T) f(t – ) d =  (t + T) * f(t) sampling period, we encounter overlapping or aliasing
– in the frequency domain [Fig. (f)]. This is due to
[This is impulse at t = +T convolved with f(t)] undersampling. If we decrease the sampling period
(increase the sampling frequency) as in Fig. 6.2.2(g),
This gives us the figures as shown in Fig. 6.2.1.
we can eliminate aliasing in the frequency domain
[Fig. 6.2.2(h)].
 If we now multiply Fig. (h) with a gating function H(u)
[Fig. 6.2.2(i)] we can select only the required part of
the spectrum [Fig. 6.2.2(k)]. We take the inverse
Fourier transform to get back the original signal f(t)
[Fig. 6.2.2(j)].
 All this was possible because the sampling period was
small (sampling frequency was large). Hence to avoid
aliasing, the sampling frequency should be greater

ge
than or equal to 2  Bandwidth (2 w).
1
 In other words,  t  2w . This is the Whittaker-
io led
Shannon sampling theorem. Only if this condition is
met, can we recover the original signal. These 1-D
Fig. 6.2.1
concepts can be easily extended to the 2-D domain. A
ic ow

 We see that convolution replicates the signal f(t) at function f(x, y) is called a band limited function if its
every impulse g(t). Fourier transform F(u, v) is zero outside a bounded
n
 Another important property that one needs to know is region in the frequency plane i.e.
bl kn

that convolution in one domain is equal to F (u, v) = 0 u > u0, v > v0

multiplication in the other i.e. convolution in the time

at

Let us consider a band limited signal as shown in

Pu ch

domain is equal to multiplication in the frequency

Fig. 6.2.2.
domain and similarly multiplication in the time domain
is equal to convolution in the frequency domain.
Te

Fig. 6.2.2 contd….

 DSIP (MU-Sem. 7-COMP) 6-3 Sampling and Quantization

ge
io led
ic ow

Fig. 6.2.2

 Fig. 6.2.3 (a) and (b) represent the Fourier Transform of a band limited signal and its region of support (the base)
n
respectively.
bl kn
at
Pu ch
Te

Fig. 6.2.3(a) : Fourier transform Fig. 6.2.3(b) : Its region of Fig. 6.2.4 : Two dimensional
of a band limited signal support (the base) comb function

(a) 2-D comb function (b) Fourier transform of COMB function

Fig. 6.2.5
 DSIP (MU-Sem. 7-COMP) 6-4 Sampling and Quantization

 Just as in the case of a 1-D function, consider a 2-D  To avoid aliasing, the sampling frequency should be
comb function (sampling function) [Refer Fig. 6.2.4] twice the maximum frequency in the signal (image)
A 2-dimensional comb function is defined as 1
 x 
+ 2u0

Comb (x, y; x, y) 

=  (x – m x, y – n y) and y 
1
2v0
...(6.2.4)
m, n = – 

…(6.2.1) Where, 2u0 and 2v0 are the bandwidths in the u and v
directions respectively. This is known as the Nyquist
 The top view of the comb function is given in Fig. 6.2.4.
condition.
Another way of representing a 2-D function or a
1 1
sampling function is shown in Fig. 6.2.5(a). x = 2u and y = 2v are the lower bounds of the
0 0
 The Fourier transform of the comb function with
equation and are known as the Nyquist rate.
spacing of x and y is another COMB function with
spacing 1/x and 1/y. As seen from the Fig. 6.2.6, there is absolutely no

ge
aliasing. If the Nyquist condition is not met,
Comb (u, v) = FT {comb(x, y; x, y)}
overlapping of frequencies occurs and this is known as
= COMB (u, v; 1/x, 1/y) aliasing. This overlapping is shown in Fig. 6.2.7.
io led
The subsampled image is defined as
fs(x, y) = f(x, y)  comb (x, y; x, y)
+
ic ow

fs(x, y) =  f(m x, n y) (x – m x, y – n y)

m, n = – 
n
bl kn

…(6.2.2)
 Since multiplication in the time domain is equal to
at

convolution in the frequency domain, we get

Pu ch

Fs(u, v) = F(u, v) * COMB (u, v) ...(6.2.3)

 We know from the 1-D case, Fs (u, v) will simply be
Te

replicating F(u, v) at all impulses. Hence for a 2-D case,

using Fig. 6.2.3(b), we have,

Fig. 6.2.7 : Aliasing due to low sampling frequency

 The shaded portions are where overlapping (aliasing)
takes place.
 We observe aliasing when we go to watch movies.
Whenever the hero of the movie (could be the villain
too) races his car in the climax, have you ever
observed the wheels of his car ?
 When the car is slow, the wheels move in the same
direction as his car. But what happens when he speeds
up his car ? The wheels start moving in the opposite
direction. This is because; the camera, which is
shooting this scene, has a particular maximum
sampling frequency.
 As the car speeds up the Nyquist condition is not met
and hence there is aliasing, i.e. high frequencies start
looking as low frequencies.
Fig. 6.2.6
 DSIP (MU-Sem. 7-COMP) 6-5 Sampling and Quantization

(a) Rectangular (b) Triangular

Fig. 6.2.8(a) : Original image

ge
io led
(c) Hexagonal
Fig. 6.2.9 : Commonly used sampling grids
ic ow

6.3 Quantization
MU - May 2012, May 2013, Dec. 2013.
n
bl kn

Q. Write short note on : Quantization.

(May 2012, May 2013, Dec. 2013, 5 Marks)

at
Pu ch

 This is also a topic that has been widely discussed in

the lower semesters. The values obtained by sampling
Fig. 6.2.8(b) : Sampled image
Te

a continuous function usually comprise of an infinite

Coming back to the summation Equation (6.2.2), set of real numbers ranging from a minimum to a
 maximum depending upon the sensors calibration.
fs(x,y) =   f(m x, n y) (x – m x, y – n y)  These values must be represented by a finite number
m, n =– of bits usually used by a computer to store or process
The equation suggests that a sampled 2-D function any data. In practice, the sampled signal values are
represented by a finite set of integer values. This is
defined over a domain of size
known as quantization. Rounding of a number is a
M x  N y can be considered as a 2-D array simple example of quantization.
{f(0,0), f(0,1),……………….}  With these concepts of sampling and quantization, we
 This means that the function is represented as an now need to understand what these terms mean
M  N uniformly spaced sample. Apart from the when we look at an image on the computer monitor.
sampling interval in the x and y directions, another  Higher the spatial resolution of the image, greater is
important choice relevant to the image sampling is the the sampling rate i.e. lower is the image area x y
spatial arrangement of the sample points called represented by each sampled point. Similarly, higher
Tessellation. the grey level resolution (tonal resolution) more are
the number of quantized levels.
 We normally consider a comb function, which is a
 Hence spatial resolution gives us an indication of the
rectangular tessellation of grid. However other types
sampling while grey level resolution (tonal resolution)
of tessellations of the image plane, namely triangular
gives us an indication of the quantization.
and hexagonal, may also be used. Throughout this
Spatial resolution  Sampling
book though, we use rectangular tessellations.
Grey level resolution  Quantization
 DSIP (MU-Sem. 7-COMP) 6-6 Sampling and Quantization

 We have already stated that an image can be

considered as a 2-D array. Image f(x,y) is arranged in
the form of N  M array
f(0, 0) f(0, 1) ……………………….. f(0, M – 1)
 f(1, 0) f(1,f(2, 1)1) ……………………….. f(1, M – 1) 
f(x, y) =  f(2, 0) ……………………….. f(2, M – 1)

f(N –∶ 1,0) f(N – 1,1) ………………………..f(N – 1,∶ M – 1)
N M
 Hence every image that is seen on the monitor, is
actually this matrix. Each element of the matrix is
called a pixel. Never forget this. (b)

 Whenever we see an image on the screen of the Fig. 6.3.1 : The first row of the image is plotted.

computer it is actually a matrix which consists of  Similarly, if we have 8 bits to represent the grey levels,
8
N  M pixels and each pixel is considered to be a we will have 256 grey levels (2 ). Hence more the bits,

ge
sample. Hence more the pixels, more the samples, more are the grey levels and better is the tonal clarity
higher the sampling rate and hence better the spatial (quantization). The total size of the image is
resolution. The value of each pixel is known as the grey N  M  m, where m is the number of bits used.
io led
level.  The x-axis is the number of samples (pixels) in the row
 The computer understands only ones and zeros. Hence (sampling) while the y-axis is the grey level of each
these grey levels need to be represented in terms of sample (quantization)
ic ow

zeros and ones. If we have two bits to represent the  Now, comes the obvious question. As we know, more
2
n
grey levels, only 4 different grey levels (2 ) can be the samples and the bits, better is the image. What
bl kn

identified viz. 00, 01 10, 11, where 00 is black, 11 is then should be the ideal values of sampling and
white and the other two are different shades of grey. quantization.
at


Pu ch

This answer will vary from image to image. Given

below is a table of sampling and the quantization
values. As the sampling and the quantization increase,
Te

the number of bits required to store the image

increases tremendously.
 The clarity increases, but storage space required
increases too. We hence need to get a trade-off
between the two. For simplicity, we consider a square
matrix of size N  N.
Fig. 6.3.1(a)

Table 6.3.1 : Number of storage bits for various values of N and m

m 1 2 3 4 5 6 7 8
N

32 1,024 2,048 3,072 4,098 5,120 6,144 7,168 8,192

64 4,096 8,192 12,288 16,384 20,480 24,576 28,672 32,768

128 16,385 32,768 49,152 65,536 81,290 98,304 114,688 131,072

256 65,536 131,072 196,608 262,144 327,680 393,216 458,752 524,288

512 262,144 524,288 786,432 1,048,576 1,310,720 1,572,864 1,835,008 2,097,152

1,024 1,048,576 2,097,152 3,145,728 4,194,304 5,242,880 6,291,456 7,340,032 8,388,608

We shall see the effects of reducing quantization levels and reducing spatial resolution separately.
 DSIP (MU-Sem. 7-COMP) 6-7 Sampling and Quantization

MATLAB code for reducing quantization levels

%% Effects of reducing the quantization values %%
clear all
clc
a=imread(‘zebra.tif’);
a=double(a);
a=a+1;
b=max(max(a)); %% gives us the maximum value in the image
i=input (‘how many bits do you want 1 2 4 8’);
j=b/(2^i); %% since total number of bits is equal to 2^i
F=floor(a/(j+1));
F1=(F*255)/max(max(F)); % normalizing %
figure(1)
imshow(uint8(a))

ge
figure(2)
imshow(uint8(F1)) io led
[row,col]=size(a);
i=1; j=1;
for x=1:2:row
ic ow

for y=1:2:col
n
c(i,j)=a(x,y);
bl kn

j=j+1;
end
at
Pu ch

j= 1; %% This needs to be done else the value of j goes

on increasing %%
(a) Original image (b) Image using 4-bits i= i+1;
Te

end
figure(1),imshow(a)
figure(2),imshow(c)
figure(3),
imagesc(a),colormap(gray)
figure(4)
imagesc(c),colormap(gray)

(c) Image using 2-bits (d) Image using 1-bit

Fig. 6.3.2

Comparing the images, we see that “false contouring”

takes place as we reduce the number of grey levels.
MATLAB code for reducing spatial resolution.
%% Down sampling
%% To see the effects of reducing the number of samples
%%
clear all (a) Original image (b) Down sampled
clc Fig. 6.3.3
a= imread(‘deepa.tif’)
 DSIP (MU-Sem. 7-COMP) 6-8 Sampling and Quantization

Fig. 6.3.3 (c) Down sampled image displayed after

zooming to match the size of the original image

It is clear from the images that the resolution reduces

ge
as number of samples reduce. To compare and understand
the actual effects, we plot them together. To make sure
io led
that they appear to be of the same size, we upsample the
second image. Comparing Fig. 6.3.3(a) and Fig. 6.3.3(c) we
see that the second image has a “Checker board” pattern
ic ow

due to the reduction of samples.

6.4 Isopreference Curves

n
bl kn

 We have seen the effects of reducing N and m in the

preceding topic. We still do not know what is the ideal
at
Pu ch

value of N (sampling resolution) and m (tonal

resolution) for images.
 An early study by T.S. Huang in 1965 attempted to
Te

quantify experimentally the effects on image quality

produced by varying N and m simultaneously. There,
different types of images were shown to a group of
people.
 The first image was that which did not have a lot of
details for example a woman’s face. The second image
was one which had intermediate amount of
information for example a small group standing
together and the third image was one which had a lot
of details example an image of a crowd.
 In these images, N and m were varied. Observers were
then asked to rank them according to their subjective
quality. These results were summarised in the form of
isopreference curves in the N-m plane.
 Points lying on an isopreference curve corresponded
to image of equal subjective quality. From the
isopreference curves Huang concluded that images
with large amount of details require few grey levels.
Since the isopreference curve of the crowd is near
vertical, it means that for a fixed value of N, the image Fig. 6.4.1
is nearly independent of m.
 DSIP (MU-Sem. 7-COMP) 6-9 Sampling and Quantization

1.6 inch by 1.2 inch when displayed or printed at

6.5 Non-Uniform Sampling
400 pixels per inch.
 We have already seen that as N and m increase, the
spatial and the grey level resolution increase. But due
to this the storage space required increases
enormously.
 To address this problem, we use non-uniform
sampling. In this, regions which have a lot of details
i.e., were there are abrupt changes in grey levels, a lot
of samples are taken while regions where the grey
levels reduce very slowly (background) only few
samples are used.
 Hence non-uniform sampling can be used to reduce
Fig. 6.6.1
the storage space required without distorting the

ge
subjective quality of the image.  Another term that we need to understand is the
 We have all seen non-uniform sampling been used. Aspect ratio.
Whenever a person’s identity needs to be concealed  The ratio of the image’s width to its height, measured
io led
on T.V., non-uniform sampling is used. in unit length or number of pixels is referred to as its
 Very few samples are taken near the eye regions aspect ratio. Both, a 3  3 inch image and a 128  128
because of which we see blocks near the person’s eye image have the same aspect ratio of 1.
ic ow

and hence can not make out who that person is. X
Aspect ratio =
Y
n
6.6 Physical Resolution
bl kn

 By now we know that a digital image, or image for 6.6.1 Solved Examples on Physical
short, is composed of discrete pixels. These pixels are Resolution
at
Pu ch

arranged in a row and column fashion to form a

rectangular picture area. Ex. 6.6.1 : Compute the physical size of a 640  480 image

Te

Clearly, the total number of pixels in an image is a when printed by a printer at 240 pixels per inch.
function of size of the image and the number of pixels Soln. : Since we have 240 pixels per inch, the physical size
per unit area (example: inch) in the horizontal as well 640 480
as the vertical direction. of the image is 240 by 240.
 The number of pixels per unit length is referred to as
the resolution of the displaying device (most of the Ex. 6.6.2 : If we want to resize a 1024  768 image to one
deskjet printers have 670 dpi i.e., 670 dots per inch.). that is 600 pixels wide with the same aspect ratio as the
 Thus a 3  2 inch image at a resolution of 300 pixels original image, what should be the height of the resized
per inch would have a total of 540,000 pixels!! image?
 In most books as well as in this book, image size is Soln. : We know
given as the total number of pixels in the horizontal
Width
direction times the total number of pixels in the Aspect ratio =
Height
vertical direction (example : 128  128, 512  512,
640  480 ....). For the original image the
 Although this convention makes it relatively straight Aspect ratio is = 1024/768 = 1.33
forward to gauge the total number of pixels in an Now for the resized image, we want the same aspect
image, it does not specify the physical size of the ratio but a width of 600 pixels.
image or its resolution as defined in the paragraph
Width 600
above. Height = = = 451
Aspect ratio 1.33
 A 640  480 image would measure 6.66 inches by
Hence the resized image will be 600  451.
5 inches when displayed or printed at 96 pixels per
inch. On the other hand, it would measure only
 DSIP (MU-Sem. 7-COMP) 6-10 Sampling and Quantization

Ex. 6.6.3 : How much storage capacity is required to store Ex. 6.6.5 : We have a C. T. image which we want to transmit
an image with size of 1024  768 and 256 gray levels ? through a voice grade telephone line. The CT image is of
Soln. : size 512  512 and has 256 grey levels. During transmission
Storage capacity required = A  B  C each byte is preceded by a start bit and succeeded by a stop
Here A  B = Physical size of image = 1024  768 bit. How many minutes will it take to transmit the image ?

C = Number of bits required to get 256 gray (Voice grade telephone lines can transmit 9600-bits/sec)
8
levels, we must have 8 bits (2 = 256) Soln. : 256 grey level required mean 8-bits.

Storage capacity required = 1024  768  8 = 6291456  one packet = [8 + 02] bits = 10 bits
bits  image size = 512  512  10
Ex. 6.6.4 : A common measure of transmission for digital  image size = 2621440
data is the baud rate, defined as the number of bits 512  512  10
 time taken =
transmitted per second. Transmission is accomplished in 9600
packets consisting of a start bit, a byte (8-bits) of information time taken = 273.06 sec 4.5 min

ge
and a stop bit.
(a) How many minutes would it take to transmit a Summary
io led
1024  1024 image with 256 grey levels if we use a
56 k baud modem ? This chapter deals with converting a scene that is
continuous in time and space into an image that can be
(b) What would be the time required if we use a 750 k
ic ow

processed by computers. The technique of converting a

band transmission line ? continuous signal into a discrete one is called digitization
Soln. : and is explained here. Digitisation comprises of two steps
n
viz. sampling and quantization. Concepts of sampling are
bl kn

(a) Since we have 256 grey levels, we need 8-bits for

representing each pixel. explained in the one-dimensional as well as in the two
dimensional domain. Relationship between the time
at

Along with these 8-bits, we also have a start bit and a

Pu ch

domain and the frequency domain i.e., the Fourier domain

stop bit. are presented here.
Hence we have (8 + 2) bits per pixel. Aliasing, which forms an important pitfall of sampling,
Te

total number of bits for transmission are has been introduced and the importance of the Nyquist
criteria has been explained. The concept of an image being
N = 1024  1024  10
stored as a matrix has been introduced here. Spatial
N = 10485760-bits resolution and grey level resolution are also explained using
These bits are transmitted at 56 k baud. MATLAB codes.
N
time taken = 3 Review Questions
56  10
time taken = 187.25 sec  3.1 minutes
Q. 1 Explain the concept of aliasing for 1-dimensional
(b) Now if we use a faster transmission line of 750 k baud
signals.
rate, then
Q. 2 Explain the concept of aliasing for 2-dimensional
Time taken to transmit the image
signals.
N 10485760
= Q. 3 How does one avoid aliasing ?
750  10
3 3
750  10
Q. 4 Explain sampling and quantization.
= 13.98 sec  14 sec
Q. 5 Explain the effects of reducing sampling and
quantization.
Q. 6 Explain isopreference curves.
Q. 7 Explain non-uniform sampling.


 Image Enhancement in
7 Spatial Domain

Syllabus :
Gray Level Transformations, Zero Memory Point Operations, Neighborhood Processing, Spatial Filtering,
Smoothing and Sharpening Filters, Median Filter.

7.2 Spatial Domain Methods

ge
7.1 Introduction
 Image enhancement is one of the first steps in image  The term spatial domain means working in the given
io led
processing. As the name suggests, in this technique, space, in this case, the image. It implies working with
the original image is processed so that the resultant the pixel values or in other words, working directly
image is more suitable than the original for specific with the raw data.
ic ow

applications i.e. the image is enhanced.

 Let f(x, y) be the original image where f is the grey
 Image enhancement is a purely subjective processing level value and (x, y) are the image coordinates. For a
n
technique. By subjective we mean that the desired 8-bit image, f can take values from 0 - 255 where 0
bl kn

result varies from person to person. represents black, 255 represents white and all the
 An image enhancement technique used to process intermediate values represent shades of grey.
at
Pu ch

images might be excellent for a person, but the same In a image of size 256  256 as shown in the Fig. 7.2.1.
result might not be good enough for another. It is also
Te

important to know at the outset that image

enhancement is a cosmetic procedure i.e. it does not
add any extra information to the original image.
 It merely improves the subjective quality of the images
by working with the existing data.
 Image enhancement can be done in two domains :
(1) The spatial domain
(2) The frequency domain.
 Let us start with explaining what is meant by the
spatial domain, discuss the various methods of spatial
domain enhancement and then move on to discuss the
frequency domain technique in chapter of Image Fig. 7.2.1
Enhancement in Frequency Domain. Note : We always consider the axis as shown in
Fig 7.2.1. The x-axis is taken in the downward vertical
direction, while the y-axis is taken in the horizontal
direction. The reason for taking this inverted coordinate
system is to make sure that the image uses the standard
matrix coordinates that we have been using right from our
school. You will remember the first value of the matrix is
Fig. 7.1.1 always at the top left corner while the last value of the
matrix is at the right bottom corner.
 DSIP (MU-Sem. 7-COMP) 7-2 Image Enhancement in Spatial Domain

 The modified image can be expressed as,  Since this operation does not require past pixel
g (x, y) = T [f(x, y)] …(7.2.1) values, we do not need additional memory allocation.
Hence point processing is also known as zero memory
Here f(x, y) is the original image and T is the
operation. Some of the common examples of point
transformation applied to it to get a new modified
processing are :
image g(x, y). For all spatial domain techniques it is
simply T that changes. The general equation remains (1) Digital negative
the same. (2) Contrast stretching
 To summarize, spatial domain techniques are those, (3) Thresholding
which directly work with the pixel values to get a new (4) Grey level slicing
image based on Equation (7.2.1).
(5) Bit plane slicing
 Spatial domain enhancement can be carried out in two
(6) Dynamic range compression
different ways :
(7) Power law transformation
(1) Point processing

ge
(2) Neighbourhood processing
 Before we proceed to explain the following examples,
it is important to understand the identity
io led
transformation. The identity transformation is given in
the Fig. 7.3.1.

ic ow

In the Fig. 7.3.1, the solid line is the transformation T.

The horizontal axis represents the grey scale of the
n
input image (r) and the vertical axis represents the
bl kn

grey scale of the output image (s). It is called an

identity transformation because it does not modify the
at

new image at all !!.

Pu ch

 As seen, the grey level 10 is modified to 10, 125 to 125

and finally 255 to 255. In general s1 = r1. Fig. 7.3.1 will
Te

Fig. 7.2.2 help us understand the point processing techniques

better.
7.3 Point Processing / Zero Memory
Operation

MU - Dec. 2015, Dec 2016, May 2017,

May 2018, Dec 2018

Q. What do you understand by zero memory operation.

(Dec. 2015, 5 Marks)
Q. Explain any five zero memory operations.
(Dec. 2016, May 2017, 10 Marks)
Q. Explain Thresholding.
(Dec. 2016, May 2017, 2 Marks)
Q. Explain zero memory operations.
(May 2018, Dec. 2018, 5 Marks) Fig. 7.3.1
 In point processing, we work with single pixels i.e. T is
(1) Digital negative
1  1 operator. It means that the new value f(x, y)
depends on the operator T and the present f(x, y). This  Digital negatives are useful in a lot of applications. A
statement will be clear as we start giving some common example of digital negative is the displaying
examples. of an X-ray image.
 DSIP (MU-Sem. 7-COMP) 7-3 Image Enhancement in Spatial Domain

 As the name suggests, negative simply means inverting

the grey levels i.e. black in the original image will now
look white and vice versa. The Fig. 7.3.2 is the digital
negative transformation for a 8-bit image.

