EE445S Real-Time Digital Signal Processing Lab Fall 2020

Finite Impulse Response Filters

Prof. Brian L. Evans

Dept. of Electrical and Computer Engineering
The University of Texas at Austin

Lecture 5 http://www.ece.utexas.edu/~bevans/courses/realtime
 Outline
• Many Roles for Filters
• Convolution
• Z-transforms
• Linear time-invariant systems
Transfer functions
Frequency responses
• Finite impulse response (FIR) filters
Filter design with demonstration
Cascading FIR filters demonstration
Linear phase
5-2
 Many Roles for Filters
• Noise removal Bandpass Spectrum In Noise

Bandpass filtering to suppress

out-of-band noise (slide 1-5) w
- wc wc
• Spectral analysis x0[m]
x[m] LPF
Two-channel filter bank on right
x1[m]
Used in interpolation (slide 13-9) HPF

• Spectral shaping
x[m] System r[m]
Data conversion (lectures 10/11) Model Equalizer
Channel equalization (slide 16-7) (Filter)
Goal is to make
Carrier frequency/phase recovery cascade all-pass
Timing recovery (slide 13-17 & 16-10)
5-3
 Spectral Analysis
• Analysis: decompose signal into frequency bands
Graphic equalizer: apply separate gain to each band
Compression: spend more bits on more important
• Synthesis: combine bands into one signal
• Example: Two-band filter bank / wavelet
LPF – averaging: x0[m] = ½ x[m] + ½ x[m-1]
HPF – first-order difference: x1[m] = ½ x[m] - ½ x[m-1]
Exact reconstruction of x[m] is possible through synthesis
x0[m] x0[m] x[m]
x[m] LPF +
x1[m] x1[m]
HPF
5-4
Analysis Synthesis
Finite Impulse Response (FIR) Filter
• Same as discrete-time tapped delay line (slide 3-15)
x[n-1]
x[n] z-1 z-1 … z-1

h[0] h[1] h[2] … h[M-1]

S y[n]

• Impulse response h[n] has finite extent n = 0,…, M-1

M-1
Discrete-time
y[n] =å h[m] x[n - m] 5-5
convolution
m=0
 Review

Discrete-time Convolution Derivation

• Output y[n] for input x[n] y n  T  x n 

• Any signal can be decomposed   

y n  T   x m   n  m 
into sum of discrete impulses m   

• Apply linearity properties y n    x m T   n  m
m  

• Apply shift-invariance y n    x m h n  m
m  

• Apply change of variables y n    h m x n  m
m  
h[n] Averaging filter
impulse response y[n] = h[0] x[n] + h[1] x[n-1]
1
2
= ( x[n] + x[n-1] ) / 2
n 5-6
0 1 2 3
 Review

Convolution Comparison
• Continuous-time convolution of x(t) and h(t)
 
y  t   x t   h t    x   h t    d   h   x t    d
 

For each t, compute different (possibly) infinite integral

• In discrete-time, replace integral with summation
¥ ¥
y [ n ] = x[n]* h[n] = å x [ m ] h [ n - m ] = å h [ m ] x [ n - m ]
m=- ¥ m=- ¥
For each n, compute different (possibly) infinite summation
• LTI system
Characterized uniquely by its impulse response
Its output is convolution of the input and its impulse response
5-7
 Convolution Demos
• Signal Processing First Matlab Demonstrations
Continuous-Time http://dspfirst.gatech.edu/matlab/#cconvdemo
Discrete-Time http://dspfirst.gatech.edu/matlab/#dconvdemo
• Convolving two signals of finite lengths
Continuous-Time y(t) = x(t) * h(t) Ly = Lx + Lh
Discrete-Time y[n] = x[n] * h[n] Ly = Lx + Lh − 1
• Convolving two causal signals?
x(t) h(t) y(t)
* =
1 1 Ts

t t t
0 Ts 0 Ts Ts 2Ts

What about convolving two pulses of different lengths? See Appendix E

 Review

Z-transform Definition
• For discrete-time systems, z-transforms play same
role as Laplace transforms do in continuous-time
Bilateral Forward z-transform Bilateral Inverse z-transform

1
H ( z)   h n z  n h[n]   H ( z ) z  n 1
dz
n   2 j R

Inverse transform requires contour integration over closed

contour (region) R
Contour integration covered in a Complex Analysis course
• Compute forward and inverse transforms using
transform pairs and properties
5-9
 Review

Three Common Z-transform Pairs

• h[n] = d[n] • h[n] = an u[n]

H  z    a u n z
 0
H  z     
n
 
n
 n z n
  n z n
1
n   n 0 n  

  n
Region of convergence: entire a
  a n z n    
n0  z 
z-plane n 0

1 a
• h[n] = d[n-1] 
a
if
z
1
1
 1 z
H  z     n  1 z
n  
n
    n  1 z  n  z 1
n 1 Region of convergence for
Region of convergence: entire summation: |z| > |a|
z-plane except z = 0 |z| > |a| is the complement
h[n-1]  z-1 H(z) of a disk (see next slide)