(a) Original image

ge
io led
Fig. 7.3.2

 The digital negative image can be obtained by using a

ic ow

simple transformation given by,

s = 255 – r (rmax = 255)
n
bl kn

Hence when
r = 0, s = 255 and when r = 255, s = 0.
at
Pu ch

In general, (b) Digital negative

s = (L – 1) – r …(7.3.1) Fig. 7.3.3
Te

Here L is the number of grey levels. (256 in this case)

(2) Contrast stretching
MATLAB program for finding the digital negative
 Many times we obtain low contrast images due to
%% MATLAB code to calculate negative %% poor illuminations or due to wrong setting of the lens
clear all aperture. The idea behind contrast stretching is to
clc increase the contrast of the images by making the dark
aa=imread (‘saturn.tif’); portions darker and the bright portions brighter.

a=double (aa)  Fig. 7.3.4 shows the transformation used to achieve

contrast stretching. In the Fig. 7.3.4, the dotted line
c=255; % for a 8-bit image %
indicates the identity transformation and the solid line
b=c–a; is the contrast stretching transformation.
figure(1)  As is evident from the Fig. 7.3.4, we make the dark
colormap(gray) grey levels darker by assigning a slope of less than one
imagesc(a) and make the bright grey levels brighter by assigning a
slope greater than one.
figure(2)
 One can assign different slopes depending on the input
colormap(gray)
image and the application. As was mentioned, image
imagesc(b)
enhancement is a subjective technique and hence
there is no one set of slope values that would yield the
desired result.
 DSIP (MU-Sem. 7-COMP) 7-4 Image Enhancement in Spatial Domain

end
end
subplot(2,1,1)
imshow(uint8(a))
title(‘Original Image’);
subplot(2,1,2)
imshow(uint8(b))
title(‘Image after Contrast Stretching’)

Fig. 7.3.4 : Original grey level r

ge
 The formulation of the contrast-stretching algorithm is
given as follows :
 lr 0r<a
io led
s =  m  (r – a) + v ar<b …(7.3.2)
 n  (r – b) + w br<L–1
where l, m and n are the slopes. It is clear from the
ic ow

Fig. 7.3.4 that l and n are less than one while m is greater (a) Original image (b) Contrast stretched image
than one. The contrast stretching transformation increases
n
Fig. 7.3.5
the dynamic range of the modified image.
bl kn

(3) Thresholding
MATLAB program for contrast stretching

at

Extreme contrast stretching yields thresholding. If we

Pu ch

%% Contrast stretching of an image %%

observe the contrast stretching diagram closely we
% Slopes taken are 0.5, 2 and 0.5 % notice that if the first and the last slope are made zero,
Te

clear all; and the centre slope is increased, we would get a

thresholding transformation i.e. if
clc;
r1 = r2, s1 = 0 and s2 = L – 1, we get a thresholding
a=imread(‘xray1.tif’) function. It is shown in Fig. 7.3.6.
a=double(a);  The formula for achieving thresholding is as follows,
[row col]=size(a);
LT=input(‘Enter the lower threshold value:’); s = 0; if r  a 
 …(7.3.3)
s=L–1; if r > a 
UT = input(‘Enter the upper threshold value:’);
for x=1:1:row
for y=1:1:col
if a(x,y)<=LT
b(x,y)=0.5*a(x,y);
else if a(x,y)<=UT
b(x,y)=2*(a(x,y)-LT)+0.5*LT;
else b(x,y)=0.5*(a(x,y)-UT)+0.5*LT+2*(UT-LT);
end
end
Fig. 7.3.6 : Original grey level r
 DSIP (MU-Sem. 7-COMP) 7-5 Image Enhancement in Spatial Domain

Where, L is the number of grey levels.

 As mentioned earlier, image enhancement being a
subjective phenomena, the value of a will vary from
image to image and from person to person. The
objective is to identify the region that he or she is
interested in. An important thing to note is that a
thresholded image has the maximum contrast as it has
only black and white grey values.
MATLAB program for thresholding
%%% Thresholding %%%
clear all
clc (a) Slicing without background

p=imread(‘spine.tif’);
a=p;

ge
[row col]=size(a);
T=input(‘Enter value of Threshold :’)
io led
for i=1:1:row
for j=1:1:col
if(p(i,j)<T)
ic ow

a(i,j)=0;
else
n
a(i,j)=255;
bl kn

end
end (b) Slicing with background
at
Pu ch

end Fig. 7.3.8

figure(1),imshow(p),
This can be implemented using the formulation
figure(2),imshow(a)
Te

s = L – 1; if a  r  b 
 …(7.3.4)
s=0; otherwise 
 This method is known as Grey level slicing without
background. This is because in this process, we have
completely lost the background. In some applications,
we not only need to enhance a band of grey levels but
also need to retain the background.
(a) Original image of spine (b) Image obtained
 This technique of retaining the background is called as
using threshold
Grey-level slicing with background. The transformation
Fig. 7.3.7
is as shown in the Fig. 7.3.8(b).
(4) Grey level slicing (Intensity slicing) The formulation for this is
 What thresholding does is it splits the grey level into s = L – 1; if a  r  b 
 …(7.3.5)
two parts. At times, we need to highlight a specific s=r; otherwise 
range of grey values like for example enhancing the MATLAB program for grey level slicing with and
flaws in an X-ray or a CT image. without background is as follows :
 In such circumstances, we use a transformation known %% Grey level slicing without background %%
as grey level slicing. The transformation is shown in clear all
the Fig. 7.3.8(a). It looks similar to the thresholding clc
function except that here we select a band of grey p=imread (‘skull.tif’);
level values. z=double (p);
 DSIP (MU-Sem. 7-COMP) 7-6 Image Enhancement in Spatial Domain

[row,col]=size(z)
for i=1:1:row
for j=1:1:col
if((z(i,j)>50))&&(z(i,j)<150)
z(i,j)=255;
else
z(i,j)=0; (a) Original image (b) Grey level slicing

end with background

end Fig. 7.3.10

end (5) Bit plane slicing

figure (1); % …...original image.
 In this technique, we find out the contribution made
imahow (p)
by each bit to the final image. As mentioned earlier, an
figure (2); % …...gray level slicing without background image is defined as say a

ge
imshow (uint8(z)) 256  256  8 image. In this, 256  256 is the number
io led of pixels present in the image and 8 is the number of
bits required to represent each pixel. 8-bits simply
means 28 or 256 grey levels.
 Now each pixel will be represented by 8-bits. For
ic ow

example black is represented as 00000000 and white is

represented as 11111111 and between them, 254 grey
n
(a) Original image (b) Gray level slicing levels are accommodated. In bit plane slicing, we see
bl kn

without background the importance of each bit in the final image. This can
Fig. 7.3.9 be done as follows.
at
Pu ch

MATLAB program  Consider the LSB value of each pixel and plot the
%% Grey level slicing with backgound %% image using only the LSBs. Continue doing this for each
bit till we come to the MSB. Note that we will get
Te

clear all
8 different images and all the 8 images will be binary.
clc
p=imread(‘skull.tif’); Solved Example
z=double(p);
Ex. 7.3.1 : Given a 3  3 image, plot its bit planes.
[row col]=size(p);
for i=1:1:row 1 2 0
for j=1:1:col 4 3 2
if(z(i,j)>50))&&(z(i,j)<150)
7 5 2
z(i,j)=255;
Soln. : Since 7 is the maximum grey level, we need only
else
3-bits to represent the grey levels.
z(i,j)=p(i,j);
end Hence we will have 3-bit planes. Converting the image
to binary we get,
end
end 001 010 000 1 0 0 0 1 0 0 0 0
figure (1); %------- original image.
imshow(p) 100 011 010 0 1 0 0 1 1 1 0 0
figure (2); %------- grey level slicing with background 111 101 010 1 1 0 1 0 1 1 1 0
imshow (uint8(z))
Binary image LSB plane Middle bit MSB
plane plane
 DSIP (MU-Sem. 7-COMP) 7-7 Image Enhancement in Spatial Domain

MATLAB program Observing the images we come to the conclusion that

%% MATLAB code for bit extraction %% the higher order bits contain majority of the visually
clear all significant data, while the lower bits contain the suitable
details in the image. Bit plane slicing can hence be used in
clc
image compression. We can transmit only the higher order
a=imread(‘warne.tif’);
bits and remove the lower order bits. Bit plane slicing is
a=double(a);
also used in steganography.
r=input(’which bit image do you want to see 1=MSB
8=LSB’); (6) Dynamic range compression (Log transformation)
[row col]=size(a);  At times, the dynamic range of the image exceeds the
for x=1:1:row capability of the display device. What happens is that
for y=1:1:col some pixel values are so large that the other low value
c=dec2bin(a(x,y),8); % converts decimal to pixels get obscured.
binary
 A simple day-to-day example of such a phenomena is

ge
d=c(r); that during daytime, we cannot see the stars. The
w(x,y)=double(d); %% since w is a char and reason behind this is that the intensity of the sun is so
cannot be plotted
io led large and that of the stars is so low that the eye cannot
if w(x,y)==49 %% since double of d will be either 49 adjust to such a large dynamic range.
or 48
 In image processing, a classic example of such large
w(x,y)=255;
ic ow

differences in grey levels is the Fourier spectrum (will

else
be discussed in detail in the frequency domain
w(x,y)=0;
n
enhancement technique).
bl kn

end
 In the Fourier spectrum only some of the values are
end
very large while most of the values are too small. The
at

end
Pu ch

6
dynamic range of pixels is of the order of 10 . Hence,
figure(1) when we plot the Fourier spectrum, we see only small
imshow(uint8(a)) dots, which represent the large values.
Te

figure(2)  Something needs to be done to be able to see the

imshow(uint8(w)) small values as well. This technique of compressing the
dynamic range is known as dynamic range
compression.
 We all know that the log operator is an excellent
compressing function. Hence the dynamic range
compression is achieved by using a log operator. C is
the normalization constant.

Fig. P. 7.3.1 : Eight images, each representing Fig. 7.3.11

contribution of a single bit
 DSIP (MU-Sem. 7-COMP) 7-8 Image Enhancement in Spatial Domain

MATLAB program for dynamic range compression  The power law transformation can also be used to
%% Dynamic range compression %% improve the dynamic range of an image. Given below
clear all is the MATLAB code for power transformation. The
clc final image has been normalized to 0 - 255 range.
aa=imread(‘saturn.tif’);
a=double(aa) ;
[row,col]=size(a);
for x=1:1:row
for y=1:1:col
c(x,y)=a(x,y)*((-1)^(x+y)); %% Needed to center the
transform
end
end

ge
d=abs(fft2(c));
Fig. 7.3.13 : Gamma correction factor
d_log=log(1+d);
%%% Plotting %% Power transformation %%
io led
figure(1) clear all
colormap(gray) clc
imagesc(d)
ic ow

img1=imread(‘test.tif’);
figure(2) [row,col]=size(img1);
n
colormap(gray)
gamma=input(‘Enter the correction factor: ’);
bl kn

imagesc(d_log)
img=double(img1);
at
Pu ch

for i=1:row
for j=1:col
Te

nuimg(i,j)=img(i,j)^gamma
end
(a) (b) end
Fig. 7.3.12 : Dynamic range compression numax=max(max(nuimg));

(7) Power law transformation

numin=min(min(nuimg));
n=255/(numax-numin);
 The basic formula for power-law transformation is
for i=1:row

g(x, y) = c  f(x, y) for j=1:col

It can also be written as s = cr ...(7.3.6) nuimg1(i,j)=n*(nuimg(i,j)-numin); % Normalisation

 Here c and  are positive constants. The end

transformation is shown Fig. 7.3.13 for different values end
of  which is also called the gamma correction factor. nuimg2=uint8(nuimg1);
We observe that by changing the value of gamma, we subplot(2,1,1)
obtain a family of transformation curves. imshow(img1)
 Non-Linearities encountered during image capturing, title(‘Original image’)
printing and displaying can be corrected using gamma subplot(2,1,2)
correction. Hence gamma correction is important if
imshow(nuimg2)
the image needs to be displayed on the computer.
title(‘Image after power transformation’)
 DSIP (MU-Sem. 7-COMP) 7-9 Image Enhancement in Spatial Domain

 To achieve neighbourhood processing, we place this

3  3 (it could also be a 5  5 or a 7  7...) mask on the
image, multiply each component of the mask with the
corresponding value of the image, add them up and
place the value that we get, at the center.
 This operation is the same as convolution. In the
convolution operation, we have two signals. Of the
two signals, we take one, flip it and then move it
Fig. 7.3.14
across the other signal step by step. The same thing is
done here. We don’t need to flip the mask as it is
7.4 Neighbourhood Processing symmetric.
 This, as mentioned before, is also a spatial domain  If f is the original image and g is the modified image,
technique in image enhancement. Unlike the point then,
processing techniques where we consider one pixel at g(x, y) = f(x – 1, y – 1)  w1 + f(x – 1, y)  w2
+ f(x – 1, y + 1)  w3 + f(x, y – 1)  w4

ge
a time and modify it depending on our requirement,
here we not only consider a pixel but also its + f(x, y)  w5 + ….……+ f(x + 1, y + 1)  w9
immediate neighbours. ...(7.4.1)
io led
 To cut a long story short, we change the value of the
pixel f(x, y) based on the values of its 8 neighbours as
shown in Fig. 7.4.1. Instead of a 3  3 neighbourhood,
ic ow

we could also use a 5  5 or a 7  7….. neighbourhood.

n
bl kn
at
Pu ch

Fig. 7.4.3

 Once g(x, y) is calculated, we shift the mask by one

Te

step towards the right to the next pixel. Now w5

coincides with f(x, y + 1).
 One of the important operations that can be achieved
using neighbourhood processing is that of image
filtering.
 We can perform low pass, high pass and band pass
Fig. 7.4.1 : A 3  3 neighbourhood filtering using neighbourhood operations.
 Before explaining the procedure of performing filtering
on images, it is imperative to understand what we
mean by frequencies in an image !! We are all well
versed with frequencies in 1-dimensional signals.
Given two signals, we can easily distinguish between
the lower frequency signal and the higher frequency
signal. Refer Fig. 7.4.3.

Fig. 7.4.2 : 3  3 mask

 There are a lot of things that can be achieved by
neighbourhood processing which are not possible with
point processing. Fig. 7.4.2 shown is called a mask or a
window or a template. Fig. 7.4.4
 DSIP (MU-Sem. 7-COMP) 7-10 Image Enhancement in Spatial Domain

 We can conclude that the lower signal is of a much (1) Gaussian noise : The Probability Density Function
higher frequency, by checking the number of (PDF) of Gaussian noise is given by the formula,
2 2
oscillations. That is, higher the frequency, greater are – (z –) / 2
1e
the number of oscillations. p(z) =
2 
 If the two signals represent voltages, then, how fast
z  Grey level
the voltages change is an indication of the frequency.
  Mean of average value of z
The same concepts can be applied to images.
  Standard deviation
 In images, instead of voltages, we have grey levels. If

the grey levels change slowly over a region, it is   Variance
considered to have low frequency, whereas, if the grey
levels change very rapidly, that region is considered to
have high frequencies.
 Hence in images, regions where the grey levels change

ge
slowly are low frequency areas and regions which have
abrupt grey level changes, are high frequency areas.
Always remember this.
io led
 In most of the images, the background is considered to
be a low frequency region, whereas the edges are Fig. 7.4.5
considered to be high frequency regions. Hence low
ic ow

If we plot this function, we notice that,

pass filtering implies removing (bluring) the edges
while high pass filtering implies removing the 70% of its value lies in the range [( – ), ( + )] and
n
background. 95% of its value lies in the range [( – 2), ( + 2)]
bl kn

Gaussian noise has a maximum value at  and then it

7.4.1 Low Pass Filtering (Smoothing)
at

starts falling off.

Pu ch

 Low pass filtering as the name suggests removes the Consider the image shown in Fig. 7.4.6.
high frequency content from the image. It is used to
Te

remove noise present in the image.

 Noise, is normally a high frequency signal and low pass
filtering eliminates the noise. Before proceeding to
explain the low-pass filtering technique, we shall
spend some time discussing the types of noise that are
fairly common in images.
Fig. 7.4.6 Fig. 7.4.7

7.4.2 Noise
The principal sources of noise in a digital image arise
during image acquisition and during transmission. No
matter how much care one takes, some amount of noise
always creeps in. Based on the shapes (Probability Density
Functions) of the noise, they are classified as :
(1) Gaussian noise (2) Salt and pepper noise Fig. 7.4.8

(3) Rayleigh noise (4) Gamma noise There are two constant regions in the image. Hence,
the histogram of the image is as shown in Fig. 7.4.7. If in
(5) Exponential noise (6) Uniform noise
this image, Gaussian noise creeps in, the histogram gets
Of these, the first two are more common than the
modified as shown in Fig. 7.4.8. Gaussian noise arise in an
others. We shall explain the Gaussian and salt and pepper
image due to factors such as circuit noise, sensor noise
noise in this chapter.
poor illumination and high temperature.
 DSIP (MU-Sem. 7-COMP) 7-11 Image Enhancement in Spatial Domain

(2) Salt and Pepper Noise

7.4.3 Low Pass Averaging Filter
The PDF of the salt and pepper noise (bipolar noise) is
MU - May 2017
Pa; for z = a
p(z) = Pb; for z = b
Q. Explain lowpass spatial filtering.
 0; otherwise
(May 2017, 5 Marks)
If Pa or Pb is zero, this noise is called unipolar noise.
The PDF of salt and pepper is shown in Fig. 7.4.9.  If an image has Gaussian noise present in it, we use a
low pass averaging filter to eliminate the noise. The
frequency response of the low-pass filter along with its
spatial response is shown in Fig. 7.4.11.

ge
Fig. 7.4.9
Generally a and b are black and white grey levels
io led
respectively. Hence for a 8-bit image, a = 0, b = 255
because of which the noise is called salt (white) and pepper
(black). Some books refer to it as speckle noise.
ic ow

Take the same image as the one taken for the

n
Gaussian example.
bl kn

Fig. 7.4.11

 From the spatial response we generate the mask that

at
Pu ch

would give us the low pass filtering operation. One

important thing to note from the spatial response is
Te

that all the coefficients are positive. The standard low

pass averaging mask (3  3) is given above. As the
name suggests, each element of the mask is the
Fig. 7.4.10 (a)
average value.
When salt and pepper creeps in, the image looks like
 We could also use a 5  5 or a 7  7 mask as per our
requirement.

Fig. 7.4.10 (b)

Salt and pepper noise creeps into images in situations

Fig. 7.4.12
where quick transients, such as faulty switching take place.
 Let us take an example of a 3  3 mask on a pseudo-
image and see how it eliminates the edges. Consider a
8  8 size image. It is clear that the image has a single
edge between 10 and 50. To get rid of this edge (high
frequency) we use a 3  3 averaging mask.
 DSIP (MU-Sem. 7-COMP) 7-12 Image Enhancement in Spatial Domain

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

ge
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10
io led 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
ic ow

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
n
bl kn

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

50 50 50 50 50 50 50 50
at
Pu ch

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
 We place a 3  3 mask on this image. We start from
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
the left hand top corner. We cannot work with the
Te

borders and hence are normally left as they are. 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

 We then multiply each component of the image with 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

the corresponding value of the mask. Since all the nine
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
values of the image are 10, the average is also ten and
the centre pixel (the underlined pixel) remains ten. 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
 We now shift the mask towards the right till we reach 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
the end of the line and then move it downwards. Some
of the mask positions are shown here. 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

Note : The resultant should be written in a new matrix The last one in the sequence being
(image). 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
50 50 50 50 50 50 50 50
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50
50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50
 DSIP (MU-Sem. 7-COMP) 7-13 Image Enhancement in Spatial Domain

The result of convolving the image with the 3  3 a1(x,y)=w(1)*a(x-1,y-1)+w(2)*a

averaging mask is shown. (x-1,y)+w(3)*…
a(x-1,y+1)+w(4)*a
10 10 10 10 10 10 10 10
(x,y-1)+w(5)*a(x,y)+w(6)*…
10 10 10 10 10 10 10 10 a(x,y+1)+w(7)*a(x+1,y-
10 10 10 10 10 10 10 10 1)+w(8)*a(x+1,y)+w(9)*…
a(x+1,y+1);
23.3 23.3 23.3 23.3 23.3 23.3 23.3 23.3
end
36.6 36.6 36.6 36.6 36.6 36.6 36.6 36.6 end
50 50 50 50 50 50 50 50 figure (1)
50 50 50 50 50 50 50 50 imshow (uint8 (a))
figure (2)
50 50 50 50 50 50 50 50
imshow (uint8 (a1))
 As we notice, the low frequency regions have

ge
remained unchanged, but the sharp edge between 10
and 50 has become blurred.

io led
The transition has reduced. Now from 10, the grey
level changes to 23.3 (23 after rounding off), from 23 it
changes to 36.6 (36 after rounding off) and finally from
ic ow

36 it changes to 50. Hence we can say that the sharp

edge has become blurred. Check for yourself, what
(a) Original image with (b) Low passed image
would happen if you took a bigger mask, say, 5  5.
n
Gaussian noise
bl kn

(You will have to take a bigger image to test your

results).
Fig. 7.4.13
at

 These kinds of averaging filters are excellent when the

Pu ch

Generalized MATLAB program for low pass averaging

image contains Gaussian noise. It achieves filtering by
using any mask size
blurring the noise. Some of the other low pass
Te

averaging masks are shown. %% Low pass filtering of an image using averaging
technique
0 1 0 1 1 1
clear all;
1 1
6 1 2 1 10  1 2 1 clc;
0 1 0 1 1 1 a=imread(‘blood.tif’);
a=double(a);
MATLAB program for low pass filtering along with the [row col]=size(a)
images with Gaussian noise
m=input(‘Enter the mask size:’);
%% Low pass filter used on an image with Gaussian noise n=m^2;
clear all for i=1:1:m
clc for j=1:1:m
aa=imread(‘xray.tif’); % size 600  800 w(i,j)=1/n;
f=double(aa); end
ab=imnoise(aa,’gaussian’); % adding noise end
a=double(ab); s=(m+1)/2;
w=[1 1 1; 1 1 1; 1 1 1]/9 for x=1:1:row
[row col]=size(a); for y=1:1:col
for x=2:1:row-1 b(x,y)=a(x,y);
for y=2:1:col-1 end
end
 DSIP (MU-Sem. 7-COMP) 7-14 Image Enhancement in Spatial Domain

for x=s:1:row-s  The steps to perform median filtering are as follows :

for y=s:1:col-s (1) Assume a 3  3 empty mask.
b(x,y)=0; (2) Place the empty mask at the left hand corner.
end
(3) Arrange the 9 pixels in ascending or descending
end
order.
for x=s:1:row-s
(4) Choose the median from these nine values.
for y=s:1:col-s
for i=1:1:m (5) Place this median at the centre.
for j=1:1:m (6) Move the mask in a similar fashion to the
b(x,y)=a(x-s+i,y-s+j)*w(i,j)+b(x,y); averaging filter.
end  In median filtering, the grey level of the centre pixel is
end replaced by the median value of the neighbourhood.
end Always write the resultant in a new matrix (image). We

ge
end shall explain median filtering with an example.
subplot (3,1,1)  In the image shown below let 250 be the salt and
imshow (uint8(a)) pepper noise. If we use an averaging filter, the noise
io led
title (‘Original Image’); would spread out and the edges would also end up
subplot (3,1,2) getting blurred. If we only want to remove the noise
without disturbing the edges, we use a median filter.
imshow (uint8(b))
ic ow

title (‘Low Pass Filtered Image’); 10 10 10 10 10 10 10 10

n
c=conv2(a,w);
bl kn

10 10 10 10 10 10 10 10
subplot(3,1,3)
imshow(uint8(c)) 10 250 10 10 10 10 10 10
at
Pu ch

title (‘Low Pass Filtered Image using MATLAB’);

10 10 10 10 10 10 10 10
7.4.4 Low Pass Median Filtering 50 50 50 50 50 250 50 50
Te

MU - Dec. 2017 50 50 50 50 50 50 50 50
Q. Write short notes on : Median Filter.
50 50 50 50 50 50 50 50
(Dec. 2017, 5 Marks)
50 50 50 50 50 50 50 50
 The averaging filter removes the noise by blurring it till
it is no longer seen. But in the process, it also blurs the  Just like the earlier case, we start by placing the mask
edges. Bigger the averaging mask more is the blurring. at the left hand corner. We again ignore the borders.
There are times when the image contains salt and
10 10 10 10 10 10 10 10
pepper noise.
 If we use an averaging filter to remove the same, it will 10 10 10 10 10 10 10 10
blur the noise but it would also ruin the edges. Hence 10 250 10 10 10 10 10 10
when we need to eliminate salt and pepper noise, we
work with a non-linear filter known as the Median 10 10 10 10 10 10 10 10
filter. 50 50 50 50 50 250 50 50
 They are also called order-statistic filters because their
response is based on the ordering or ranking of the 50 50 50 50 50 50 50 50
pixels contained within the mask. 50 50 50 50 50 50 50 50
 In this case, we use a mask similar to the averaging
50 50 50 50 50 50 50 50
filter except that the mask has no values. So it’s like
working directly with the 8 neighbours of the centre
pixel.
 DSIP (MU-Sem. 7-COMP) 7-15 Image Enhancement in Spatial Domain

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 250 10 10 10 10 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10

50 50 50 50 50 250 50 50 50 50 50 50 50 50 50 50

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

50 50 50 50 50 50 50 50 50 50 50 50 50 50 50 50

ge
 Arranging the 9 pixels in ascending or descending Solved Example
order we get,
Ex. 7.4.1 : If x = {2 3 4 3 4 5 6} and w = {–1 0 1}, perform
io led
(10, 10, 10, 10, 10, 10, 10, 10, 250)
th median filtering.
 The median is the 5 value since the mask is 3  3
mask. This value is placed at the center. We now move Soln. : w = {–1, 0, 1} simply means that the size of the mask
ic ow

the mask to the new location and continue the is 1  3 and the term 0 indicates the position from where
procedure. Performing this operation on the entire
n
the filtering starts.
image, we get the final image.
bl kn

Hence the final answer is y = {2 3 3 4 4 5 6}. The

 As can be seen, the salt and pepper noise gets
principal function of median filtering is to force points with
eliminated without distorting the edges. Median filters
at
Pu ch

distinct intensities to be more like their neighbours.

give astonishing results even when the amount of
noise is relatively large. MATLAB program for median filtering along with the
images with salt and pepper noise
Te

Fig. P. 7.4.1
 DSIP (MU-Sem. 7-COMP) 7-16 Image Enhancement in Spatial Domain

%% Median filter on image with salt and pepper noise %%

7.5 Highpass Filtering
clear all
clc  Highpass filtering eliminates the low frequency regions
I=imread (‘deepa.tif’); while retaining or enhancing the high frequency
J=imnoise (I, ‘salt & pepper’,0.02); % Adding noise components.
a=double(J);
 An image, which is high-passed, would have no
b=a;
background (as background are low frequency regions)
[row col]=size(a);
and would have enhanced edges. Hence high pass
for x=2:1:row-1;
filters are used to sharpen blurred images.
for y=2:1:col-1;
%% To make a 3  3 mask into a 1  9 mask  Once we have understood as to how a mask moves
a1=[a(x–1,y–1) a(x–1,y) a(x–1,y+1) a(x,y–1) over the entire image for the averaging filter, the same
a(x,y)… applies to the high pass filter as well. All that needs to
a(x,y+1) a(x+1,y-1) a(x+1,y) a(x+1,y+1)]; be changed are the mask coefficients.

ge
a2=sort(a1);  The frequency response and the spatial response of a
med=a2(5); % the fifth value is the median high pass filter are shown in Fig. 7.5.1.
b(x,y)=med
io led  As seen from the spatial response, it is clear that the
end mask coefficients have to be such that they have a
end positive value at the centre and negative values at the
figure(1)
ic ow

periphery.
Imshow(uint8(J))
 One of the high pass masks is shown in Fig. 7.5.1.
figure(2)
n
imshow(uint8(b))  The important thing to note is that the sum of the
bl kn

coefficients of the high pass mask has to be equal to

zero. This is because, when we place this mask over
at
Pu ch

the low frequency regions, the result should be zero.

Let us take an example of a pseudo-image.
Te

10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
10 10 10 10 10 10 10 10
100 100 100 100 100 100 100 100
(a) Original image with (b) Median filtered image
100 100 100 100 100 100 100 100
salt and pepper noise
100 100 100 100 100 100 100 100
Fig. P. 7.4.1 100 100 100 100 100 100 100 100

Fig. 7.5.1
 DSIP (MU-Sem. 7-COMP) 7-17 Image Enhancement in Spatial Domain

 By now we know what to expect when we do a high Note that 1/9 is simply a scaling function unlike the
pass operation. The background, which is low low pass mask in which the 1/9 term is a part of the
frequency, gets eliminated while the edges get averaging operation. In this case we could also work with
enhanced. 1/8 or 1/7 depending on the image. Some of the other high
 We move the mask in a similar fashion as in the low pass masks are as follows
pass filter example. The output that we get is shown
below. Note that the background has been eliminated 0 –1 0 –1 –2 –1
and we get all zero values. This is because the sum of
the mask coefficients is zero. –1 4 –1 –2 12 –2
0 0 0 0 0 0 0 0
0 –1 0 –1 –2 –1
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 The result using the first mask and putting negative
values as zeros is.
–270 –270 –270 –270 –270 –270 –270 –270
270 270 270 270 270 270 270 270 0 0 0 0 0 0 0 0

ge
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
io led
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
There are two issues to be dealt with, when working 30 30 30 30 30 30 30 30
ic ow

with high pass masks.

0 0 0 0 0 0 0 0
(1) As we can see, there are negative values in the output
0 0 0 0 0 0 0 0
n
image. Pixel values cannot be negative. They always
bl kn

start with zero. Hence we need to get rid of the 0 0 0 0 0 0 0 0

negative values. If we consider the mod operation, do
MATLAB program for high pass filtering
at

you think it would solve the problem ? NO. The reason

Pu ch

for that is –270 is a value that is lower than zero, that %% High pass filtering %%
is, it is supposed to be darker than the darkest value clear all
Te

i.e. zero. If we now take the mod value i.e. clc

MOD [–270], we get +270, which is definitely not aa=imread (‘xray.tif’);
darker than the zero value. What would happen is that a=double (aa);
all large negative values would be shown as bright
[row col]=size(a);
spots. Hence taking the mod of negative values will
distort the image grey levels. A simple way to get rid of w=[-1 -1 -1; -1 8 -1, -1 -1 -1]
this problem is to let all negative values be zero. Hence for x=2:1:row-1
–270 would be written as 0. for y=2:1:col-1
(2) The values of the original image at the edge are very a1(x,y)= w(1)*a(x-1,y-1)+w(2)*
large and there is a tendency of them going out of a(x-1,y)+w(3)* …
range due to the centre weight of +8. We always a(x-1,y+1)+w(4)*a(x,y-
encounter this problem in practice. To eliminate this 1)+w(5)*a(x,y)+w(6)* …
problem, we use a mask with a scaling function. a(x,y+1)+w(7)*a(x+1,y-
Hence for practical purposes we use a mask are as 1)+w(8)*a(x+1,y)+w(9)* …
follows, a(x+1,y+1);
end
–1 –1 –1
end
1
9 –1 8 –1 figure(1)
imshow(uint8(a))
–1 –1 –1
figure(2)
33 imshow(uint8(al))
High pass mask
 DSIP (MU-Sem. 7-COMP) 7-18 Image Enhancement in Spatial Domain

We have a mask which is as shown

–1 –1 –1
1
9 –1 8.9 –1

–1 –1 –1

3  3 High boost mask

(a) Original image (b) High pass filtered image
 We can select different values of A and see the
Fig. 7.5.2
difference.
7.6 High-Boost Filtering  Use the same pseudo image as the one used in the
high pass example to see the results. What you will
 As can be seen, the high pass filter gives great results. notice is that places which had a zero in the high pass
But there is one problem. It gets rid of the complete example will now have some positive values. Hence it
background.

ge
does not eliminate the background completely. This
 There are times, when we need to enhance the edges technique is one of the basic tools that is used in the
but also retain some of the background. To do that, we printing industry.
io led
use a modified version of the high pass filter known as  There are other neighbourhood techniques like
High-Boost filtering. Robert’s filtering, Sobel’s filtering and Prewitt’s
 In High-Boost filtering, we pass some of the filtering, which are similar to the methods discussed
ic ow

background along with the high frequency content. above. They form a separate chapter called
We know that, High pass = Original – Low pass Segmentation and hence would be discussed later.
n
bl kn

 To pass some of the background, we multiply the MATLAB program for high-boost filtering
original image with a multiplicative factor A. This gives %% High boost filtering
at

us high boost filtering. clear all

Pu ch

Hence, clc
High Boost = (A) original – Low pass aa=imread (‘xray.tif’);
Te

= (A –1) Original + Original – Low pass a=double(aa);

High Boost = (A –1) Original + High pass [row col]=size (a);
If A = 1, then w=[-1 -1 -1; -1 8.9 -1; -1 -1 -1]; %% High boost
High Boost = High pass mask

 If A > 1, then some of the original signal is added back for x=2:1:row-1
to the high pass result. This process restores some of for y=:1:col-1
the background into the high passed image. This a1(x,y)= w(1)*a(x-1,y-1)+w(2)*
technique is also known as unsharp masking. The a(x-1,y)+w(3)* …
mask coefficients for high boost filtering are as a(x-1,y+1)+w(4)*a(x,y-
follows, 1)+w(5)*a(x,y)+w(6)* …
Here X= 9A – 1 a(x,y+1)+w(7)*a(x+1,y-
1)+w(8)*a(x+1,y)+w(9)* …
Hence if A= 1, X = 8 which is a high pass mask.
a(x+1,y+1);
–1 –1 –1
1 end
9 –1 X –1 end
–1 –1 –1 figure(1)
imshow(uint8(a))
3  3 High boost mask
figure(2)
If A = 1.1, X = 8.9.
imshow(uint8(a1))
 DSIP (MU-Sem. 7-COMP) 7-19 Image Enhancement in Spatial Domain

end
end
n=0
for x=s:1:row-s
for y=s:1:col-s
if b(x,y) > n
n=b(x,y);
(a) Original (b) High boost end
Fig. 7.6.1 end
The generalized MATLAB program for high pass end
filtering using any mask size is given next. c=255/n;
%% High pass filtering of an image %% for x=s:1:row-s
clear all; for y=s:1:col-s

ge
clc; b(x,y)=c*b(x,y);
a=imread(‘mri2.tif’); end
a=double(a); end
io led
[row col]=size(a); subplot(2,1,1)
m = input(‘Enter the mask size:’); imshow(uint8(a))
title(‘Original Image’);
ic ow

for i=1:1:m
for j=1:1:m subplot(2,1,2)
n
w(i,j)=-1 imshow(uint8(b))
bl kn

end title(‘High Pass Filtered Image’)

end
at

7.7 Zooming
Pu ch

s=(m+1)/2;
w(s,s)=m^2-1;
 Another important application in image enhancement
for x=1:1:row
Te

where the spatial domain neighbourhood operation is

for y=s:1:col used image zooming.
b(x,y)=a(x,y);
 Zooming of images is not something that is new to
end you. If you have used Microsoft paint or Photoshop
end editor, you would be aware that there is an option
for x=s:1:row-s which allows us to zoom an image by 25%, 50%, and
for y=s:1:col-s 100% ……Have you ever imagined as to how an image
b(x,y)=0; that looked small can suddenly look so big ? The
end technique is quite simple and after reading the next
end couple of pages, you would be able to do it yourself.
for x=s:1:row-s  Zooming can be carried out using two different
for y=s:1:col-s methods,
for i=1:1:m (1) Replication (2) Interpolation
for j=1:1:m
 We will discuss both these methods in detail. We shall
b(x,y)=a(x-s+i,y-s+j)*w(i,j)+b(x,y); first start with zooming using replication.
end
end
if b(x,y) < 0
b(x,y)=0;
end Fig. 7.7.1
 DSIP (MU-Sem. 7-COMP) 7-20 Image Enhancement in Spatial Domain

7.7.1 Replication 1 0 2 0 3 0 4 0
 This is also known as zero interlacing the columns, as
 In replication, we simply replicate each pixel and then
we add zeros at every other column. Now inserting a
replicate each row. Consider the pseudo-image shown
row full of zeros gives us. This is known as zero
below. interlacing the rows as we add zeros at every other
1 2 3 4 row.

5 6 7 8 1 0 2 0 3 0 4 0
0 0 0 0 0 0 0 0
9 8 6 7
1 0 2 0 3 0 4 0
0 1 2 3
0 0 0 0 0 0 0 0
 As stated earlier, we start from the first row. We
replicate each pixel and then replicate each row. The 5 0 6 0 7 0 8 0
first row now looks like 0 0 0 0 0 0 0 0

ge
1 1 2 2 3 3 4 4
9 0 8 0 6 0 7 0
We now replicate this row to get
io led 0 0 0 0 0 0 0 0
1 1 2 2 3 3 4 4
1 1 2 2 3 3 4 4 0 0 1 0 2 0 3 0
ic ow

Performing this operation on the entire image we get 0 0 0 0 0 0 0 0

1 1 2 2 3 3 4 4  This image is known as the zero interlaced image. On

n
this image, we run a replication mask given below to
bl kn

1 1 2 2 3 3 4 4
get the zoomed image.
5 5 6 6 7 7 8 8
at

1 1
Pu ch

5 5 6 6 7 7 8 8
1 1
9 9 8 8 6 6 7 7
Te

 The final zoomed image is as follows,

9 9 8 8 6 6 7 7
1 1 2 2 3 3 4 4
0 0 1 1 2 2 3 3

0 0 1 1 2 2 3 3 1 1 2 2 3 3 4 4

 Hence a 4  4 image is zoomed to a 8  8 image. This

5 5 6 6 7 7 8 8
method can be repeated to get bigger images.
Remember, this is an image enhancement technique
5 5 6 6 7 7 8 8
and hence no new data is added.
 In the zoomed pseudo-image, we observe that as we 9 9 8 8 6 6 7 7
increase the size of the image, clusters of grey levels
are formed which are disconcerting to the observer. 9 9 8 8 6 6 7 7
Hence zooming increases the size of the image no
doubt, but it also gives the image a patchy look. 0 0 1 1 2 2 3 3
 Zooming by replication can be implemented on the
computer by using a replication mask. The first step is 0 0 1 1 2 2 3 3
to interlace the original image with zeros. This is
known as zero interlacing. Note : To get the above result, we need to add a row of
 In this we add zeros after every pixel and then add a zeroes at the top and a column of zeroes at the beginning
complete row of zeros. Adding zeros to every other of the zero interlaced image making it 9  9.
pixel of the first row we get,
 DSIP (MU-Sem. 7-COMP) 7-21 Image Enhancement in Spatial Domain

 As mentioned earlier, zooming by replication gives the  This is the first step. We now proceed to find the
final image a patchy look since clusters of grey levels average values along the columns. Hence the final
are formed. This can be substantially reduced by using image is as shown. In practice, we need to round of
a better method of zooming known as interpolation. the pixel values.

7.7.2 Linear Interpolation 1 1.5 2 2.5 3 3.5 4 2

 In this method, instead of replicating each pixel, 3 3.5 4 4.5 5 5.5 6 3

average of the two adjacent pixels along the rows is 5 5.5 6 6.5 7 7.5 8 4
taken and placed between the two pixels. The same
7 7 7 6.25 6.5 7 7.5 3.25
operation is then performed along the columns. This

ge
operation is done on the interlaced image. 9 8.5 8 7 6 6.5 7 3.5
1 0 2 0 3
io led 0 4 0 4.5 4.5 4.5 4.25 4 4.5 5 2.5
0 0 0 0 0 0 0 0
0 0.5 1 1.5 2 2.5 3 1.5
5 0 6 0 7 0 8 0
0 0 0 0 0 0 0 0 0 0.25 0.5 0.75 1 1.25 1.5 0.75
ic ow

9 0 8 0 6 0 7 0  If we compare the results of replication and

0 0 0 0 0 0 0 0 interpolation, it is clear that the patchiness that was
n
present in the replicated image is much less in the
bl kn

0 0 1 0 2 0 3 0
interpolated image. Hence we can conclude that
0 0 0 0 0 0 0
zooming by interpolation is more effective than
at
Pu ch

Interpolation along rows is given by, zooming by replication.

v1 (m, 2n) = u (m, n); 0  m  M –1, 0  n  N –1  To implement zooming by interpolation, we simply run
Te

1 an interpolation mask on the zero interlaced image.

v1 (m, 2n + 1) = [u(m, n) + u (m, n + 1)];
2
1/4 1/2 1/4
0mM –1; 0  n  N –1
1/2 1 1/2
Interpolation of the column is given by,
v(2m, n)= v1 (m, n); 1/4 1/2 1/4
1 MATLAB program for zooming by replication as well as
v(2m + 1, n) = [v (m, n) + v1 (m + 1, n)];
2 1 interpolation
0  m  M –1; 0  n  N –1 %% Zooming using replication as well as interpolation %%
Hence the first row now becomes the modified image clc
as follows : clear all
x=imread (‘warne.tif’);
1 1.5 2 2.5 3 3.5 4 2
s=size(x);
0 0 0 0 0 0 0 0
ss=s(1)+s(2); %% To set the size of the output image
5 5.5 6 6.5 7 7.5 8 4 %%
0 0 0 0 0 0 0 0 y=zeros(ss);
9 8.5 8 7 6 6.5 7 3.5 i=l;
0 0 0 0 0 0 0 0 j=1;
for n=1:2:ss
0 0.5 1 1.5 2 2.5 3 1.5
for m=1:2:ss
0 0 0 0 0 0 0 0
y (m, n)= x (i, j);
 DSIP (MU-Sem. 7-COMP) 7-22 Image Enhancement in Spatial Domain

i=i+1; We get the digital negative using the above equation.

end
134 50 38 99 104
i=1; %% Else the value of i goes on
increasing%% 116 128 98 138 130
j=j+1;
3 138 19 117 113
end
h1=[1 1;1 1]; % Mask for replication 28 73 77 58 13
h2=[1/4 1/2 1/4; 1/2 1 1/2; 1/4 1/2 1/4]; 54 149 136 4 15
%% Mask for interpolation %%
A1=conv2 (y,h1); %% We can use the formula that is used Ex. 7.8.2 :For a given image find -
in Low pass filtering %%
(i) Digital negative of an image.
A2=conv2 (y,h2);
(ii) Bit plane slicing.
figure(1),imshow(x)
figure(2),imshow(uint8 (A1)) 4 3 2 1

ge
figure(3),imshow(uint8 (A2))
3 1 2 4

5 1 6 2
io led
2 3 5 6

Soln. : We assume the image to be 3-bit

ic ow

(a) Original (b) Zooming by (c) Zooming by (i) Digital negative

replication interpolation
n
bl kn

Fig. 7.7.2 s = (L – 1) – r

or g (x, y) = (L – 1) – f (x, y)
7.8 Solved Examples
at
Pu ch

3
Here L = 2 = 8
Ex. 7.8.1 : Obtain the digital negative of the following 8 Bits
 s = 7–r
Te

Per Pixel - BPP image.

i.e., g (x, y) = 7 – f (x, y)
121 205 217 156 151
 The digital negative of the image is
139 127 157 117 125
3 4 5 6
252 117 236 138 142
4 6 5 3
227 182 178 197 242
2 6 1 5
201 106 119 251 240
5 4 2 1
MU - May 2014, 10 Marks.
Soln. : (ii) Bit plane slicing
It is known that it is a 8-bit image. Hence the number In this, we convert the given image into binary and
8
of grey levels that this image can hold is 2 = 256 then separate the planes.
 L = 256. We assume the image to be 3 bit.
Hence the minimum grey level is 0
100 011 010 001
while the maximum grey level is 255
011 001 010 100
s(x,y) = (L – 1) – r(x,y) = (256 – 1) – r(x,y)
101 001 110 010
s(x,y) = 255 – r(x,y)
010 011 101 110
 DSIP (MU-Sem. 7-COMP) 7-23 Image Enhancement in Spatial Domain

Separating the bit planes we get,

1 0 0 0 0 1 1 0 0 1 0 1

0 0 0 1 1 0 1 0 1 1 0 0

1 0 1 0 0 0 1 1 1 1 0 0

0 0 1 1 1 1 0 1 0 1 1 0

MSB plane Center bit plane LSB plane

Ex. 7.8.3 : For following image find :

Contrast stretching r2 = 5, r1 = 3, s2 = 6, s1 = 2. Fig. P. 7.8.3

ge
4 3 2 1 Hence we have the required slopes to compute
contrast stretching. The contrast stretching formula is given
io led
3 1 2 4 below.
f (x, y) =
5 1 6 2
  r 0r<3
s =    (r – r1) + s1 3  r < 5
ic ow

2 3 5 6   (r – r ) + s 5  r < 7
2 2
n
MU - May 2016, 10 Marks. Here r1 = 3, r2 = 5, s1 = 2, s2 = 6
bl kn

Soln. :  = 0.66,  = 2,  = 0.5

at

We make a table of values for r and s using the

Pu ch

Contrast stretching r2 = 5, r1 = 3, s2 = 6, s1 = 2 :
contrast stretching formula.
4 3 2 1
r s
Te

f (x, y) = 3 1 2 4
0 s = r = 0.66  0 = 0
5 1 6 2
1 s = r = 0.66  1 = 0.66
2 3 5 6
2 s = r = 0.66  2 = 1.32
The contrast stretching transformation is shown in 3 s =  (r – r1) + s1 = 2 (3 – 3) + 2 = 2
Fig. P. 7.8.3.
4 s =  (r – r1) + s1 = 2 (4 – 3) + 2 = 4
From the values of r1 and s1 we can find the slope of .
In a similar manner, we can find the value of  using the 5 s =  (r – r2) + s2 = 0.5 (5 – 5) + 6 = 6
values of r1, r2 and s1, s2. 6 s =  (r – r2) + s2 = 0.5 (6 – 5) + 6 = 6.5
From the values of r2 and s2, we can find the value of .
7 s =  (r – r2) + s2 = 0.5 (7 – 5) + 6 = 7
s1 2
 = = = 0.66 Hence the contrast stretched image is
r1 3
y2 – y1 6 – 2 4 2 1.32 0.66
 = = =2
x2 – x1 5 – 3
2 0.66 1.32 4
y2 – y1 7 – 6
Finally,  = = = 0.5 6 0.66 6.5 1.32
x2 – x1 7 – 5
1.32 2 6 6.5
 DSIP (MU-Sem. 7-COMP) 7-24 Image Enhancement in Spatial Domain

4 2 1 1
0 if f(x, y) 4
If we round off, the final  2 1 1 4
image would be 6 1 7 1 g(x, y) =
1 2 6 7 L–1 if f(x, y) 4

Ex. 7.8.4 : For the 3-bit 4  4 size image, perform the Here
following operations : L–1 = 7
(i) Negation  The thresholded image is as follows,
(ii) Thresholding with T = 4
7 7 7 7
(iii) Intensity level slicing with background r1 = 2 and r2 = 5
(iv) Bit plane slicing for MSB and LSB planes
7 0 0 0
(v) Clipping with r1 = 2 and r2 = 5 0 7 0 7

ge
1 2 3 0 7 7 0 7
2 4 6 7
(iii) Bit plane slicing
5 2
io led 4 3
We convert the given image into binary and then
3 2 6 1 separate the plane. Since the given image is 3-bit, the
Soln. : Let the given image be f(x, y) binary image would be as follows,
ic ow

(i) Negation 001 010 011 000

n
s = (L – 1) – r 010 100 110 111
bl kn

OR 101 010 100 011

at

g (x, y) = (L – 1) – f(x, y)
Pu ch

3
011 010 110 001
Since the image is 3-bit, L = 2 = 8
Separating the bit planes we obtain.
L–1 = 7
Te

 g(x, y) = 7 – f(x, y) 0 0 0 0 0 1 1 0 1 0 1 0

Hence the digital negative of the image is 0 1 1 1 1 0 1 1 0 0 0 1

6 5 4 7 1 0 1 0 0 1 0 1 1 0 0 1

g(x, y) = 5 3 1 0 0 0 1 0 1 1 1 1 1 0 0 1

2 5 3 4 MSB plane Center bit plane LSB plane

4 5 1 6 (iv) Clipping with r1 = 2, r2 = 5

(ii) Thresholding with T = 4 The clipping transformation is shown in Fig. P. 7.8.4(a).