Finite extent sequences 5 - 10

Infinite extent sequence
 Review

Region of Convergence
• Region of the complex z- • Four possibilities (z = 0 is
plane for which forward z- special case that may or
transform converges may not be included)
Im{z} Im{z}

Entire Disk
plane Re{z} Re{z}

Im{z} Im{z}

Intersection
Complement of a disk and
of a disk Re{z} complement Re{z}

of a disk
5 - 11
 Review

System Transfer Function

• Z-transform of system’s impulse response
Impulse response uniquely represents an LTI system
• Example: FIR filter with M taps (slide 5-4)
 M 1
H ( z )   h n z   h n z  n  h[0]  h[1] z 1  ...  h[ M  1] z ( M 1)
n

n   n 0

Transfer function H(z) is polynomial in powers of z -1

Region of convergence (ROC) is entire z-plane except z = 0
• Since ROC includes unit circle, substitute z = e j w
into transfer function to obtain frequency response
M 1
H (e )  H ( z ) | z e   h n e  j  n  h[0]  h[1] e  j   ...  h[ M  1] e  j ( M 1) 
j
j

n 0
5 - 12
 Example: Ideal Delay
• Continuous Time • Discrete Time
Delay by T seconds Delay by 1 sample
x(t) y(t) x[n] y[n]
T z 1

y  t   x(t  T ) y[n]  x[n  1]

Impulse response Impulse response
h t    (t  T ) h[n]   [n  1]
Frequency response Frequency response
H    e j  T H    e j 

| H   |  1 | H   |  1

H       T H       5 - 13
 Linear Time-Invariant Systems
• Fundamental Theorem of Linear Systems
If a complex sinusoid were input into an LTI system, then
output would be input scaled by frequency response of LTI
system (evaluated at complex sinusoidal frequency)
Scaling may attenuate complex sinusoid and shift it in phase
(complex sinusoids are eigenfunctions of LTI systems)
Continuous-time example: see handout F (lowpass RC filter)
Discrete-time derivation: Input x[n] = e j w n into LTI system
¥ ¥
y[n] = å e ( )
å
j w n- m jw n
h[m] e - j w m =e
jw n
h[m] =e H (w )
m=- ¥ m=- ¥

x[n] * h[n] H(w)

H(w) is frequency response (DT Fourier transform of h[n])
 Frequency Response
• Continuous-time e jt
H  j  e j  t
h(t )
LTI system cos  t  H  j  cos  t  H  j  
• Discrete-time e j n H e j   e j  n
h[n]
LTI system
cos  n  H  e j   cos n  H  e j   
• For real-valued impulse response H(e -j ω) = H*(e j ω)
Input e  j  n  e j  n  2 cos( n)
Output H  e  j   e  j  n  H  e j   e j  n  H *  e j   e  j  n  H  e j   e j  n 
He ej
e  H e  e
 jH ( e j )  j n j jH ( e j )
e j n 
2 H  e  cos  ω n  H  e  
j j
5 - 15
 Frequency Response
• System response to complex sinusoid e j w n for all
possible frequencies w in radians per sample:
|H(w)| |H(w)|
stopband
Linear
stopband
phase

w w
-wstop -wp wp wstop
d
passband delay ( )   H ( )  kdelay
d
Lowpass filter: passes low and attenuates high frequencies
Linear phase / constant group delay: FIR filter whose impulse
response is symmetric or anti-symmetric about its midpoint
• Not all FIR filters exhibit linear phase 5 - 16
 FIR Filter Design
• Specify desired piecewise Lowpass Specification
constant magnitude resp. dB
forbidden
Apass
• Lowpass filter example
0
w  [0, wpass], magdB  [-Apass, Apass] -Apass forbidden
w  [wstop, p], magdB ≤ Astop
Achtung!
Transition band unspecified but forbidden
should not amplify Astop
• Symmetric (linear phase) w
wpass wstop p
FIR filter design methods Passband Transition Stopband
Windowing band
Least squares Apass passband tolerance (dB)
Remez (Parks-McClellan) Astop stopband attenuation (dB)
AdB = 20 log10(A) 5 - 17
 Example: Two-Tap Averaging Filter
• Input-output relationship h[n] Two-tap averaging filter
1 1
y[ n]  x[n]  x[n  1]
2 2 ½
• Impulse response
1 1
n
h[n]   [n]   [n  1] 1 2 3
2 2
• Frequency response x[n-1]
1 1
H     e j  x[n] z-1
2 2
1  j 2   j 2 j  
1 1 1
H    e e
  e 2 

h[0] h[1]
2  
    j 2
1
H     cos  e S y[n]
linear phase
2
1
Ð H (w ) =- w 5 - 18
magnitude 2
 Example: First-Order Difference
• Input-output relationship h[n] First-order difference
1 1
y[ n]  x[n]  x[ n  1]
2 2 ½
• Impulse response
1 1
n
h[n]   [n]   [n  1] 2 3
2 2 -½
• Frequency response
1 1
H     e j 
2 2 generalized
p 1
1  j 2   j 2 j   Ð H (w ) = - w
1 1 1
H    e
linear phase
e  e 2  2 2
2  
 
p 1 æ ö
æw ö - j 1w æw ö j
p
- j
1
w æw ö jç - w ÷
H w =sin ç ÷j e 2 =sin ç ÷e 2 e 2 =sin ç ÷ e è 2 2 ø
( ) è2 ø è2 ø è2 ø
5 - 19
amplitude
 Complementary Frequency Responses