The thresholding function is shown in Fig. P. 7.8.4

along with thresholding formula.

Fig. P. 7.8.4 Fig. P. 7.8.4(a)

 DSIP (MU-Sem. 7-COMP) 7-25 Image Enhancement in Spatial Domain

r 2  r 5 s=L–1 r1  r  r2 s=L–1 r1  r  r2
s = 
0 otherwise s=0 otherwise s= r otherwise
Hence the final clipped image is as follows,
0 0 0 0 0 2 1 2 2 1
0 2 3 0
2 4 0 0 0 7 7 7 0 2 7 7 7 2

5 2 4 3 0 0 0 0 0 6 2 7 6 0
3 2 0 0
0 0 0 7 0 2 6 6 7 1
Ex. 7.8.5 : Perform intensity level (grey level) slicing on the
3 BPP image. Let r1 = 3 and r2 = 5. Draw the modified image 0 7 0 0 0 0 7 2 2 1

using with background and without background Without background With background
transformations.
Ex. 7.8.6 : The grey levels in an image range from 10 to 50.

ge
Soln. : Let us draw the grey level transformation.
We need to display the image on a device that has a grey
2 1 2 2 1 level range of 0 to 255. Give a transformation which would
io led
2 3 4 5 2 accomplish this.

6 2 7 6 0 Soln. : The transformation that we use is known as the

ic ow

2 6 6 5 1 Range normalization.

0 3 2 2 1 It is given by,
n
d–c
bl kn

s (x, y) = [r (x, y) – a] + c
b–a
at

This transformation converts the original range [a, b]

Pu ch

to the new range [c, d].

Te

In this sum a = 10, b = 50, c = 0, d = 255.

255
s(x, y) = [r(x, y) – 10] + 0
40
51 255
s(x, y) = r(x, y) –
8 4

When r = 10, s = 0
(a)
r = 50, s = 255

Ex. 7.8.7 : The image shown below has 8 different grey

levels. Plot this image using only 4 grey levels.

0 1 1 1 1 4

1 1 2 3 2 2

1 1 2 2 3 3

1 2 4 6 2 3

1 2 4 2 4 4
(b)
1 2 3 7 2 5
Fig. P. 7.8.5
 DSIP (MU-Sem. 7-COMP) 7-26 Image Enhancement in Spatial Domain

Soln. : When we apply the LPF on the image, the center pixel

This can be achieved using the Grey level reduction z5 changes to

1
transformation which is given by,
9 [z1 + z2 + z3 + z4 + z5 + z6 + z7 + z8 + z9 ]
s(x, y) = Gr(x , y) K 1
 K G Original – Low pass = z5 – [z + z2 + z3 + z4 + z5 + z6 + z7
9 1
Here K Original number of grey levels + z8 + z 9 ]
G  Modified number of grey levels z1 z2 z3 z4 z5 z6 z7 z8 z9
= z5 – – – – – – – – –
9 9 9 9 9 9 9 9 9
[.]  Choosing the floor value. 8z5 1
= – [z1 + z2 + z3 + z4 + z5 + z6 + z7 + z8 + z9]
s(x, y) = 4r(x , y) 8 9 9
 8 4
This is nothing but a high pass mask
 r(x, y) s(x, y)

ge
0 0 1 –1 –1 –1
1 0 1
2 2 2 9  –1 8 –1
io led
3 2
–1 –1 –1
4 4 3
5 4 Note : We can also prove this by taking a pseudo-image.
ic ow

6 6 4 We move a low pass mask on the original image and

n
7 6 subtract it from the original. We now move a high pass
bl kn

mask on the original image. The results of both these

 The modified image is operations will be the same.
at
Pu ch

0 0 0 0 0 4 Ex. 7.8.9 : Filter the following image using 3  3

neighbourhood averaging by assuming (i) Zero padding
0 0 2 2 2 2
Te

(ii) Pixel replication

0 0 2 2 2 2 1 2 3 2

0 2 4 6 2 2 f (x, y) = 4 2 5 1

0 2 4 2 4 4 1 2 6 3

0 2 2 6 2 4 2 4 6 7

Soln. :
Ex. 7.8.8 : Show that a high pass filtered image can be
A 3  3 averaging mask is shown below :
obtained in the spatial domain as High pass = Original – Low
pass. 1 1 1
Soln. : 1
w (x, y) =9 1 1 1
Original image Low pass filter High pass filter 1 1 1

z1 z2 z3 1/9 1/9 –1/9 –1/9

Filtering is performed using the convolution operation,
1/9 –1/9
i.e., g (x, y) = f (x, y)  w (x, y).
z4 z5 z6 1/9 1/9 1/9 –1/9 8/9 –1/9
(i) Zero padding

z7 z8 z9 1/9 1/9 1/9 –1/9 –1/9 –1/9 In this, we zero pad the image before performing the
filtering operation. This gives us a 6  6 image.
 DSIP (MU-Sem. 7-COMP) 7-27 Image Enhancement in Spatial Domain

1 1 2 3 2 2
0 0 0 0 0 0
1 2 2.55 2.44 2.33 2
0 1 2 3 2 0 4 2 2.88 2.88 2.88 1
g (x, y) =
1 2.44 3.55 4 4.33 3
0 4 2 5 1 0
f (x, y) = 2 2.22 3.66 5 5.77 7
0 1 2 6 3 0 2 2 4 6 7 7

0 2 4 6 7 0 Ex. 7.8.10 : Perform discrete convolution on the following

image arrays. Assume that the left bottom corner elements
0 0 0 0 0 0 are at the origin in both cases. f1 and f2 are two images.

We move the averaging mask over this image starting

from the top left corner as follows,

ge
 g (x, y) = f (x, y)  w (x, y)

0 0 0 0 0 0
io led
0 1 1.88 1.66 1.22 0
Fig. P. 7.8.10
ic ow

0 1.33 2.88 2.88 2.22 0

Soln. :
n
0 1.66 3.55 4 3.11 0
f1 (x, y) * f2 (x, y) =  f1 (m, n) f2 (x –m, y – n)
bl kn

0 1 2.33 3.11 2.44 0 m n

Let f1 (x, y) * f2 (x, y) = g(x, y)
at
Pu ch

0 0 0 0 0 0
g(x, y) =  f1 (m, n) f2 (x –m, y – n)
(ii) Pixel replication
Te

m n

In this we replicate the border pixels to generate a

6  6 image. g(0, 0) =  f1 (m, n) f2 (–m, – n)
m n

1 1 2 3 2 2

1 1 2 3 2 2

4 4 2 5 1 1
f (x, y) =

1 1 2 6 3 3
Fig. P. 7.8.10(a)
2 2 4 6 7 7 The basic shift operations are

2 2 4 6 7 7

We move the averaging mask over this image starting

from the top left corner as shown.

 g (x, y) = f (x, y)  w (x, y)

Fig. P. 7.8.10(b)
DSIP (MU-Sem. 7-COMP) 7-28 Image Enhancement in Spatial Domain

f2(m, n)  fliping about m-axis f2(m, n)  fliping about n-axis

ge
io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 7.8.10(c)cont…
 DSIP (MU-Sem. 7-COMP) 7-29 Image Enhancement in Spatial Domain

corel 2

ge
io led
ic ow
n
bl kn
at
Pu ch
Te

Fig. P. 7.8.10(c)

Another method : It is convenient to represent digital Soln. :

images as matrices and exploit the benefits of linear Zero padding F1 and F2 so that both are of size
algebra. We modify the matrix so as to suit convolution.
(m1 + m2 – 1)  (n1 + n2 – 1) = 3  4
Let F1 and F2 be the two images that need to be
convolved :
Let F1 be a m1  n1 image and F2 be a m2  n2 image.
We zero pad F1 and F2 so that both of them are of size
(m1 + m2  1)  (n1 + n2  1).

Ex. 7.8.11 : Perform discrete convolution on the given image

arrays

Fig. P. 7.8.11(a)

Now, of these two, arrange one as a circular matrix

and the other one as a row stacking matrix. Let the first
element of these be their respective origins, then
G (x, y) = F1 (x, y) F2 (x, y)
Fig. P. 7.8.11
 DSIP (MU-Sem. 7-COMP) 7-30 Image Enhancement in Spatial Domain

ge
We see that the results are the same as obtained from the previous
method.
io led
ic ow

Fig. P. 7.8.11(b)
n
Ex. 7.8.12 : Apply Low and High Pass Spatial masks on the following Fig. P. 7.8.12 matrix.
bl kn

Prove that High pass = Original – Low pass. Assume virtual Rows and Columns.

30 31 32
at
Pu ch

f(x,y) = 33 120 30

32 32 31
Te

Fig. P. 7.8.12

Soln. : We zero padd the image to make it 5  5

0 0 0 0 0

0 30 31 32 0

f(x,y) = 0 33 120 30 0

0 32 32 31 0

0 0 0 0 0

We convolve this image with a standard 3  3 Low pass and a High pass mask.

(i) Low Pass

1 1 1
Low pass mask h1 (x,y) = 1 1 1

1 1 1
 DSIP (MU-Sem. 7-COMP) 7-31 Image Enhancement in Spatial Domain

23.77 30.66 23.66

g1 (x,y) = f(x,y) * h1(x,y) = 30.88 41.22 30.66

24.11 30.88 23.667

(ii) High Pass

–1 –1 –1
High pass mask h2 (x,y) = 1
–1 8 –1
9
–1 –1 –1

6.22 0.33 8.33

ge
g2 (x,y) = f (x,y) * h2 (x,y) =
io led 2.11 78.77 – 0.667

7.889 1.11 7.33

We now show that High pass = Original – Low pass

ic ow
n
30 31 32 23.77 30.66 23.66
bl kn

Original – Lowpass = 33 120 30 – 30.88 41.22 30.66

at
Pu ch

32 32 31 24.11 30.88 23.667

Te

6.22 0.33 8.33

= 2.11 78.77 – 0.667

7.889 1.11 7.33

This is the same as the High pass result. Therefore, High pass = Original – Low pass
Here m1  n1 is the size of f(x, y) [ 2  and m2  n2 is
Ex. 7.8.13 : Given f (x, y) = 
5 6 7
, 1 2

 h(x, y) =   the size of h(x, y) [2  2], Therefore m1 = 2, m2 = 2, n1 = 3,
8 9 10  3 4
n2 = 2. Hence zero padding should be done in such a
Find linear convolution of input image f(x, y) with filter of
manner so that f1 (x, y) and h(x, y) are of size
h(x, y).
(2 + 2 – 1)  (3 + 2 – 1).
Soln. : Let the output image be g (x, y)
= 34
Here, g (x, y) = f(x, y)  h(x, y)
 
0 0 0 0
We shall solve this using the circular matrix method.
(The other method is the graphical technique).  f (x, y) =  5 6 7 0  ; h(x, y)
We zero pad f(x, y) and h(x, y) in such a way that both   9 10 0 
 
of them are of size. 0 0 0 0
(m1 + m2 – 1)  (n1 + n2 – 1)
=  1 2 0 0

  4 0 0
 DSIP (MU-Sem. 7-COMP) 7-32 Image Enhancement in Spatial Domain

We now arrange one of these as a circular matrix and the other one as a row slacking matrix. Let us convert f(x, y) to a
circular matrix. Let the encircled terms be the starting points

8 0 0 0 0 0 7 6 5 0 10 9 3
9 8 0 0 0 0 0 7 6 5 0 10 4

10 9 8 0 0 0 0 0 7 6 5 0 0

0 10 9 8 0 0 0 0 0 7 6 5 0

5 0 10 9 8 0 0 0 0 0 7 6 × 1

6 5 0 10 9 8 0 0 0 0 0 7 2

7 6 5 0 10 9 8 0 0 0 0 0 0

0 7 6 5 0 10 9 8 0 0 0 0 0

ge
0 0 7 6 5 0 10 9 8 0 0 0 0

0 0 0 7 6 5 0 10 9 8 0 0 0
io led
0 0 0 0 7 6 5 0 10 9 8 0 0

0 0 0 0 0 7 6 5 0 10 9 8 0
ic ow

Multiplying the two matrices gives us Ex. 7.8.14 : For the given 3 bit, 4  4 size image perform
n
bl kn

24 the following operations :

(i) Thresholding (T = 4)
59
at

(ii) Intensity level slicing with background for r1 = 2,

Pu ch

66 r2 = 5
40 (iii) Bit plane slicing for LSB and MSB planes
Te

23 (iv) Negation.

4 2 3 0
63
1 3 5 7
73
5 3 2 1
48
2 4 6 7
5
MU - Dec. 2011, 10 Marks)
16
Soln. :
19
(i) Thresholding (T = 4)
14 3
Since the given image is 3-bit, L = 2 = 8
We re-arrange this matrix to generate a filtered image. The transformation for thresholding is shown in

 
5 16 19 14 Fig. P. 7.8.14.

 g(x, y) =  23 63 73 48  Here,

 24 59 66 40  s = L–1=7 r4
We would get the same answer if we make h(x, y) into s = 0 r<4
a circular matrix instead of f(x, y). Go ahead and try it out
yourself.
 DSIP (MU-Sem. 7-COMP) 7-33 Image Enhancement in Spatial Domain

 The final image would be, (iii) Bit plane slicing

We convert the image to binary. Since there are 8 grey
levels (0 to 7), we require 3 bits to represent each
pixel.

4 2 3 0 100 010 011 000

1 3 5 7 Binary 001 011 101 111

Fig. P. 7.8.14
5 3 2 1 ––– 101 011 010 001
7 0 0 0
2 4 6 7 010 100 110 111
0 0 7 7
The LSB and MSB planes are as follows,
7 0 0 0

ge
0 0 1 0 1 0 0 0
0 7 7 7
1 1 1 1 0 0 1 1
(ii)
io led
Intensity level slicing with background for
r1 = 2, r2 = 5 1 1 0 1 1 0 0 0

The transformation for intensity level slicing is shown 0 0 0 1 0 1 1 1

ic ow

in Fig. P. 7.8.14(a).
LSB plane MSB plane
Here,
n
bl kn

s = L–1=7 2r5 (iv) Negation

s = r otherwise The transformation for negation is shown in
at
Pu ch

Fig. P. 7.8.14(b).
Here,
Te

s = (L – 1) – r
s = 7–r
Here the grey levels get inverted.

 The final image is as shown in Fig, 7.8.14(b)

Fig. P. 7.8.14(a)

Hence all values between 2 and 5 are made equal to 7

and the remaining values remain unchanged. Hence
the final result is,

Fig. P. 7.8.14(b)
7 7 7 0
3 5 4 7
1 7 7 7
6 4 2 0
7 7 7 1
2 4 5 6
7 7 6 7
5 3 1 0
 DSIP (MU-Sem. 7-COMP) 7-34 Image Enhancement in Spatial Domain

7.9 Comparison of Contrast Stretching and Thresholding

Sr. Contrast Stretching Thresholding
No.

1. Contrast stretching increases the dynamic range by Thresholding also increase that dynamic range by
making the dark pixels darker and bright pixels setting a threshold. All values less than the threshold
brighter. are made black while values greater than the threshold
are made equal to white.

2. The formula of the contrast-stretching algorithm is The formula for achieving thresholding is as follows :
given below : s=0; if ra
lr 0r<a s=L–1; if r>a
s =  m  (r – a) + v a  r < b
 n  (r – b) + w b  r < L – 1

ge
3. The transformation for contrast stretching is shown The transformation for thresholding is shown below :
below :
io led
ic ow
n
bl kn
at
Pu ch

4. We can adjust the contrast in an image by changing the Thresholding is extreme contrast stretching. In other
Te

values of the slope. words, a thresholded image has maximum contrast.

5. The final result of contrast stretching gives us a grey The final result of thresholding gives us a binary image.
scale image. The output image has different shades of Which means that the thresholded image has only
grey. black and white values.

Summary Review Questions

In this chapter, we introduce the term spatial domain Q. 1 Explain spatial domain processing.
and image enhancement in the spatial domain. Different
image enhancement techniques that improve the Q. 2 Compare and contrast average filtering and median
subjective quality of the image are explained. The methods filtering.
described here can be classified into two broad groups :
Q. 3 Explain the difference between operations involving
point processing and neighborhood processing. The
3  3 mask for median filtering and average filtering.
concept of frequencies in an image is explained using
simple illustrations. Basic filtering operations such low pass Q. 4 Develop a procedure for computing the median in
filtering and high pass filtering are presented. Advantages n  n neighbourhood.
of each of the operations are stated using MATLAB codes.
Any particular technique that may be applied depends on Q. 5 What do we mean by Gaussian noise and why is an
the subjective quality of the image as well as the purpose averaging filter used to eliminate it ?
behind the processing. It must be noted that more than one
technique can be used to get the desired result.
 DSIP (MU-Sem. 7-COMP) 7-35 Image Enhancement in Spatial Domain

Q. 6 What is salt and pepper noise (binary noise) and how Q. 14 Given a image, what is the output using a 3 H 3
does a median filter remove it ? averaging filter and a medain filter ? Explain the
result.
Q. 7 Explain the output and application of the following
3 3 3 33 3
zero memory enhancement techniques.
3 3 3 33 3
(1) Contrast stretching
3 3 10 10 3 3
(2) Thresholding
3 10 3 33 3
(3) Range compression
3 3 3 38 3
(4) Bit extraction.
3 3 3 33 3
Q.8 Compare and contrast the low pass and high pass 3 3 3 33 3
spatial filters. 3 3 3 33 3

Q. 9 Explain why a median filter is called a salt and pepper Q. 15 How is image subtraction carried out ? Give an

ge
filter. example of image subtraction in the field of medicine.

Q. 10 Compare and contrast the smoothing and sharpening Q. 16 Explain the procedure of zooming an image using
io led
filters. replication and interpolation.

Q. 11 Low pass filter is a smoothing filter. Explain. Q. 17 Why should the sum of any high pass mask be zero ?
ic ow

Q. 12 What do we mean by frequencies in an Q. 18 Perform discrete convolution on the following image

n
image ? arrays.
bl kn

Q. 13 What is meant by unsharp masking and crisping ?

1 3 2 2 4
Explain with relevant figures.
at

2 1
Pu ch

5 6 3
Te

Fig. Q. 18

Q. 19 Explain the power-law transformation.


 8 Histogram Modelling

Syllabus
Histogram Processing, Histogram equalization

8.1 Introduction Method 2

 Histogram of images provide a global description of In the second method, instead of plotting the number
the appearance of an image. The information obtained of pixels directly, we plot their probability of occurrence

ge
from histograms is enormous and hence histogram i.e.,
modelling; though a spatial domain technique is nk
p(rk) =
introduced as a separate chapter. n
io led th
 By definition, histogram of an image represents the Where, rk k grey level
relative frequency of occurrence of the various grey th
nk  Number of pixels in the k grey level
levels in an image. Histogram of an image can be
ic ow

plotted in two ways. n  Total number of pixels in an image


n
In the first method, the x-axis has the grey levels and Grey level Number of pixels nk
bl kn

the y-axis has the number of pixels in each grey level, (nk) p(rk) = n
while in the second method, the x-axis represents the
0 40 0.4
at

grey levels, while the y-axis represents the probability

Pu ch

1 20 0.2
of the occurrence of that grey level.
2 10 0.1
Method 1 3 15 0.15
Te

4 10 0.1
Grey level Number of pixels (nk)
5 3 0.03
0 40 6 2 0.02
1 20
n = 100
2 10
3 15
4 10
5 3
6 2

Fig. 8.1.2

 This is known as a normalized histogram. The

advantage of the second method is that the maximum
value to be plotted will always be 1. Generally black is
considered as grey level 0 and white as the maximum.
Fig. 8.1.1
 DSIP (MU-Sem. 7-COMP) 8-2 Histogram Modelling

 Just by looking at the histogram of the image, a great  Our aim in this chapter would be to transform the first
deal of information can be obtained. Some of the three histograms into the fourth histogram. In other
typical histograms are shown in Fig. 8.1.3. words, we try to increase the dynamic range of the
image. Though MATLAB has a inbuilt function to plot
the histogram (hist) we shall write our own code.
MATLAB code for plotting the histogram of an image.
%% Program for histogram %%
clear all
clc
a=imread(‘test.tif’);
a=double(a);
(a) [row col]=size(a);
h=zeros(1,300);
for n=1:1:row

ge
for m=1:1:col
if a(n,m)==0
%% To ensure that the values of a are not zero
io led
a(n,m)=1;
end
end
ic ow

(b) end
for n=1:1:row
n
for m=1:1:col
bl kn

t=a(n,m);
%% Takes the value of the pixel ex. 12
h(t)=h(t)+1;
at
Pu ch

%% Incrementing each counter h(12)

end
Te

end
(c) %% PLOTTING %%
figure (1)
imshow(uint8(a))
figure(2)
bar(h)

(d)
Fig. 8.1.3

 In Fig. 8.1.3(a) all the pixels belong to the lower grey

levels 0, 1, .... and hence we can be sure that the
image, represented by this histogram, is a dark image.
Similarly, Fig. 8.1.3(b) is the histogram of a bright
image. Fig. 8.1.3(c) is a low contrast image since only a
small range of grey levels are present. Fig. 8.1.3(d) is a
high contrast image of all the 4 histograms shown, the
last histogram represents the best image.
Fig. 8.1.4(a) : Original image
 DSIP (MU-Sem. 7-COMP) 8-3 Histogram Modelling

8.2 Linear Stretching

 One way to increase the dynamic range is by using a

technique known as histogram stretching.

Fig. 8.2.1

 In this method, we do not alter the basic shape of the

histogram, but we spread it so as to cover the entire

ge
dynamic range. We do this by using a straight line
Fig. 8.1.4(b) : Histogram of image
equation having slope (smax – smin) / (rmax – rmin).
io led
8.1.1 Mean and Standard Deviation of Where, smax  Maximum grey level of output image
Histogram smin  Minimum grey level of output image
rmax Maximum grey level of input image
ic ow

Histogram of an image provides a global description of

the appearance of an image. By definition, histogram of an rmin  Minimum grey level of input image
n
image represents the relative frequency of occurrence of smax – smin
bl kn

the various grey levels in an image. s = T(r) = (r – rmin) + smin …(8.2.1)

rmax – rmin
1. Mean (μ) : The Mean is a measure of central
at
Pu ch

tendency. Mean is calculated by multiplying and

adding the columns of the frequency table and diving
it by the total number of elements present in the
Te

image. The mean value best reflects the typical score

of a data set. Mean is also known as the Expected
value.
Consider the given frequency table.
Fig. 8.2.2
Grey level 1 2 3 4
 This transformation function shifts and stretches the
Number of pixels 5 12 10 5
grey level range of the input image to occupy the
The mean in this case will be = ((1  5) + (2  12) + (3  10) entire dynamic range (smin , smax).

+ (4  5))/32 = 2.46
8.2.1 Solved Examples on Linear Stretching
2. Standard deviation (σ) : In image processing, standard
deviation is a measure that is used to measure the Ex. 8.2.1 : Perform histogram stretching so that the new
amount of variation of grey levels from the mean. A image has a dynamic range of [0, 7].
standard deviation close to 0 indicates that the data
Grey 0 1 2 3 4 5 6 7
points tend to be very close to the mean of the image,
level
while a high standard deviation indicates that the data
points are spread out over a wider range of grey level Number 0 0 50 60 50 20 10 0
values. of pixels
 DSIP (MU-Sem. 7-COMP) 8-4 Histogram Modelling

Soln. : Ex. 8.2.2 : Perform histogram stretching so that the new

Here, rmin = 2 rmax = 6 image has a dynamic range of [0, 7].
smin = 0 smax = 7 Grey level 0 1 2 3 4 5 6 7
7 Number of pixels 100 90 85 70 0 0 0 0
  s = (r – 2) + 0
4
Soln. :
When, r = 2, s = 0
r = 3, s = 1.75  2
r = 4, s = 3.5  4
r = 5, s = 5.2  5
r = 6, s =7
 We have,
r=2 s=0
r=3 s=2

ge
r=4 s=4 Fig. P. 8.2.2

r=5 s=5 This histogram of the image is shown. As can be

io led
concluded, it is a dark image.
r=6 s=7
Here , rmin = 0, rmax = 3,
 Modified histogram is, smin = 0, smax = 7
ic ow

smax – smin
Grey level 0 1 2 3 4 5 6 7 s = (r – rmin) + smin
rmax – rmin
n
Number of pixels 50 0 60 0 50 20 0 10 7
bl kn

 s = (r – 0) + 0
3
When, r = 0, s = 0
at
Pu ch

7
r = 1, s = (1 – 0) = 2.3  2
3
7
r = 2, s = (2 – 0) = 4.67  5
Te

3
7
r = 3, s = (3 – 0) = 7
3
 We have,
r=0 s=0
r=1 s=2
r=2 s=5
r=3 s=7
Hence the histogram tends to spread out as shown in
Fig. P. 8.2.2 (a).

Fig. P. 8.2.1

We see that the shape of the histogram is retained but

it is spread out or stretched.
Let us take another example of histogram linear
stretching. Fig. P. 8.2.2(a)
 DSIP (MU-Sem. 7-COMP) 8-5 Histogram Modelling

MATLAB code for histogram stretching %%

%% Histogram stretching %%
clear all
clc
a=imread(‘pout.tif’);
a=double(a);
[row col]=size(a);
a=a+1;
w=min(min(a));
constant=255/(max(max(a))-min(min(a)));
cmin=0;
for x1=1:1:row (a) Original image

ge
for y1=1:1:col
c(x1,y1)=constant*(a(x1,y1)-w)+cmin;
io led
end
end
ic ow

%% Plotting histogram of the original image %%

n
h=zeros(1,300);
bl kn

for n=1:1:row
at

for m=1:1:col
Pu ch

t=a(n,m); %% Takes the value of the pixel for example 12

h(t)=h(t)+1; %% Incrementing the corresponding counter
Te

h(12)
end
end (b) Image after histogram stretching

figure(1), bar(h)
% Plotting histogram of the modified image %
c=c+1;
h=zeros(1,400);
for n=1:1:row
for m=1:1:col
t=round(c(n,m));
h(t)=h(t)+1;
end
end
figure(2),bar(h)
figure(3),imshow(uint8(a))
figure(4),imshow(uint8(c)) (c) Original histogram

Fig. 8.2.3 : Cont…

 DSIP (MU-Sem. 7-COMP) 8-6 Histogram Modelling

We know, s = T(r)
What is this T which produces equal values in each grey
level ?
 The transformation that we are looking for must
satisfy the following two conditions :
(a) T(r) must be single valued and monotonically
increasing in the interval 0  r  1 and
(b) 0  T(r)  1 for 0  r  1 i.e., 0  s  1 for 0  r  1
Here the range of r is taken as [0, 1]. This is called the
normalized range. This range is taken for simplicity. So
instead of r being in the range [0, 255] we take [0, 1]
(d) Linear stretched histogram Let us see what these two conditions :
Fig. 8.2.3

ge
(1) If T (r) is not single valued. Refer Fig. 8.3.2(a). Here r1
8.3 Histogram Equalization and r2 are mapped as a single grey level s1 i.e., two
different grey levels look the same in the modified

io led
Linear stretching is a good technique but as mentioned image. This is a big drawback, hence the
earlier, the shape remains the same. transformation has to be single valued. This preserves
 There are many applications, wherein we need a flat the order from black to white.
ic ow

histogram. This cannot be achieved by histogram A grey level transformation that is both single valued
stretching. and monotonically increasing is shown in Fig. 8.3.2(b).
n
 We now introduce a new technique known as
bl kn

histogram equalization. A perfect image is one which

has equal number of pixels in all its grey levels. Hence
at
Pu ch

our objective is not only to spread the dynamic range,

but also to have equal pixels in all the grey levels.
Te

 This technique is known as histogram equalization. We

now search for a transformation that would transform
a bad histogram to a flat histogram
(a) (b)
Fig. 8.3.2

(2) If condition (b) is not satisfied, then the mapping will

not be consistent with the allowed range of pixel
values. In other words, s will go beyond the valid
range.
Since the transformation is single valued and
monotonically increasing, the inverse transformation
exists.
–1
 r = T (s) ; 0  s  1
 The grey levels for continuous variables can be
characterized by their probability density functions
pr(r) and ps(s).
 From probability theory we know that if pr(r) and T(r)
–1
are known and if T (s) satisfies condition (a), then the
probability density of the transformed grey level is
Fig. 8.3.1
 DSIP (MU-Sem. 7-COMP) 8-7 Histogram Modelling

ps (s) = pr(r) dr …(8.3.1) Note : pr(r) is the PDF and hence cannot be greater
 ds –1
r =T (s) than 1.
i.e., the probability density of the transformed image is Soln. :
equal to the probability density of the original image
From the diagram, we know that the above histogram
multiplied by the inverse slope of the transformation. represents a dark image.
 We now need to find a transformation which would – r + 1 ;0r1
pr(r) = 
give us a flat histogram. Let us consider the Cumulative 0 ; otherwise
Density Function (CDF). Cumulative density function is We shall use the CDF that we just discovered.
obtained by simply adding up all the Probability r r
Density Functions (PDF).
s = T (r) =  pr (r)dr =  (– r + 1) dr
s = T (r) 0 0
r –r
2

=  pr (r) dr ; 0r1 …(8.3.2) s =

2
+r

ge
0 2s = – r + 2r
2

Differentiating with respect to r, we get 2

2s –1 = – r + 2r –1
ds
io led 2
= pr (r) …(8.3.3) (1 – 2s) = (r – 2r + 1)
dr
(1 – 2s) = (r – 1)2
Using Equation (8.3.3) in Equation (8.3.1)
(r – 1) =  1 – 2s
ic ow

ps(s) = [ 1 ] = 1; 0  s  1
i.e., ps(s) = 1 r = 1  1 – 2s ··· 0  r  1
n
This is nothing but a uniform density function !!  r = 1 – 1 – 2s
bl kn

Let us see what we have got here. We know,

pr (r) dr = (– r + 1) dr
at

ps (s) =
 ds  ds
Pu ch

Substituting the value of r,

Te

d
ps (s) = (– 1 + 1 – 2s + 1)
ds
{1 – (1 – 2s) 1/2}
d
Fig. 8.3.3 ps (s) = 1 – 2s 
ds
{– (1 – 2s) 1/2}
A bad histogram becomes a flat histogram when we
∵ d (1) = 0 
find the cumulative density function!! Let us cement this  ds 
concept by taking an examples. –1 
1 – 2s   (1 – 2s) (– 2) 
–1/2
=
 2 
8.3.1 Solved Examples on Histogram 1
= 1 – 2s 
Equalization 1 – 2s
 ps (s)= 1 ;
Ex. 8.3.1 For the given histogram. Perform histogram Hence ps(s) = 1; 0  r  1.
equalization.

Fig. P. 8.3.1
Fig. P. 8.3.1(a)
 DSIP (MU-Sem. 7-COMP) 8-8 Histogram Modelling

We see that pr(r) which represented a dark histogram

has been transformed to an equalized histogram.

Ex. 8.3.2 : Though pr(r) cannot be greater than 1, you could

just consider it as another problem and solve it
pr(s) = {
– 2r + 2
0
;0r1
; otherwise

Fig. P. 8.3.3 : Original dark histogram

Grey nk PDF CDF (L – 1)  sk Rounding

level nk sk =  i.e. 7  sk off
pr (rk) = n

ge
Fig. P. 8.3.2 pr(rk)

Soln. : 0 790 0.19 0.19 1.33 1

io led
Try out this problem yourself and see what you get.
1 1023 0.25 0.44 3.08 3
We see that CDF gives us a flat response no matter
2 850 0.21 0.65 4.55 5
what the shape of the original histogram is.
ic ow

So far, we have developed the technique of histogram 3 656 0.16 0.81 5.67 6

equalization using continuous domain mathematics. For

n
4 329 0.08 0.89 6.23 6
bl kn

image processing, we need to work in the discrete domain.

5 245 0.06 0.95 6.65 7
Hence if n is the total number of pixels in an image,
at

then the PDF is 6 122 0.03 0.98 6.86 7

Pu ch

nk 7 81 0.02 1 7 7
pr (rk) =
n
Te

N = 4096
r

st nd
We know, s = T (r) = pr (r) We take 1 , 2 and the last column
0
Old grey level Equalized grey level New grey level
In the discrete domain,
r 0 790 1
sk = T (rk) =  pr (r). This is the C. D. F
1 1023 3
0
Let us take up an example to see how this works in the 2 850 5
discrete domain. Some additional steps need to be taken
into account while working in the discrete domain. 3 656 6

4 329 6
Ex. 8.3.3 : Equalize the given histogram
Grey level 0 1 2 3 4 5 6 7 5 245 7
Number of 790 1023 850 656 329 245 122 81
pixels
6 122 7

Soln. : 7 81 7
L = 8 (Number of grey levels) We notice that the new grey levels have pixels only at
We first plot the original histogram 1, 3, 5, 6, 7. There are no pixels in grey levels 0, 2 and 4.
 DSIP (MU-Sem. 7-COMP) 8-9 Histogram Modelling

Equalized grey level Number of pixels Soln. :

Grey nk nk sk = sk  Rounding
0 0 pr (rk) = n
level  pr(rk) (L – 1) off
1 790 i.e. sk  7

2 0 0 100 0.384 0.384 2.688 3

1 90 0.346 0.73 5.11 5
3 1023
2 50 0.1923 0.9223 6.456 6
4 0 3 20 0.0769 1 7 7
5 850 4 0 0 1 7 7
5 0 0 1 7 7
6 656 + 329 = 985
6 0 0 1 7 7
7 245 + 122 + 81 = 448
7 0 0 1 7 7
n = 260
st nd
We take the 1 , 2 and the last column.

ge
Old nk New grey levels New Modified
grey nk grey levels
io led levels
0 100  3  100 00

1 90  5 90 10
ic ow

2 50  6 50 2 0
n
Fig. P. 8.3.3(a) : Equalized histogram
3 20 7 3  100
bl kn

So here we are. A dark histogram becomes an evenly 4 7 40

0
spaced histogram. But wait a minute, we had started out to
at

5 0 7 20 + 0 + 0 20 5  90
Pu ch

get a flat histogram as in the continuous case but the

+ 0 + 0 = 20
histogram obtained is not flat; why?
6 0 7 6  50
Te

Discrete domain is an approximation of the continuous

7 0 7 7  20
domain i.e., values between 0 and 1, 1 and 2 and so on are
not known. Due to this reason, perfectly flat results are The original histogram showed that the image had
never obtained. only 0-3 grey levels. After histogram equalization we see
that the dynamic range has increased. Hence histogram
Ex. 8.3.4 equalization does increase the dynamic range.
Grey nk Next question that is usually asked is what happens
level when we equalize an already equalized histogram. How
many times can we equalize an image ?
0 100
1 90
2 50
3 20
4 0
5 0
6 0
7 0

Fig. P. 8.3.4

Equalize the above histogram (Here L = 8) Fig. P. 8.3.4(a)

 DSIP (MU-Sem. 7-COMP) 8-10 Histogram Modelling

Equalize Equalize again ? change in the histogram. Go ahead, try out your own
i.e. Dark historgram –––––– Flat historgram –––––––––––
examples. They will all give similar results.
We shall see what happens by considering an example.
Histogram equalization is guaranteed to yield exact
Ex. 8.3.5 : Given a histogram, see what happens when we results only in the continuous case. As the number of levels
decrease, the error between the specified and resulting
equalize it twice.
histogram increase and hence we do not get a flat
Grey level 0 1 2 3 histogram.
MATLAB code for histogram equalization
nk 70 20 7 3
%% Histogram equalization %%
Soln. : clear all
clc
nk
Grey nk pr (rk) = n sk =  pr(rk) L=4 Rounding New nk a=imread(‘kids2.tif’);
level sk  (L – 1) off % You can try the program with ‘pout.tif’%

ge
0 70 0.7 0.7 2.1 2  70
a=double(a);
big=max((max(a)));
1 20 0.2 0.9 2.7 3
%% Gives maximum value in the image
io led
2 7 0.07 0.97 2.91 3  20 + 7 + 3 [row col]=size(a);
C=row*col; %% Gives the total number of pixels required
3 3 0.03 1 3 3 = 30
ic ow

h=zeros(1,300);
L=4 100 %% Defining two arrays (h and z) to store the histogram
n
The modified grey levels are values
bl kn

z=zeros(1,300);
Grey level 0 1 2 3 for n=1:1:row
at

for m=1:1:col
Pu ch

nk 0 0 70 30
if a(n,m)==0
We now equalize this again. %% To ensure that the values of ‘a’ are not zero
Te

Grey nk pr (rk) sk L = 4 sk Rounded Modified

a(n,m)=1;
level  (L – 1) off grey level end
end
0 0 0 0 0 0 end
1 0 0 0 0 0 %% Plotting the histogram of the original image %%
for n=1:1:row
2 70 0.7 0.7 2.1 2 70 for m=1:1:col
3 30 0.3 1 3 3 30 t=a(n,m);
%% Takes the value of the pixel example 12 %%
100 h(t)=h(t)+1;
%% Incrementing the corresponding counter h(12)%%
 we get, end
end
Grey level 0 1 2 3 pdf=h/C; %% Probability function %%
nk 0 0 70 30 cdf(1)=pdf(1);
%% So that we don’t have the condition x(0) in the
Which is the same as what we had got after the first for loop below %
equalization.
for x=2:1:big
Hence, equalizing again gives us the same result i.e., cdf(x)=pdf(x)+cdf(x-1); %% Calculating the cdf %%
equalizing an already equalized histogram causes no
 DSIP (MU-Sem. 7-COMP) 8-11 Histogram Modelling

end
new=round(cdf*big); %% New is the new gray level
new=new+1; %% if new = 0, it is not to be taken in
the loop
for p=1:1:row
for q=1:1:col
temp=a(p,q);
b(p,q)=new(temp);
%% b has the equalized image %%
%% Plotting the equalised histogram %%
t=b(p,q);
%% Takes the value of the pixel example 12 %%
z(t)=z(t)+1; %% Incrementing the corresponding
counter z(12)%%

ge
end
end
io led
b=b-1;
%% Since we had initially incremented the value of n %% (b) Image after histogram equalization
%% Plotting %%
ic ow

figure (1), imshow(uint8(a))

figure (2), bar(h)
n
figure (3), imshow(uint8(b))
bl kn

figure (4), bar(z)

at
Pu ch
Te

(c) Original histogram

Fig. P. 8.3.5(a) : Original image

(d) Histogram equalized

Fig. P. 8.3.5
 DSIP (MU-Sem. 7-COMP) 8-12 Histogram Modelling

8.4 Histogram Specification Soln. :

 From the earlier method (Histogram equalization) we

note that histogram equalization is automatic. It is not
interactive. i.e., it always gives one result - an
approximation to an uniform histogram. It is at times
desirable to have an interactive method in which
certain grey level are highlighted.
 It should be noted that if we modify the grey level of
an image that has a uniform PDF, ps(s), using the
–1
inverse transformation r = T (s), we get back the
original histogram pr(r). Exploiting this knowledge, we Fig. P. 8.4.1(a)
can obtain any shape of the grey level histogram by

ge
processing the given image in the following way.
 Suppose pr(r) and ps(s) represent grey level PDFs of
io led
input and output images and r and s are their
respective grey levels. Suppose k represents grey level
of some intermediate image result i.e.,
ic ow

r s
k = T1 (r) =  pr (r) dr and k = T2 (r) =  ps (s)ds
n
0 0
bl kn

 In both these cases pk(k) is uniform. Hence the

–1
transformation s = T 2 (T1(r)) achieves the desired Fig. P. 8.4.1(b)
at
Pu ch

result. The procedure of histogram specification can Now equalize (a) as well as (b)
be computed as follows : Histogram (a) : L = 8
Te

(1) Equalize the levels of the original image. Grey nk pr (rk) = CDF = CDF Rounding New
nk  pr(rk)
(2) Specify the desired density function and obtain level  off grey
n
the transformation function. level
(L – 1)
(3) Apply the inverse transformation function.
0 790 0.19 0.19 1.33 1  790

8.4.1 Solved Example Histogram 1 1023 0.25 0.44 3.08 3  1023

2 850 0.21 0.65 4.55 5  850

Ex. 8.4.1 : Given histograms (a) and (b), modify histogram a
as given by histogram (b). 3 656 0.16 0.81 5.67 6  656 +
329 =
Histogram (a) 4 329 0.08 0.89 6.23 6
985
Grey 0 1 2 3 4 5 6 7
5 245 0.06 0.95 6.25 7 
Number 790 1023 850 656 329 245 122 81
of pixels 6 122 0.03 0.98 6.86 7  245 +
122 +
Histogram (b)
81
Grey 0 1 2 3 4 5 6 7
7 81 0.02 1 7 7  = 448
Number of 0 0 0 614 819 1230 819 614
pixels
4096
 DSIP (MU-Sem. 7-COMP) 8-13 Histogram Modelling

Note that there are 5 distinct levels after histogram 8.5 Solved Examples on Histogram
equilization of histogram (a) Modelling
Grey 0 1 2 3 4 5 6 7
level Ex. 8.5.1 : Two images can have the same histogram

Number 0 790 0 1023 0 850 985 448 (Justify/ Contradict with reason). MU - Dec 2017, 5 Marks
of pixels Soln. :

Now equalize (b) : L = 8 Histograms give us information about the numbers of

grey levels present in the image, it does not give us any
Grey nk nk CDF CDF  New
pr (rk) = n spatial information. This means, histograms do not tell us
level (L – 1) grey
where the grey levels lie in an image. If all pixels in an
level
image are shuffled, there will be no change in the
0 0 0 0 0 0 histogram. For example, the given two images are
1 0 0 0 0 0 completely different but their histograms will be equal as
the total number of black pixels and white pixels are the

ge
2 0 0 0 0 0
same in both the images. Hence two images can have the
3 614 0.149 0.149
io led 1.05 1 same histogram.
4 819 0.20 0.35 2.45 3
5 1230 0.30 0.65 4.55 5
ic ow

6 819 0.20 0.85 5.97 6

7 614 0.15 1 7 7 Fig. P. 8.5.1
n
4096
bl kn

Ex. 8.5.2 : What effect would setting to zero the lower order
To obtain histogram specification, we apply the bit planes have on the histogram of an image in general ?
at

inverse transform comparing equalized (a) and (b).