Magnitude
response

Phase
response

Averaging filter (lowpass) in white

First-order difference filter (highpass) in yellow/blue
First-order difference has p phase shift at zero mag response frequency
 Cascading FIR Filters Demo
• Five-tap discrete-time averaging FIR filter with
input x[n] and output y[n]
y n  x[n]  x[n  1]  x[n  2]  x[n  3]  x[n  4]
Standard averaging filtering scaled by 5
Lowpass filter (smooth/blur input signal)
Impulse response of {1, 1, 1, 1, 1} removes multiplications
• First-order difference FIR filter
y n  x[n]  x[n  1] h[n] First-order difference
impulse response
Highpass filter (sharpens
input signal) 1
Impulse response is {1, -1} n
removes multiplications 2 3
-1 5 - 21
 Cascading FIR Filters Demo

Apply 1-D five-tap filter across each row to produce 256 x 260 image
and then apply the same 1-D five-tap filter down each column of 256
x 260 image to produce a 260 x 260 image
 Cascading FIR Filters Demo
• DSP First, 2nd ed, ch. 6, cascading FIR Filters
http://dspfirst.gatech.edu/chapters/06firfreq/demos/blockd/index.html
From lowpass filter to highpass filter:
original image  blurred image  sharpened/blurred image
From highpass to lowpass filter:
original image  sharpened image  blurred/sharpened image
• Frequencies that are zeroed out can never be
recovered (e.g. DC is zeroed out by highpass filter)
• Order of two LTI systems in cascade can be switched
under the assumption that computations are
performed in exact precision

5 - 23
 Cascading FIR Filters Demo
• Input image is 256 x 256 matrix
Each pixel represented by eight-bit number in [0, 255]
0 is black and 255 is white for monitor display
• Each filter applied along row then column
Averaging filter adds five numbers to create output pixel
which requires 3 extra bits (worst case) for each pass
Difference filter subtracts two numbers to create output pixel
which requires 1 extra bit (worst case) for each pass
• Full output precision 16 bits/pixel is maintained
Demonstration uses double-precision floating-point data and
arithmetic (53 bits of mantissa + sign; 11 bits for exponent)
5 - 24
 Cascading FIR Filters Demo
• How to compare images pixel by pixel?
Input image is 256 x 256 and output image is 261 x 261
• Unit sample delay example
If observing output y[n], how to x[n] 1
y[n]
z
recover input x[n] exactly?
x[n] y[n]
3 3
2 2
… …
1 1
n n
0 1 2 0 1 2 3
• FIR filters in demo have constant group delay
Remove first 2 rows/columns and last 3 rows/columns 5 - 25
 Importance of Linear Phase
• Speech signals • Linear phase crucial
Use phase differences in Audio
arrival to locate speaker Images
Once speaker is located, ears Communication systems
are relatively insensitive to •
Linear phase response
phase distortion in speech
from that speaker Need FIR filters
Used in speech compression Realizable IIR filters
(in cell phones) cannot achieve linear
phase response over all
d = c t
frequencies

5 - 26
 Importance of Linear Phase
code
• For images, vital visual information in phase

Take FFT of image Take FFT of image Take FFT of image

Set phase to zero Set magnitude to one Set magnitude to one
Take inverse FFT Take inverse FFT Take inverse FFT
(which is real-valued) Keep imaginary part Keep real part
• Original image is from Matlab
http://users.ece.utexas.edu/~bevans/courses/realtime/lectures/05_FIR_Filters/Original%20Image.tif 5 - 27
 Finite Impulse Response Filters
• Duration of impulse response h[n] is finite, i.e.
zero-valued for n outside interval [0, M-1]:
 M 1
y n   x n  h n   h m x n  m   h m x n  m
m   m 0
Output depends on current input and previous M-1 inputs
Summation to compute y[k] reduces to a vector dot product
between M input samples in the vector
 x[n], x[n  1], ..., x[n  ( M  1)] 
and M values of the impulse response in vector
 h[0], h[1], ..., h[M  1] 
• What instruction set architecture features would you
add to accelerate FIR filtering?
5 - 28
 Takeaways
• Linear time-invariant systems
Uniquely defined by its impulse response
Do not create new frequencies
System must be initially “at rest” (zero initial conditions)
• Some FIR filters have linear phase
FIR coefficients (impulse response) must be either even
symmetric or odd symmetric about its midpoint
Constant group delay in samples for all input frequencies
• Useful discrete-time FIR filters
Averaging: Lowpass, linear phase
First-order difference: highpass, linear phase
Make all coefficients 1 for efficient multiplier-free version
5 - 29