Pu ch

Soln. : The number of pixels having different grey level

values would decrease, thus causing the number of
components in the histogram to decrease. Since the
Te

1  3 number of pixels is constant, this would cause the height of

the remaining histogram peaks to increase. Typically, less
3  4
variability in grey level values will reduce the contrast.
5  5
Let us check this out with an example.
6  6
Given below is a 4  4 image having grey levels 0-18.
7  7
0 1 2 3
4 5 6 7
8 9 10 11

Fig. P. 8.4.1(c) 12 13 14 15

Hence the final result is as shown Histogram of the original image is as shown in
Fig. P. 8.5.2.
Grey 0 1 2 3 4 5 6 7
Number 0 0 0 790 1023 850 985 448
of pixels
Hence we see that histogram (a) has been equalized
and mapped in such a way that it resembles the histogram
(b). This is histogram specification.
Fig. P. 8.5.2
 DSIP (MU-Sem. 7-COMP) 8-14 Histogram Modelling

We convert the original image to a binary image. Since Comparing this histogram with the original histogram,
the maximum grey level is 15, we need 4-bits for its we see that the variability is reduced, i.e. number of grey
representation. levels are reduced and the height of these grey levels has
increased.
0000 0001 0010 0011
Ex. 8.5.3 : What would be the effect on the histogram if we

0100 0101 0110 0111 set to zero, the higher order bit planes.
Soln. :
1000 1001 1010 1011 Proceeding in a similar manner as in Ex. 8.5.2 and
working with the same image, we get
1100 1101 1110 1111
Converting to binary

Suppose we make the last two bits equal to zero, we

0 1 2 3 0000 0001 0010 0011
get

ge
4 5 6 7 0100 0101 0110 0111
0000 0000 0000 0000
8 9 10 11 1000 1001 1010 1011
io led
0100 0100 0100 0100
12 13 14 15 1100 1101 1110 1111

1000 1000 1000 1000 (a) (b)

ic ow

Converting the upper two Converting back to grey

1100 1100 1100 1100
n
bits to zero levels we get
bl kn

Converting back to the grey level format we get, 0000 0000 0000 0000 0 1 2 3
at

0100 0100 0100 0100 0 1 2 3

Pu ch

0 0 0 0
1000 1000 1000 1000 0 1 2 3
4 4 4 4
Te

1100 1100 1100 1100 0 1 2 3

8 8 8 8
(c) (d)
12 12 12 12
Fig. P.8.5.3
Plotting the histogram as shown in Fig. P. 8.5.3(e).
its histogram is as shown in Fig. P. 8.5.2(a).

Fig. P. 8.5.3 (e)

Fig. P. 8.5.2(a) In this case too, the grey levels reduce. But one
important thing that happens is that the image becomes
much darker.

Ex. 8.5.4 : A digital image with 8 quantization levels is given below :

Perform histogram equalization.

f(x,y) = |x – y| for x = 0 to 7 and y = 0 to 7 MU - Dec. 2018, 10 Marks

 DSIP (MU-Sem. 7-COMP) 8-15 Histogram Modelling

Soln. :

In this we form an image by taking the modulus of the difference of the spatial coordinates. The image that is formed
using the equation f(x,y) = |x – y| is shown below :

0 1 2 3 4 5 6 7

0 0 1 2 3 4 5 6 7

1 1 0 1 2 3 4 5 6

2 2 1 0 1 2 3 4 5

f(x, y) = 3 3 2 1 0 1 2 3 4

4 4 3 2 1 0 1 2 3

ge
5 5 4 3 2 1 0 1 2

6 6 5 4 3 2 1 0 1
io led
7 7 6 5 4 3 2 1 0
ic ow

We generate the frequency table / histogram of the original image.

n
Grey levels 0 1 2 3 4 5 6 7
bl kn

Number of pitch 8 14 12 10 8 6 4 2
at
Pu ch

The original histogram is shown in Fig. P.8.5.4.

Te

Fig. P. 8.5.4
DSIP (MU-Sem. 7-COMP) 8-16 Histogram Modelling

We now, generate a table to perform histogram equalization

Grey level nk nk Sk =pr (rk) Sk (L – 1) Rounding off

pr (rk) = n
Sk 7

0 8 0.125 0.125 0.875 1

1 14 0.218 0.343 2.401 2

2 12 0.187 0.53 3.71 4

3 10 0.156 0.686 4.802 5

4 8 0.125 0.811 5.677 6

5 6 0.093 0.904 6.328 6

ge
6 4 0.062 0.966 6.762 7

7 2 0.031 1 7 7
io led
nk= 64
ic ow
n
bl kn

st nd
We consider the 1 , 2 and last column
at

Greg level nk Rounding off New nk

Pu ch

0 8 1 8
Te

1 14 2 14

2 12 4 12

3 10 5 10

4 8 6 8 + 6 14

5 6 6

6 4 7 4+2 6

7 2 7

Hence the modified frequency table is

Grey levels 0 1 2 3 4 5 6 7

nk 0 8 14 0 12 10 14 6

The equalized histogram is shown in Fig. P. 8.5.4(a).

 DSIP (MU-Sem. 7-COMP) 8-17 Histogram Modelling

ge
io led
Fig. P. 8.5.4(a)
ic ow

From the above formulation

Ex. 8.5.5 : Given below is the slope transformation and the
image. Draw the frequency tables for original and new when, m = 2, l = 0
n
image. Assume lmin = 0 and lmax = 7. m = 3, l = 1.75  2
bl kn

2 3 4 2 m = 4, l = 3.5  4
at

m = 5, l = 5.25  5
Pu ch

5 5 2 4
m = 6, l = 7
3 6 3 5
The modified image is
Te

5 3 5 5 0 2 4 0

5 5 0 4

2 7 2 5

5 2 5 5

Histogram of the original and modified images are

shown in Fig. P. 8.5.5(a).

Fig. P. 8.5.5

Soln. : This is a sum of histogram linear stretching. By

frequency tables, we mean histogram. From the image, it is
clear that mmin = 2, mmax = 6, lmin = 0 and lmax = 7.
Using the formula
lmax – lmin
l = (m – mmin) + lmin
mmax – mmin
7–0
 l = (m – 2) + 0
6–2 Fig. P. 8.5.5(a)
7
 l = (m – 2)
4
 DSIP (MU-Sem. 7-COMP) 8-18 Histogram Modelling

Soln. :
We draw the original histogram of the image.

Fig. P. 8.5.5(b)

Ex. 8.5.6 : A popular procedure for image enhancement Fig. P. 8.5.7

combines high frequency emphasis and histogram Here L = 8 We now perform histogram equalization.

ge
equalization to achieve edge sharpening and contrast
Grey nk PDF CDF sk  (L – 1) Round
enhancement. Does it matter which of the two processes are
level nk i.e., i.e., sk  7 off
applied first ? pr (k) =
n
io led sk =
Soln. :
 pr(k)
High frequency emphasis is the same as a high boost
filter. Let us see what we get when we apply a high 0 0 0 0 0 0
ic ow

frequency emphasis filter. 1 50 0.05 0.05 0.35 0

n
2 100 0.1 0.15 1.05 1
bl kn

3 200 0.2 0.35 2.45 2

at

4 400 0.4 0.75 5.25 5

Pu ch

i.e., the output g(x, y) is a convolution of the input

image and the impulse response of the high boost filter. 5 200 0.2 0.95 6.65 7
If we now perform histogram equalization we get
Te

6 50 0.05 1 7 7
s = T [g (x, y)] 7 0 0 1 7 7
s = T [f (x, y) * h (x, y)] …(1)
n = 1000
If we decide to apply histogram equalization prior to st nd
Consider the 1 , 2 and last column
high boost filter, we get

Grey nk New grey level

level
…(2) 0 0 0 0 + 50
We see that Equation (1)  Equation (2). 1 50 0 = 50
2 100 1 100
The order of the operation does matter i.e., the final
3 200 2 200
result varies if we interchange the two operations.
4 400 5 400
Ex. 8.5.7 : The grey level distribution of an image is 5 200 7
shown in the table below. Perform histogram equalization 6 50 7 200 + 50 = 250
and plot the original and equalized histograms. 7 0 7
Grey level 0 1 2 3 4 5 6 7 Hence the equalized frequency table is,
Frequency of 0 50 100 200 400 200 50 0 Grey level 0 1 2 3 4 5 6 7
occurrence nk 50 100 200 0 0 400 0 250
(MU - Dec. 2013, 10 Marks The equalized histogram is shown in Fig. P. 8.5.7(a)
 DSIP (MU-Sem. 7-COMP) 8-19 Histogram Modelling

Old grey nk New grey level Modified grey

level level
0 128 0  128 + 0  303
1 75 0 175 1  280
2 280 1  280 2  416
3 416 2  416 3  635
4 635 3  635 4  1058
Fig. P. 8.5.7(a) 5 1058 4  1058 50

Ex. 8.5.8 : A 64  64 image, represented by 3 bit / pixel 6 820 6  820 6  820

has the following grey level distribution : 7 684 7  684 7  684

ge
Gray 0 1 2 3 4 5 6 7 We now draw the original as well as the equalized
level histogram.
No. of 128 75 280 416 635 1058 820 684
io led
Pixel
ic ow

Perform histogram equalization and give new distribution of

gray level. Show plots of original and equalized image.
n
bl kn

(MU - May 2011, Dec. 2012, 10 Marks) (a) Original histogram (b) Equalized histogram
Soln. : Fig. P. 8.5.8
at
Pu ch

We generate the histogram table

Summary
Grey nk PDF CDF (L – 1) Rounding
Te

level pr (rk) = sk =  sk off In this chapter, the concept of histograms is

nk  pr(rk) introduced. Though histogram processing is a spatial
n
domain technique, it is presented here as a separate
0 128 0.031 0.031 0.217 0 chapter because of its wide spread applications. Histograms
give us an idea about the quality of an image. One of the
1 75 0.018 0.0496 0.347 0
most common defects found in recorded images is poor
2 280 0.068 0.1179 0.825 1 contrast. This degradation may be caused by inadequate
lighting, aperture size and/or shutter speed. Procedures of
3 416 0.101 0.2195 1.536 2
modifying the histogram to get a high contrast image are
4 635 0.155 0.3745 2.622 3 discussed. Emphasis has been laid on the two important
techniques: linear stretching and histogram equalization.
5 1058 0.258 0.6328 4.430 4

6 820 0.201 0.8330 5.831 6 Review Questions

7 684 0.1670 1 7.00 7 Q. 1 Plot the graph of a cumulative density function (CDF).

n= Q. 2 Explain how histogram can aid in the process of

4096 thresholding and contrast stretching.
Q. 3 Equalize the given histogram.
st nd
We take the 1 , 2 and the last column
 DSIP (MU-Sem. 7-COMP) 8-20 Histogram Modelling

Q. 8 Equalize the given histogram.

Grey 0 1 2 3 4 5 6 7

Number of pixels 50 0 50 0 50 0 50 0

Q. 9 Compare between contrast stretching and histogram

equalization.
Fig. Q. 3
Q. 10 Suppose we have a dark image which needs to be
Note : pr(r) cannot be greater than 1 but solve this as any
compressed and also equalized. Which operation
other mathematical problem.
would we use first ? Will the results be the same if the
Q. 4 Explain why a discrete histogram technique does not, order of operations is reserved ?
in general, yield a flat histogram.
Q. 11 Equalize the given image.

ge
Q. 5 Suppose that a image is subjected to a histogram
0 1 2
equalization. Show that a second pass of histogram
equalization will produce exactly the same result as
io led 3 4 5
the first pass.
6 7 8
Q. 6 What are the two conditions that a transformation has
ic ow

to satisfy ? Q. 12 Given the frequency table,

n
Q. 7 Write down an algorithm to plot the histogram of an Grey 0 1 2 3 4 5 6 7
bl kn

image. Number of pixels 0 0 a b c d e 0

at
Pu ch

Perform linear stretching.

Te


 9 Segmentation

Syllabus :
Segmentation based on Discontinuities (point, Line, Edge), Image Edge detection using Robert, Sobel, Previtt masks,
Image Edge detection using Laplacian Mask. Connectivity

 In an automated blood cell counting machine, this

9.1 Introduction classification has to be made by the machine

ge
May 2016, Dec. 2016, Dec. 2018, 2 Marks (computer). The computer scans the slide and based
on certain algorithms classifies them as RBC’s,
Q. What is image segmentation
WBC’s....etc.
io led
(May 2016, Dec. 2016, Dec. 2018, 2 Marks)  We hence needed to do image segmentation to divide
 Segmentation, as the name suggests, subdivides the image into various parts.
(segments) an image into its constituent regions or (2) Finger print matching in forensic studies
ic ow

objects. Unlike image enhancement, were our primary

 In this, there exists a data base of finger prints of
concern was to improve the subjective quality of
n
people who have been convicted of crime at some
images for display, in segmentation, we address some
bl kn

aspects of analyzing the content of an image. point in time.

 This simply means we endeavour to find out what is in  Now if there has been a crime at a place and some
at
Pu ch

the picture. Segmentation forms a section of computer finger prints have been found, images are taken, and
vision i.e., we use segmentation, when we want the these images are run through an algorithm to see if
computer to make decisions. this particular finger print matches with the one in the
Te

 Segmentation is used when we need to automate a data base.

particular activity. The ultimate aim in an automated  Here too, since the system is automated we need to
system is to extract important features from image use segmentation to divided the image into various
data, from which description, interpretation, or parts.
understanding of the scene can be provided by the
 We hence say that image segmentation is used for
machine.
computer decision making (machine vision) and not
 Segmentation is not required when the images have to
for human interpretation and the main purpose of
be shown to a human being. This is because a human
segmentation is to partition the image.
visual system has an inherent quality to segment the
image shown to it.  Image segmentation algorithms are generally based on
Let us take two simple examples where image one of the two basic properties i.e. image
segmentation is used. segmentation can be achieved in any one of the two
ways viz.
(1) Automated blood cell counting :
(1) Segmentation based on discontinuities in
 We all know blood is made up of RBC’s, WBC’s,
intensity.
platelets etc.
(2) Segmentation based on similarities in intensity.
 In manual machines, the technician looks at the slide
and based on various parameters classifies them as In the first method, the approach is to partition an
RBC’s, WBC’s etc. As stated earlier, the human visual image based on abrupt changes in intensity, such as
system automatically segments the image as RBC’s, edges, while in the second method, we partition an
WBC’s... etc. image into regions that are similar according to a set
of predefined criteria.
 DSIP (MU-Sem. 7-COMP) 9-2 Segmentation

9.2.2 Line Detection

 Detection of lines can be done using the masks shown
below. In an image, lines can be in any direction and
detecting these lines would need different masks.
Fig. 9.1.1
 The syllabus only contains Segmentation based on –1 –1 –1 –1 –1 2
discontinuities 2 2 2 –1 2 –1
9.2 Image Segmentation based on –1 –1 –1 2 –1 –1
Discontinuities
Horizontal + 45
When we look at an image, we immediately observe 3
basic types of discontinuities in the grey level viz. points, –1 2 –1 2 –1 –1
lines and edges. The easiest way is to use spatial masks

ge
–1 2 –1 –1 2 –1
which have properties to detect these discontinuities.
–1 2 –1 –1 –1 2
9.2.1 Point Detection
io led
Vertical – 45
 Remember one thing, points, lines and edges are all
high frequency components and hence we need masks  We notice that all these masks have a sum equal to
ic ow

which are basically high pass. Hence the masks that we zero, and hence all of them are high pass masks. The
present here for point, line and edge detection have first mask would respond strongly to lines that are
n
the properties of a high mask mainly, the sum of the oriented horizontally. The second mask would respond
bl kn

coefficients of the mask has to be equal to zero in all to lines at an angle of + 45, the third mask would
the three cases. respond strongly to vertical lines and the last mask
would respond strongly to lines at an angle of 45.
at

 Detecting points is fairly simple. We use the standard

Pu ch

high pass mask for it. And so that this mask detects
only points and not lines, we set a threshold value i.e.,
Te

we say a point has been detected at the location on

which the mask is centered only if

–1 –1 –1

–1 8 –1

–1 –1 –1
|R|  T …(9.2.1) Fig. 9.2.1

Where R is derived from the standard convolution Let us take a simple example of a 8  8 pseudo image.
formula
0 0 0 10 0 0 0 0
R = w1 z1 + w2 z2 + ...........+ w9 z9
9 0 0 0 10 0 0 0 0
R =  wi zi  0 0 0 10 0 0 0 0
i=1
0 0 0 10 0 0 0 0
 We take |R| because we want to detect both the kinds
of points i.e. white points on a black background as 0 10 10 10 10 10 10 0
well as black points on a white background. 0 0 0 10 0 0 0 0
T is a non negative threshold which is defined by the
0 0 0 10 0 0 0 0
user.
0 0 0 10 0 0 0 0
 DSIP (MU-Sem. 7-COMP) 9-3 Segmentation

 It is quite evident that in this image we have two lines,  We see that the horizontal mask detects lines only in
one horizontal and the second vertical. the horizontal direction and cleans up the other lines
 Let us see what happens when we move the first (vertical line in this case).
(Horizontal) mask over the image. By now you should  In a similar fashion we can show that vertical, + 45,
be quite comfortable with moving masks on the entire and – 45 masks respond strongly to lines in their
image. We leave the borders and hence start from the respective directions. It would be a good idea to try
encircled pixel. this out yourself.
The result that we obtain is as shown
9.2.3 Edge Detection
0 0 0 0 0 0 0 0  More than isolated points and lines, it is the detection
0 0 0 0 0 0 0 0 of edges that form an important part of image
segmentation. An edge can be defined as a set of
0 0 0 0 0 0 0 0
connected pixels that form a boundary between two
0 – 20 – 20 – 20 – 20 – 30 – 20 0 disjoint regions.

ge
0 + 40 + 40 + 40 + 40 + 60 + 40 0  A typical example of an edge is shown in Fig. 9.2.2.

0 – 20 – 20 – 20
io led
– 20 – 30 – 20 0
0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0
ic ow

 Since line detection masks are high pass masks, we

n
Fig. 9.2.2
encounter negative values.
bl kn

 These values are made zero as was discussed in  Here the set of pixels that separate the black region
chapter of Image Enhancement in Spatial Domain. from the grey region is called an edge. The line profile
at
Pu ch

Hence the final image obtained as follows, of such an edge is shown along with edge.
 The image shown has a step edge. Such step edges
0 0 0 0 0 0 0 0
occur only in artificially generated images such as test
Te

0 0 0 0 0 0 0 0 patterns. Images obtained from digitization of optical

0 0 0 0 0 0 0 0 images of real scenes do not possess step edges.
Recollect what happens when we digitize an image.
0 0 0 0 0 0 0 0
 We first sample the image using the Nyquist condition
0 40 40 40 40 60 40 0 and then quantize it. Real world signals are not
band-limited i.e. F(u, v)  0 outside an interval and
0 0 0 0 0 0 0 0
maximum frequency present in it is infinite. Hence the
0 0 0 0 0 0 0 0 Nyquist criteria is not met. We therefore pass the
0 0 0 0 0 0 0 0 original signal through an anti-aliasing filter (which is a
low pass filter) to remove the very high frequencies.

Fig. 9.2.3

 It is this anti-aliasing filter which reduces the slope of the edge. Due to this, real world images when captured through
a camera will not have step edges. In practice, edges are modelled as having a ramp like profile.
DSIP (MU-Sem. 7-COMP) 9-4 Segmentation

Fig. 9.2.4

Here the slope of the ramp is inversely proportional to the degree of blurring.
Given below are some of the two dimensional, discrete domain edge models.

a a a a a b b b b b a a a b b b b b b b

a a a a a b b b b b a a a a b b b b b b

a a a a a b b b b b a a a a a b b b b b

ge
a a a a
io led a b b b b b a a a a a a b b b b

a a a a a b b b b b a a a a a a a b b b

Vertical step edge Diagonal step edge

ic ow

a a a a a a a a a a a a a a c b b b b b

a a a a a a a a a a a a a a c b b b b b
n
bl kn

a a a a a b b b b b a a a a c b b b b b
a a a a a b b b b b a a a a c b b b b b
at
Pu ch

a a a a a b b b b b a a a a c b b b b b

Corner step edge Vertical ramp edge

Te

a a c b b b b b b b a a a a a a a a a a

a a a c b b b b b b a a a a c c c c c c
a a a a c b b b b b a a a a c b b b b b

a a a a a c b b b b a a a a c b b b b b

a a a a a a c b b b a a a a c b b b b b

Diagonal ramp edge Single pixel transition Corner ramp edge

a a a a c b b b b b a a d e b b b b b b
a a a a c b b b b b a a a d e b b b b b

a a a a c b b b b b a a a a d e b b b b

a a a a c b b b b b a a a a a d e b b b

a a a a c b b b b b a a a a a a d e b b

Vertical ramp edge Diagonal ramp edge Smoothed transition

 DSIP (MU-Sem. 7-COMP) 9-5 Segmentation

a a a a a a a a a a

a a a a d c c c c c

a a a a c b b b b b

a a a a c b b b b b

a a a a c b b b b b

Corner ramp edge

a+b 3a + b a + 3b
c = , d= , e= , ba
2 4 4

Fig. 9.2.5
 If we now want to plot the derivative image, we need
9.3 Detection of Edges
to take the magnitude of Fig. 9.3.1(c). This gives us

ge
How do these edges come in an image ? Fig. 9.3.1(d). As can be seen, we get non-negative
 Variation of scene features, usually brightness, give values only at the two edges of the image. We can
io led
rise to edges. In other words, edges are thus conclude that computing the derivative helps us
representations of the discontinuities of the scene in detecting the edges.
intensity function.  Although this discussions has been limited to the
ic ow

 There could be various reasons such as type of 1-D horizontal profile, a similar argument applies to an
material, surface texture, lighting conditions, etc., edge of any orientation in an image.
n
which play an important role in forming these
 The first derivative at any point in an image is obtained
bl kn

discontinuities.
by using the magnitude of the gradient at that point.
 Our aim in this chapter is to detect these
at

discontinuities as edges. The process of edge detection

Pu ch

9.3.1 Computing the Gradient

is carried out by the derivative approach.
 The basic idea of the derivative approach is the Let us spend some time understanding the basics of
Te

computation of the local derivative operator. the gradient. This discussion will help us understand edge
Consider the image shown in Fig. 9.3.1. detection better.
Case 1 : Output is a function of one variable
(Example : 1-dimensional signal) y = f(x).
We know that the derivative of y w.r.t. x is
dy lim f (x + h) – f(x)
= h0 = f (x)
dx h
dy
The slope is dx

Fig. 9.3.1
 Fig. 9.3.1 shows that the first derivative of the grey line
profile is positive at the leading edge of a transition,
negative at the trailing edge and zero in areas of
Fig. 9.3.2
constant grey levels.
 DSIP (MU-Sem. 7-COMP) 9-6 Segmentation

Case 2 : Output is a function of two variables f lim f (x, y, z + l) – f(x, y, z)

= l0
(2-dimensional example), f(x, y). In the two z l
variable case, x and y are the variables. Both ^ f ^ f ^ f
  gradient F = i +j +k ...(9.3.2)
these variables change to say x + h and y + k. To x y z
find the gradient, we work with each separately Case 3 will not be used in this book
i.e., we take partial derivatives. Now, magnitude of a vector is given by ,
Let us first keep y fixed and alter only x. A
2
= Ax +Ay
2 2

f lim f (x + h, y) – f(x, y) |A| = Ax+Ay

2 2
= h0
x h
Similarly, magnitude of a 2-D gradient is,
f 1/2
f f 2 2
| F | =   +   
Here is the rate of change of f w.r.t. x keeping y
x …(9.3.3)
 x  y 
fixed.
f
Similarly, keeping x fixed and changing y, we get Here = Rate of change of ‘f’ w.r.t the x-direction
x

ge
f lim f (x, y + k) – f(x, y) f
= k0 = Rate of change of ‘f’ w.r.t the y-direction
y k y
io led
f
is the rate of change of f w.r.t. y keeping
y
ic ow

x constant.
Hence the final gradient is
n
^ f ^ f
F = i
bl kn

+j ...(9.3.1)
x y
Fig. 9.3.4
at
Pu ch

9.4 Finding Gradients using Masks

Te

Consider a 3 × 3 neighbourhood with Z5 as the origin.

Corel 2

Fig. 9.3.3
f lim f (x + h, y) – f(x, y)
Case 3 : Output is a function of three variables We know, =
x h0 h
(3-dimensional, e.g.. Density at different points in
f lim f (x, y + k) – f(x, y)
the atmosphere or pressure at different points) =
0
y k k
(Not useful in image processing)
In the discrete domain, h = k = 1
P = f(x, y, z)
f
 = f(x + 1, y) – f (x, y)
  the three partial derivatives are x
f lim f (x + h, y, z) – f(x, y, z) f
= h0 and = f(x, y + 1) – f (x, y)
x h y
f lim f (x, y + k, z) – f(x, y, z)
= k0
y k
 DSIP (MU-Sem. 7-COMP) 9-7 Segmentation

From the 3 × 3 neighbourhood we have, end

f f a3=a1+a2; %% To find the resultant gradient
x
= Z8 – Z5 and
y
= Z6 – Z5 image
1/2 figure(1)
 f  +  f  
2 2
Since, | F | = imshow(uint8(a1)) %% The x-gradient image, you might
 x  y 
need normalization
 | F | = [(Z8 – Z5) + (Z6 – Z5)2]
2 1/2
figure(2)
So as to make it easier for implementation, imshow(uint8(a2)) %% The y-gradient image, you might
need normalization
| F | = |Z8  Z5| + |Z6  Z5|
figure(3)
i.e. | F | = |Z5  Z8| + |Z5  Z6| ...(9.4.1) imshow(uint8(a3)) %% Final image
This is the first order difference gradient. This can be
implemented using two masks.

ge
io led
Corel 3
ic ow

(a) Original image (b) X-gradient image

n
This is known as the Ordinary operator.
bl kn

Steps to compute the gradient of an image are as follows

at

(1) Convolve the original image with mask 1. This gives us

Pu ch

the gradient along the x-direction.

(2) Convolve the original image with mask 2. This gives us
Te

the gradient along the y-direction.

(3) Add the results of 1 and 2.
The advantage of this prolonged method is that, we (c) Y-gradient image (d) Resultant gradient image
can see the effect of each mask separately. Fig. 9.4.1
MATLAB code for ordinary operator There is another direct method of implementing the
%% Ordinary operator %% ordinary operator practically. It gives results that are almost
clear all similar.
clc (1) Add masks 1 and 2
aa = imread(‘test.jpg’);
a = double(aa); Corel 4
[row col] = size(a);
w1=[1 0;-1 0];
w2=[1 -1; 0 0]; (2) Convolve the original image with this resultant mask to
for x=2:1:row-1; get the gradient image.
for y=2:1:col-1;
9.4.1 Roberts Mask
a1(x,y)=w1(1)*a(x,y)+w1(2)*a(x,y+1)+w1(3)*a(x+1,y)
+w1(4)*a(x+1,y+1); Robert L.G in 1965 published a paper in which he
a2(x,y)=w2(1)*a(x,y)+w2(2)*a(x,y+1)+w2(3)*a(x+1,y) stated that better results could be obtained if cross
+w2(4)*a(x+1,y+1); differences were taken instead of the straight difference.
end
 DSIP (MU-Sem. 7-COMP) 9-8 Segmentation

Ordinary masks
| F | = | Z5  Z8 | + | Z5  Z6 |
Roberts masks
| F | = | Z5  Z9 | + | Z6  Z8 | …(9.4.2)
(a) Original image (b) X-gradient image

ge
The results of the Roberts mask are shown along with
the program.
io led
The sum of the two Roberts masks is

1 1
ic ow

–1 –1 (c) Y-gradient image (d) Resultant image

n
MATLAB code for Roberts operator Fig. 9.4.2
bl kn

% Roberts operator %% 9.4.2 Prewitts and Sobel Operators

clear all
at

Roberts mask, as is evident, is an even sized mask

Pu ch

clc
(2  2). Masks of even sizes are awkward to implement.
aa=imread(‘cameraman.jpg’);
Note that the differential along the diagonals of a 2  2
Te

a=double(aa); mask is used and the edge value after the convolution
[row col]=size(a);
corresponds to the central point  r – 2  c – 2  .
1 1
w1= [1 0; 0 -1];  
w2= [0 1; -1, 0]; This problem can be avoided using 3  3 masks.
for x=2:1:row-1;
Prewitts Operator
for y=2:1:col-1;
a1(x,y)=w1(1)*a(x,y)+w1(2)*a(x,y+1)+w1(3)*a(x+1,y) Prewitt J.M.S in his paper “Object Enhancement and
+w1(4)*a(x+1,y+1); Extraction” in 1970 came up with a 3  3 mask. The Prewitt
a2(x,y)=w2(1)*a(x,y)+w2(2)*a(x,y+1)+w2(3)*a(x+1,y) operator, as is now called, while approximating the first
+w2(4)*a(x+1,y+1); derivative, assigns similar weights to all the neighbours of
end the candidate pixel whose edge strength is being
end calculated.
a3=a1+a2; %% Final Gradient Image %% F  | (Z7 + Z8 + Z9)  (Z1 + Z2 + Z3) | + | (Z3 + Z6 + Z9)  (Z1 + Z4 + Z7) |
figure(1) …(9.4.3)
imshow(uint8(a1)) %% The x-gradient image, you might
x-gradient y-gradient
need normalization
figure(2) From this equation, the masks that we obtain are
imshow(uint8(a2)) %% The y-gradient image, you might
need normalization
figure(3)
imshow(uint8(a3)) %% Final image %%
 DSIP (MU-Sem. 7-COMP) 9-9 Segmentation

These masks are known as the Prewitts masks. Sobel mask

–1 –2 –1 –1 0 1 –2 –2 0

0 0 0 + –2 0 2 = –2 0 2

1 2 1 –1 0 1 0 2 2
Fx Fy Fx + Fy

Sobel Operator MATLAB code for Prewitts operator

Duda R.O and Hart P.E in 1973 published a paper %% Prewitts operator %%
“Pattern Classification and Scene Analysis” in which they clear all
used a new operator known as the Sobel Operator. clc
In the Sobel Operator higher weights are assigned to aa=imread(‘test.tif’);
the pixels close to the candidate pixels.
a=double(aa);

ge
F | (Z7 + 2Z8 + Z9)  (Z1 + 2Z2 + Z3)| + | (Z3 + 2Z6 + Z9)  (Z1 + 2Z4 + Z7) |
[row col]=size(a);
w2=[-1 0 1;-1 0 1;-1 0 1];
io led
x-gradient y-gradient
w1=[-1 -1 -1; 0 0 0; 1 1 1];
…(9.4.4)

From this equation, the masks that we obtain are

ic ow

for x=2:1:row-1;
for y=2:1:col-1;
n
a1(x,y)=w1(1)*a(x-1,y-1)+w1(2)*a(x-1,y)
bl kn

Corel 8
+w1(3)*a(x-1,y+1)+w1(4)* …
at

a(x,y-1)+w1(5)*a(x,y)+w1(6)*a(x,y+1)
Pu ch

Implementation of Prewitt and Sobel Operators +w1(7)*a(x+1,y-1)+w1(8)* …

a(x+1,y)+w1(9)*a(x+1,y+1);
Te

(1) Convolve original image with the Fx mask to get the

x-gradient image. a2(x,y)=w2(1)*a(x-1,y-1)+w2(2)*a(x-1,y)
(2) Convolve original image with Fy mask to get the y- +w2(3)*a(x-1,y+1)+w2(4)* …
gradient image. a(x,y-1)+w2(5)*a(x,y)+w2(6)*a(x,y+1)
(3) Add the results of step (1) and (2) +w2(7)*a(x+1,y-1)+w2(8)* …
As mentioned earlier, the advantage of using three a(x+1,y)+w2(9)*a(x+1,y+1);
steps is that we can see the results of the Fx mask and
end
the Fy mask separately.
end
A shorter way of doing this which would give almost
similar results is a3=a1+a2; %% The final gradient value %%
figure(1)
(i) Add the Fx and Fy masks first.
imshow(uint8(a1))%% The x-gradient image,
(ii) Convolve the new mask with the original image.
Normalisation might be required
Prewitt mask
figure(2)
–1 –1 –1 –1 0 1 –2 –1 0 imshow(uint8(a2))%% The y-gradient image %%
0 0 0 + –1 0 1 = –1 0 1 figure(3)

1 1 1 –1 0 1 0 1 2 imshow(uint8(a3))%% Final gradient image,

Normalisation might be required
Fx Fy Fx + Fy
 DSIP (MU-Sem. 7-COMP) 9-10 Segmentation

Normalisation might be required

figure(2)
imshow(uint8(a2)) %% The y-gradient image,
Normalisation might be required
figure(3)
imshow(uint8(a3)) %% Final gradient image,
Normalisation might be required

(a) Original image (b) X-gradient image

ge
io led
(a) Original image (b) X-gradient image
(c) Y-gradient image (d) Resultant gradient image
ic ow

Fig. 9.4.3
n
MATLAB code for Sobel operator
bl kn

%% Sobel operator %%
clear all
at
Pu ch

clc
aa=imread(‘warne.tif’);
a=double(aa);
Te

[row col]=size(a);
(c) Y-gradient image (d) Resultant gradient image
w1=[-1 -2 -1; 0 0 0; 1 2 1 ];
Fig. 9.4.4
w2=[-1 0 1; -2 0 2; -1 0 1 ];
for x=2:1:row-1; Both results can be obtained using absolute values.
for y=2:1:col-1; Have you noticed one thing that is common to all the
a1(x,y)=w1(1)*a(x-1,y-1)+w1(2)*a(x-1,y) edge detection masks that we have studied so far ?
+w1(3)*a(x-1,y+1)+w1(4)* … Let us draw all the masks together.
a(x,y-1)+w1(5)*a(x,y)+w1(6)*a(x,y+1) 1 0 1 –1 1 0 0 1
+w1(7)*a(x+1,y-1)+w1(8)* …
a(x+1,y)+w1(9)*a(x+1,y+1); –1 0 0 0 0 –1 –1 0
a2(x,y)=w2(1)*a(x-1,y-1)+w2(2)*a(x-1,y) Ordinary operator Robert’s operator
+w2(3)*a(x-1,y+1)+w2(4)* …
a(x,y-1)+w2(5)*a(x,y)+w2(6)*a(x,y+1) –1 –1 –1 –1 0 1 –1 –2 –1 –1 0 1
+w2(7)*a(x+1,y-1)+w2(8)*
…a(x+1,y)+w2(9)*a(x+1,y+1); 0 0 0 –1 0 1 0 0 0 –2 0 2

end 1 1 1 –1 0 1 1 2 1 –1 0 1
end
Prewitt mask Sobel mask
a3=a1+a2; %% Final gradient values %%
figure(1)
imshow(uint8(a1)) %% The x-gradient image,
 DSIP (MU-Sem. 7-COMP) 9-11 Segmentation

Sum of the coefficients of each of these masks is  Hence the Prewitt operator performs uniform
zero !! smoothing in one direction with edge detection in the
 This is a very important property to note. Edges are perpendicular direction.
abrupt discontinuities in the gray levels and hence are  In a similar manner, we can split up the Sobel operator
high frequency regions. Since the sum of the into smaller and simpler operators.
coefficients of all these masks is zero, they eliminate
all the low frequency components of the image i.e., i.e –1 –2 –1 –1
when these masks are placed on low frequency 0 0 0 = 0  1 2 1
regions, the output is zero. Hence these masks give
1 2 1 1 LP
edges without any low frequency regions in the final
output image. HP
 Derivative filters (Gradient filters), as the name and –1 0 –1 1
suggests, calculate the gradient of the image. –2 0 2 = 2  –1 0 1
 Since noise is also high frequency, the derivative of the

ge
1 0 1 1 HP
noise terms will always be large values and hence
derivative filters are very sensitive to noise.
io led LP
 There is a huge advantage in using Prewitt and Sobel  One can verify this by comparing the results of a
masks for edge detection. Both these operators Prewitt/ Sobel filter with that of standard highpass
provide a smoothing effect along with providing filter on an image corrupted by Gaussian noise. (Use
ic ow

differentiation. imnoise command to corrupt the image).

 Hence Prewitt and Sobel operators perform well even
n
when the image is noisy because they smoothen the 9.4.3 Compass Operators
bl kn

noise !!
It is seen that edges in the horizontal as well as in the
How do they achieve this ? vertical direction are enhanced when Prewitt’s or Sobel’s
at
Pu ch

Let us consider the Prewitt operator Fx and Fy operator is used. There are applications, in which we need
edges in all the directions. A simple method would be to
–1 –1 –1 –1 0 1 rotate the Prewitts or Sobel’s mask in all the possible
Te

0 0 0 and –1 0 1 directions.
Considering a Prewitts operator
1 1 1 –1 0 1

These operators can be factored into the successive

application of two simpler operators. Corel 9

i.e. –1 –1 –1 –1

0 0 0 = 0  1 1 1
We move this mask in the anticlockwise direction to
1 1 1 1
get all other masks

–1 –1 –1 –1 –1 0 –1 0 1 0 1 1
and –1 0 1 1
0 0 0 –1 0 1 –1 0 1 –1 0 1
–1 0 0 = 1  –1 0 1
1 1 1 0 1 1 –1 0 1 –1 –1 0
–1 0 1 1

In 1 1 1 is a low pass or a –1 0 1 1 1 1 1 1 0 1 0 –1 0 –1 –1
this smoothing operator,
0 0 0 1 0 –1 1 0 –1 1 0 –1
while
–1 –1 –1 0 –1 –1 1 0 –1 1 1 0
is a high pass operator.
 DSIP (MU-Sem. 7-COMP) 9-12 Segmentation

–1 –1 –1 C(x,y)=[c(1)*I(x-1,y-1)+c(2)*I(x-1,y)
0 0 0 +c(3)*I(x-1,y+1)+c(4)*I(x,y-1) …
+c(5)*I(x,y)+c(6)*I(x,y+1)+c(7)*I(x+1,y-1)
1 1 1
+c(8)*I(x+1,y)+c(9)*I(x+1,y+1)];
This operator is known as the compass operator and is
very useful for detecting weak edges. D(x,y)=[d(1)*I(x-1,y-1)+d(2)*I(x-1,y)
Compass operators can also be implemented using the +d(3)*I(x-1,y+1)+d(4)*I(x,y-1)+ …
Sobel operator. d(5)*I(x,y)+d(6)*I(x,y+1)+d(7)*I(x+1,y-1)
%% Program of compass operator %% +d(8)*I(x+1,y)+d(9)*I(x+1,y+1)];
clear all, clc E(x,y)=[e(1)*I(x-1,y-1)+e(2)*I(x-1,y)
aa=imread(‘lily.tif’); +e(3)*I(x-1,y+1)+e(4)*I(x,y-1) …
I=double(aa); +e(5)*I(x,y)+e(6)*I(x,y+1)+e(7)*I(x+1,y-1)

ge
s=size(I); io led +e(8)*I(x+1, y)+e(9)*I(x+1,y+1)];

a=[-1 -1 -1; 0 0 0; 1 1 1]’ F(x,y)=[f(1)*I(x-1,y-1)+f(2)*I(x-1,y)

b=[a(1,2) a(1,3) a(2,3); a(1,1) a(2,2) a(3,3); a(2,1) a(3,1) +f(3)*I(x-1,y+1)+f(4)*I(x,y-1) …

a(3,2)] +f(5)*I(x,y)+f(6)*I(x,y+1)+f(7)*I(x+1,y-1)
ic ow

c=[b(1,2) b(1,3) b(2,3); b(1,1) b(2,2) b(3,3); b(2,1) b(3,1) +f(8)*I(x+1,y)+f(9)*I(x+1,y+1)];

n
b(3,2)] G(x,y)=[g(1)*I(x-1,y-1)+g(2)*I(x-1,y)
bl kn

d=[c(1,2) c(1,3) c(2,3); c(1,1) c(2,2) c(3,3); c(2,1) c(3,1) +g(3)*I(x-1,y+1)+g(4)*I(x,y-1) …

c(3,2)] +g(5)*I(x,y)+g(6)*I(x,y+1)+g(7)*I(x+1,y-1)
at
Pu ch

e=[d(1,2) d(1,3) d(2,3); d(1,1) d(2,2) d(3,3); d(2,1) d(3,1) +g(8)*I(x+1,y)+g(9)*I(x+1, y+1)];
d(3,2)]
Te

H(x,y)=[h(1)*I(x-1,y-1)+h(2)*I(x-1,y)
f=[e(1,2) e(1,3) e(2,3); e(1,1) e(2,2) e(3,3); e(2,1) e(3,1)
+h(3)*I(x-1,y+1)+h(4)*I(x,y-1) …
e(3,2)]
+h(5)*I(x,y)+h(6)*I(x,y+1)+h(7)*I(x+1,y-
g=[f(1,2) f(1,3) f(2,3); f(1,1) f(2,2) f(3,3); f(2,1) f(3,1) 1)+h(8)*I(x+1,y)+h(9)*I(x+1, y+1)];
f(3,2)]
end
h=[g(1,2) g(1,3) g(2,3); g(1,1) g(2,2) g(3,3); g(2,1) g(3,1)
end
g(3,2)]
W=max(max(max(max(max(max(max(A,B),C),D),E),F),G),
for x=2:1:s(1)-1;
H); %% To calculate maximum
for y=2:1:s(2)-1;
%% Plotting %%
A(x,y)=[a(1)*I(x-1,y-1)+a(2)*I(x-1,y)
figure(1),imshow(uint8(A)),figure(2),imshow(uint8(B))
+a(3)*I(x-1,y+1)+a(4)*I(x,y-1)+ …
figure(3),imshow(uint8(C)),figure(4),imshow(uint8(D))
a(5)*I(x,y)+a(6)*I(x,y+1)+a(7)*I(x+1,y-1)
figure(5),imshow(uint8(E)),figure(6),imshow(uint8(F))
+a(8)*I(x+1,y)+a(9)*I(x+1,y+1)];
figure(7),imshow(uint8(G)),figure(8),imshow(uint8(H))
B(x,y)=[b(1)*I(x-1,y-1)+b(2)*I(x-1,y)
figure(9),imshow(uint8(W)),figure(10),imshow(uint8(I))
+b(3)*I(x-1,y+1)+b(4)*I(x,y–1) …
+b(5)*I(x,y)+b(6)*I(x,y+1)+b(7)*I(x+1,y-1) %% Normalisation may be required %%

+ b(8)*I(x+1,y)+b(9)*I(x+1,y+1)];
 DSIP (MU-Sem. 7-COMP) 9-13 Segmentation

ge
io led
ic ow
n
bl kn
at

Resultant image
Pu ch

Fig. 9.4.5 : Compass operated images

 
9.5 Image Segmentation using the  f f
F =
Te

2 +
x y
2
Second Derivative - the Laplacian

 |  F | = [f(x + 1, y) + f(x  1, y) + f(x, y + 1)
 We have already seen the effects of using the first
derivative of the image. It was shown that the first + f(x, y  1)  4f(x, y)] ...(9.5.1)
derivative does enhance the edges of the image. We
now try to find out the effects of computing the Considering the 3  3 neighbourhood
second derivative on the edges.
y–1 y y+1
We know,
x–1 Z1 Z2 Z3
f f
F = + 
x y x Z4 Z5 Z6
f
 = f(x + 1, y)  f(x, y) and x+1 Z7 Z8 Z9
x
f This equation in the discrete form reduces to,
 = f(x, y + 1)  f(x, y)
y
|  F | = [Z8 + Z2 + Z6 + Z4  4Z5]
2

The second derivative is given by



f f
  This equation can be implemented using a mask i.e.
F = 2 +
x y
2


f
Where 2 = f(x + 1, y) + f(x – 1, y) – 2f(x, y) and
x C-11(a)

f
2 = f(x, y + 1) + f(x, y – 1) – 2f(x, y)
y
 DSIP (MU-Sem. 7-COMP) 9-14 Segmentation

 This is known as the Laplacian operator. Some books (1) Since the Laplacian is a second derivative filter, it
reverse the signs of the coefficients of the mask i.e., is very sensitive to noise (much more than the
first derivative). Hence in an image, if there is any
C-11(b) noise present, the Laplacian gives very large
values and ruins the entire image.
(2) The magnitude of the Laplacian produces double
edges, which is an undesireable effect. Let us
 There might be cases when, we would need to add explain what we mean by double edges. Consider
higher weights at the center pixel (>4). While doing a strip of image shown in Fig. 9.5.1. It is seen that
nd
this we must make sure that the sum of the the magnitude of the 2 derivative produces two
coefficients of the mask is zero since edges are high peaks for a single edge.
frequency components. Hence if we increase the The zero crossing property of the Laplacian is
weights at the centre, we would also have to change used to detect edges.
the values of the coefficients at the borders.

ge
9.5.1 Laplacian of Gaussian Operator
 As stated earlier, the Laplacian mask evokes very
io led
strong response to stray noise pixels. Hence if some
kind of noise cleaning is done prior to the application
of the Laplacian operator, better results could be
ic ow

obtained. In practice the Laplacian is preceded by a

smoothing operation (Gaussian smoothing). The
n
resultant algorithm is commonly known as a Laplacian
bl kn

of Gaussian (LoG) or Marr-Hildreth operator. (Marr

D.C. and Hildreth presented a paper “Theory of Edge
at
Pu ch

Detection” in 1980 which explains in this concept)

Let us explain the mathematics of this.
Te

 What we just said was that, we first smoothen the

image using a Gaussian function (Remember low pass
Gaussian filter?) and then take the Laplacian (second
derivative of the image). A simple method to
implement this practically, is to combine the two i.e.
combine the action of a gaussian function and a
Fig. 9.5.1
second derivative function.
 One important thing to note about the Laplacian mask Consider the gaussian function
is that unlike the Sobel and Prewitt operator, they are 2 2
– 
x +y
isotropic filters. i.e., their response is independent of  2
2 
the direction of the discontinuties in the image. In h(x, y) = e …(9.5.2)
other words, isotropic filters are rotation invarient. where  is the standard deviation and x and y are the
i.e., rotating the image and then applying the mask two variables (spatial coordinates). It is  that determines
gives the same result as applying the mask to the the degree of blurring.
image first and then rotating the result. 2 2 2
Let x + y = r
 2
– 
Now that we have the second derivative Laplacian r
mask with us, can we use it directly on the image ? the 22
 h(r) = e
answer is No. The Laplacian is not generally used in its
 The Laplacian of h i.e., the second derivative of h with
original form as an edge detector for mainly two
respect to r is now calculated. We first calculate the
reasons.
first derivative w.r.t. r
 DSIP (MU-Sem. 7-COMP) 9-15 Segmentation
2
  – r2/22 – r2/22  – r  MATLAB code for obtaining Laplacian of a Gaussian image
h(r) = e =e ·
r r r  2 
2
clear all
 – 2/ r
· 2
– r2/22 – r2/22 –r clc
h(r) = e · 2 = e
r  2/     a=imread(‘alumgrns.tif’);
Now taking the second derivative, we get a=double(a);
[row col]=size(a);

2
 – r – r2/22  – 1  – r2/22
r
2 h(r) =
r  
2 e = 2
  r
(re ) lap=[0 1 0;1 -4 1;0 1 0]; %% Laplacian mask
log=[0 0 -1 0 0;0 -1 -2 -1 0;-1 -2 16 -2 -1;0 -1 -2 -1

2
2 h(r) =
–1 re– r2/22 – r2  + e– r2/22  0;0 0 -1 0 0];%% LOG
r 
2
   
for x=3:1:row-2;
2 – r2/22
1 r e 2 2
for y=3:1:col-2;
=  2 – e– r /2 

2
  %% Calculating only the Laplacian %%
2
=
1
e
2
– r /22
 2 – 1
r c(x,y)=[lap(1)*a(x-1,y-1)+lap(2)*a(x-1,y)

2
  +lap(3)*a(x-1,y+1)+lap(4)*a(x,y-1)+…

ge
r –  – r2/22
2 2

2
2 h(r) = e lap(5)*a(x,y)+lap(6)*a(x,y+1)
r 
4
+lap(7)*(x+1,y-1)+lop(8)*a(x+1,y)+...
r – 4  e– r2/22
2 2
 lap(9)*a(x+1,y+1)];
 h =
io led …(9.5.3)
   %% Calculating the Laplacian of a Gaussian %%

 This is the Laplacian of a Gaussian (LoG).  h can be dee(x,y)=log(1)*a(x-2,y-2)+log(2)*a(x-2,y-1)
easily plotted using MATLAB. Due to the shape of the +log(3)*a(x-2,y)+log(4)*a(x-2,y+1)+…
ic ow

LoG, it commonly referred to as the Mexican Hat log(5)*a(x-2,y+2)+log(6)*a(x-1,y-2)

function. +log(7)*a(x-1,y-1)+log(8)*a(x-1,y)+...
n
log(9)*a(x-1,y+1)+log(10)*a(x-1,y+2)
bl kn

+log(11)*a(x,y 2)+log(12)*a(x,y-1)+...
log(13)*a(x,y)+log(14)*a(x,y+1)+log(15)*a(x,y+2)
at
Pu ch

+log(16)*a(x+1,y-2)+...
log(17)*a(x+1,y-1)+log(18)*a(x+1,y)
Te

+log(19)*a(x+1, y+1)+log(20)*a(x+1,y+2)+…
log(21)*a(x+2,y-2)+log(22)*a(x+2,y-
1)+log(23)*a(x+2,y)+log(24)*a(x+2,y+1)+...
log(25)*a(x+2,y+2);
Fig. 9.5.2
end

 It can be seen that  h has a large positive value at the end
centre and negative values at the peripheries. %% Thresholding the output of the LoG operator %
We can approximate this by using a 5  5 mask as for x=1:1:row-2
shown. for y=1:1:col–2
if dee(x,y)<40
 Remember, this mask is not unique. We could use any
deep(x,y)=0;
values that approximate the shape of the LoG.
else
 One thing that needs to be kept in mind is that since
deep(x,y)=255
our intention is to enhance the edges, the sum of
end
coefficients of the mask should be zero.
end
0 0 –1 0 0 end
%%%%%%%%%%%%%%%%%%%%%%%%%
0 –1 –2 –1 0
%%%%%%%%ZERO CROSSINGS%%%%%%%%%%
–1 –2 16 –2 –1 for x=2:1:row-2
0 –1 –2 –1 0 for y=2:1:col-2
a1=deep(x,y);
0 0 –1 0 0
a2=deep(x-1,y-1);
DSIP (MU-Sem. 7-COMP) 9-16 Segmentation

if a1==0 && a2==0 9.5.2 Canny Edge Detector

deepa(x,y)=0;
deepa(x-1,y-1)=0;  This is the last of the edge detectors that we would
else if a1==255 && a2==255 study here. Canny operator (Named after J. K. Canny)
deepa(x,y)=0; is a first derivative edge detector coupled with noise
deepa(x-1,y-1)=0; cleaning. Canny has, over the years, become one of
the most popular derivative operators. Like in the LoG,
else
a Gaussian function is used to smoothen the noise.
deepa(x,y)=255;
 In the Canny edge detector, we first smoothen the
deepa(x-1,y-1)=255;
image using a Gaussian low pass filter and then take
end
the first derivative.
end
end Consider the Fig. 9.5.4,
end
end

ge
figure(1),imshow(uint8(a)) %% Original image
figure(2),imshow(uint8(c)) %% Only Laplacian
figure(3),imshow(uint8(dee)) %% LoG
io led
figure(4),imshow(uint8(deep)) %% LoG+Threshold
figure(5),imshow(uint8(deepa)) %% Zero crossings (a) A ramp edge, (b) First derivative of ramp edge, (c)
Second derivative of ramp edge
ic ow
n
bl kn
at
Pu ch

(d) Gaussian function, (e) First derivative of

Gaussian function
Te

(a) Original image Fig. 9.5.4

Compare Fig. 9.5.4(c) and Fig. 9.5.4(e). Notice the
similarity between the shape of the first derivative of the
Gaussian and the second derivative of the ramp edge.
Hence the derivative of the bell shape of the Gaussian
function approximates the second derivative or zero
crossing operator (LoG).

9.6 Connectivity

(b) Image using LoG operator MU : Dec. 2016, Dec. 2017

Q. Write short note on : 4, 8 and m-connectivity.

(Dec. 2016, Dec. 2017, 5 Marks)

 Consider a pixel p(x, y). This pixel p has two horizontal

and two vertical neighbours which share a surface with
p(x, y).
 This set of four pixels is known as the 4-
neighbourhood of pixel p(x, y), where each pixel is said
to be at a unit distance from p(x, y). This set of pixels is
(c) Zero crossing image denoted as N4(p). In a similar manner, apart from
Fig. 9.5.3 these 4-neighbours, there are another four diagonal
pixels which touch p(x, y) at the corners ND(p).
 DSIP (MU-Sem. 7-COMP) 9-17 Segmentation

p(x – 1, y)

p(x, y – 1) p(x, y) p(x, y + 1)

p(x + 1, y)
N4(p)
 This set of eight pixels (The 4-neighbours as well as the Corel 16
diagonal pixels) is called the 8-neighbourhood of pixel
p(x, y) and is denoted as N8(p)
 N8 (p)= N4(p) + ND(p).
Corel 17
 Connectivity is an important concept that is used in
segmentation. When we started off with this chapter,
it was stated that segmentation is used for machine  Let us take an example. Given the criteria that p = q = 1
vision.

ge
for the region to be connected, check whether the two
p(x –1, y – 1) p(x –1, y + 1) region S1 and S2 are connected.
 Even though p and q are both equal to one, these two
io led
p(x, y)
regions will not be shown connected, if we use a 4-
p(x + 1, y – 1) p(x + 1, y + 1) connectivity algorithm. This is because, in the 4-
connectivity algorithm, we only check the 4-
ic ow

ND(p)
neighbours and see if any of these four neighbours has
 In segmentation, the machine (computer) scans the
a pixel with the same value 1, in this case
n
image to form different regions. This scanning of the

bl kn

If we use an 8-connectivity algorithm, S1 and S2 would

image by the machine is based on the connectivity
that the user or the code specifies. To establish if two be shown as a single connected region. Apart from 4-
at

connectivity and 8-connectivity, we also have m-

Pu ch

pixels are connected, it must be determined whether

they are neighbours (4-neighbours or 8-neighbours) adjacent or m-connectivity (mixed connectivity). It is
and if their grey levels satisfy a specified criteria. used to eliminate the double paths that arise when we
Te

use 8-connectivity.
 Two regions that touch only at a corner can be
considered to be a single region or two distinct Let us take an example.
regions : How they are considered depends on the Consider the image shown.
definition of connectivity used ?
 Two pixels p and q with some common criteria are said
Corel 18
to be 4-connected if they share a side. i.e., two pixels p
and q with some common criteria are 4-connected if q
is in the set N4(p).
Here the condition used is that pixels belong to a
region if their value is 2.
When we use 8-connectivity here, we get

Corel 19

We notice that the pixel at the right hand upper corner

 Two pixels p and q with some common criteria are said gets connected twice due to 8-connectivity. The ambiguity
to be 8-connected if they share either a side or a is removed by using m-connectivity.
corner. i.e., two pixels p and q with some common
criteria are 8-connected if q is in the set N8(p).
 DSIP (MU-Sem. 7-COMP) 9-18 Segmentation

Hence m-connectivity can be defined as follows. 9.7 Distance Transform

Two pixels p and q are said to be m-connected if
Though not directly related to segmentation, Distance
(i) q is in N4(p) or
transform is introduced here as it is related to connectivity.
(ii) q is in ND(p) and the set N4(p) ⋂ N4(q) is empty Distance transforms are required in image processing
Using this definition we get projects where measurements have to be made. The
distance transform provides a measure of the separation of
points in an image.
Corel 20 (1) Euclidean Distance : Euclidean distance is the straight
line distance between two pixels. If p and q are the
two pixels with coordinates (x1, y1) and (x2, y2), then
Any connectivity can be used as long as one is 2 2 1/2
DE = [(x1 – x2) + (y1 – y2) ]
consistent. Often 8-connectivity yields results that lie closer
to one’s intuition. Diagrammatically it can be shown as

ge
9.6.1 Solved Example on Connectivity
Ex. 9.6.1 : Consider two image subsets S1 and S2
io led Corel 22

S1 S2

0 0 0 0 0 0 0 0 2 2 
2 0 0 2 0 0 2 0 0 2  for a 3  3 region, we have the Euclidian distance as
ic ow

2 0 0 2 0 2 2 0 0 0 (2) City Block Distance (D4 Distance) : For the same points
n
0 0 2 2 2 0 0 0 0 0 p(x1, y1) and q(x2, y2), the city block distance is defined
bl kn

as
0 0 2 2 2 0 0 2 2 2 DCITY (p, q) = D4(p, q) = |x1  x2| + |y1  y2|
at
Pu ch

For V = {2}, determine whether S1 and S2 are The city block distance measures the path between
(a) 4-connected (b) 8-connected (c) m-connected the pixels based on a 4-connected neighbourhood.
Te

Soln. : Pixels whose edges touch are 1 unit apart and pixels
touching diagonally are 2 units apart.
V = {2} simply means that two pixels p and q are
connected if their values are equal to 2.
S1 S2
Corel 23
0 0 0 0 0 0 0 2 2 0

2 0 0 2 0 0 2 0 0 2
(3) Chess Board Distance (D8 Distance) : For pixels p(x1,
2 0 0 2 0 2 q 2 0 0 0 y1) and q(x2, y2), the chess board distance is defined as
0 0 2 2 2 p 0 0 0 0 0 DCHESS, (p, q) = D8 (p, q) = max(|x1  x2|, |y1  y2|)
The chessboard distance measures the path between
0 0 2 2 2 0 0 2 2 2
the pixels based on a 8-connected neighbourhood.
(a) S1 and S2 are not 4-connected because q is not in the Pixels whose edges or corners touch are 1 unit apart.
set of N4(p).
2 2 2 2 2
(b) S1 and S2 are eight connected because q is in the set of
N8(p). 2 1 1 1 2

(c) S1 and S2 are m-connected because 2 1 0 1 2

(i) q is in the set ND(p) and 2 1 1 1 2
(ii) the set N4(p) ⋂ N4(q) is empty
2 2 2 2 2
 DSIP (MU-Sem. 7-COMP) 9-19 Segmentation

(4) Dm Distance (m-adjacency Distance) : This distance (b) D8(p, q) = max (|x1  x2|, |y1  y2|)
measure is based on m-adjacency. Pixels p and q are = max (|4  0|, |0  4|)
m-adjacent if
= max (4, 4) = 4
(a) q is in N4(p) or
(b) q is in ND(p) and N4(p) ⋂ N4(q) is empty. Ex. 9.7.3 : Consider the image segment shown

3 1 2 1q
9.7.1 Solved Examples on Distance for V = {0, 1}, compute the
2 2 0 2
Transform length of the shortest
1 2 1 1 m-path between p and q.
Ex. 9.7.1 : Measure the Dm distance between p and q for the
1p 0 1 2
following images.
Soln. : The path taken is shown below :
Soln. :

ge
io led
Hence the shortest m-distance is equal to 5.
ic ow

9.8 Additional Solved Examples

Corel 24
n
Ex. 9.8.1 : Develop an algorithm for converting a one-pixel
bl kn

thick 8-path to a 4-path.

Soln. : This can be done by defining all possible
at
Pu ch

neighbourhood shapes to go from a diagonal segment to a

corresponding 4-connected segment, as shown in
Fig. P. 9.8.1.
Te

Ex. 9.7.2 : Given below is a 5  5 image. Find out D4, D8.

Soln. :

Corel 25

p(x1, y1) = p(4, 0)

q(x2, y2) = q(0, 4)
(a) D4(p, q) = |x1  x2| + |y1  y2|
= |4  0| + |0  4|
Fig. P. 9.8.1
= 4+4=8
The algorithm then simply looks for the appropriate
match everytime a diagonal segment is encountered.
 DSIP (MU-Sem. 7-COMP) 9-20 Segmentation

Ex. 9.8.2 : A binary image contains straight lines oriented  We get

2
horizontally, vertically, at 45 and at 45. Give a set of 3 H 3   r  
r –   e – 22 dr =
2
2  2
masks that can be used to detect 1 pixel long breaks in     
  –

 =0
these lines. Assume that the grey level of the lines is 1 and –

that the grey level of the background is 0.

Ex. 9.8.4 : Show that the Laplacian mask is a Isotropic filter.
Soln. : (Hint : use a 3  3 mask)
The 4 masks that we could use are Soln. :
By Isotropic filters we mean, filters that are rotation
invariant. It means that rotating the image and then
Corel 30 applying the filter gives the same result as applying the
filter to the image first and then rotating the result.
A 3  3 Laplacian mask is given
Note that the sum of the coefficients along the

ge
0 –1 0
required direction is zero.
Hence each mask would yield a value of 0 when –1 4 –1
io led
centered on a pixel of an unbroken 3-pixel segment 0 –1 0
oriented in the direction of the required mask. The
response of the mask would be +2 when the mask is
ic ow

centered on a one pixel gap in a 3-pixel segment oriented in 1 2 1

the direction of the required mask. Consider a 3  3 image 2 4 2
n
bl kn

Ex. 9.8.3 : Show that the average value of the Laplacian 4 8 4


operator  h is zero.
Case 1: Image Convolve with Laplacian mask Rotate
at
Pu ch

Soln. :
Applying the mask we get,
We have seen that
2
Te

 r
r –  
2
 h (r) = + 
– 2
 2
e
   Corel 31
What we need to prove is
2
 2  r
–  
 
r 2 –
 e 2 dr = 0
  Case 2 : Image Rotate Convolve with Laplacian mask
–
2 2 Rotating image we get,
 r  r
1 – –
2 dr – 1
2 2


  2
r e

  e 2 dr = 0
– –

We know from the definition of the Guassian density Corel 32

that
2
 r
1 – 2
  e 2 dr = 1
We see that (1) = (2)
2 – 
Hence we conclude that the Laplacian mask is an
We also know that the variance of a Guassian random
Isotropic filter (Invariant to rotation).
variable is
2
 r Note : Convolution was performed after zero padding the
– 2
Variance (r) =  =

 2
r e 2 dr image.
–
 DSIP (MU-Sem. 7-COMP) 9-21 Segmentation

Ex. 9.8.5 : Show that the Laplacian operation defined by Soln. :

 
 f f Let f(x, y) be the original image.
f = 2 + 2 is istotropic
x y
The Laplacian mask is given by
Soln. :
0 1 0
We use the following equations relating coordinates
after axis rotation by angle . 1 –4 1
x = x cos   y sin  0 1 0
y = x sin  + y cos  
 f(x, y)   f(x, y)= f(x, y)  [f(x + 1, y) + f(x  1, y)
where (x, y) are the unrotated and (x, y) are the + f(x, y + 1) + f(x, y  1)  4f(x, y)]
rotated coordinates.
= 6 f(x, y)  [f(x + 1, y) + f(x  1, y) + f(x, y + 1)
Laplacian operation for the unrotated coordinates are,
  + f(x, y  1) + f(x, y)]
f f
f
[ ]
2
= 2 + 1
x y
2
= 5 1.2 f(x, y)  5 [f(x + 1, y) + f(x –1, y) + f(x, y + 1) + f(x, y 1) + f(x, y)]

ge
and the Laplacian operation for the rotated –

coordinates are, f(x, y)   f(x, y) = 5 {1.2 f(x, y) – f (x, y)}
  –
f f
f
2
= 2 +
io led Here f (x, y) denotes the average of f(x, y) .
x y
2

Treating the constants as factors of proportionality we

Let us start with
get.
f f x f y f f
cos  + sin 
ic ow

= · + · = –
x x x y x x y 
f(x, y)   f(x, y)  f(x, y) – f (x, y)
We take the partial derivative of this expression again
n
The RHS is the equation for unsharp masking. Hence
with respect to x
bl kn

 
subtracting the Laplacian from an image is equivalent to
f f   f  unsharp masking.
2 cos  +
2
2 = sin  cos 
x x x y
at
Pu ch


  f  f 2 Ex. 9.8.7 : Segment the following image into two regions
+ cos  sin  + 2 sin  …(1)
y x y based on the edge-orientations derived through a gradient
Te

f operator clearly depicting the boundary.

We now find
y
27 27 27 27 27
f f x f y
= · + ·
y x y y y 25 25 25 25 25
f f
= – sin  + cos  23 23 30 33 36
x y
Taking the partial derivatives of this expression with 21 21 33 36 39
respect to y gives us
  19 19 36 39 42
f f   f 
2 sin  –
2
= cos  sin 
y x y Soln. : The easiest way to segment this image into two
2
x
regions would be use a thresholding function with T  28.
  f  f
2
sin  cos  + 2 cos 
2
– But since we have been asked to do this using gradient
y x y
operators, let us use a Prewitts compass operator.
Adding the two equations we get We use the following Prewitts masks and convolve
   
f f f f each one of them with the image.
2+ 2 = 2 +
x y x y
2
We finally calculate the maximum.
This proves that the Laplacian operation is –1 –1 –1 –1 –1 0
independent of rotation.
M1 = 0 0 0 M2 = –1 0 1
Ex. 9.8.6 : Show that subtracting the Laplacian from an 1 1 1 0 1 1
image is proportional to unsharp masking.
DSIP (MU-Sem. 7-COMP) 9-22 Segmentation

–1 0 1 0 1 1 0 0 0 0
M3 = –1 0 1 M4 = –1 0 1 0 5 –5 – 18
–1 0 1 –1 –1 0 E = f  M5 =
0 0 – 15 – 33
0 2 –8 – 18
1 1 1 1 1 0
M5 = 0 0 0 M6 = 1 0 –1 0 0 0 0

–1 –1 –1 0 –1 –1 0 8 1 –9
F = f  M6 =
0 15 6 – 13

1 0 –1 0 –1 –1 0 27 23 0
M7 = 1 0 –1 M8 = 1 0 –1

ge
1 0 –1 1 1 0 0 0 0 0
0 7 10 6
io led
Let the original image be f (x, y).
G = f  M7 =
While convolving we ignore the corner pixels. 0 19 25 12
0 36 45 18
0 0 0 0
ic ow

0 –5 5 18
0 0 0 0
n
A = f  M1 =
0 0 15 33
bl kn

0 –1 9 15
0 –2 8 18 H = f  M8 =
at

0 11 29 31
Pu ch

0 0 0 0 0 21 37 24
Te

0 –8 –1 9 We now take the maximum from each of the pixel

B = f  M2 =
0 – 15 –6 13 locations.
Hence we have
0 – 27 – 23 0
0 0 0 0
0 0 0 0 0 8 10 18
0 –7 – 10 –6 W=
0 19 29 33
C = f  M3 =
0 – 19 – 25 – 12
0 36 45 24
0 – 36 – 45 – 18
We zero pad one row and one column to get the final
image.
0 0 0 0
0 0 0 0 0
D = f  M4 = 0 1 –9 – 15
0 0 0 0 0
0 – 11 – 29 – 31 Final =
0 0 8 10 18
0 – 21 – 37 – 24
0 0 19 29 33

0 0 36 45 24
 DSIP (MU-Sem. 7-COMP) 9-23 Segmentation

Hence we have two regions, one being the zero region Q. 3 Explain edge detection in detail.
(say a) and the other being the non-zero region (say b).
Q. 4 Show that the Laplacian operator is invariant to
a a a a a rotation.

a a a a a Q. 5 Explain the need of a LoG operator.

a a b b b Q. 6 Explain the following edge extraction operators :

a a b b b (a) Sobel (b) Prewitt

(c) Roberts (d) Laplacian
a a b b b
Use these masks on an image of your choice.
Summary
Q. 7 A binary image contains straight lines oriented
horizontally, vertically, at + 45° and at 45°. Give a
Segmentation is an essential step in almost all image

ge
analysis systems. It is imperative to understand the set of 3 H 3 masks that can be used to detect 1-pixel

applications of image segmentation. Segmentation long breaks in these lines. Assume that the grey level
partitions an image into meaningful regions having certain of the lines is 1 and that of the background is 0.
io led
characteristics unique to that region. In this, the output is
Q. 8 Explain the concept of compass operators.
an abstract representation of the input image. There exists
no general segmentation algorithm which will work Q. 9 Explain whether poorly illuminated image can be
ic ow

satisfactorily for all images. Users have to choose a easily segmented.

particular segmentation algorithm which would be suitable
n
Q. 10 Write in detail on connectivity of pixels.
bl kn

for the problem in hand. Segmentation based on

dissimilarities as well as segmentation based on similarities Q. 11 What is the distance transform ?
have been discussed in detail. MATLAB codes have been
at
Pu ch

provided for gradient operators such as Prewitts and Sobel. Q. 12 Explain :

Special emphasis has been given to Hough transforms (a) Euclidean distance
Te

which forms an important topic of edge linking. (b) City block distance
(c) Chess board distance
Review Questions
(d) Dm distance.

Q. 1 Explain the term Image Segmentation.

Q. 2 Explain, segmentation based on discontinuties and
segmentation based on similarities.


 Te
Pu ch
bl kn
ic ow
Note

at
io led
n ge