6 Image Compression

IMAGE PROCESSING
IMAGE COMPRESSION
By
DR. FERDA ERNAWAN
Faculty of Computer Systems & Software Engineering
ferda@ump.edu.my
OER Digital Image Processing by Ferda Ernawan (editor) work is under licensed Creative Commons Attribution-
NonCommercial-NoDerivatives 4.0 International License.
Today’s Lesson
• Introduction to Image Compression
• General JPEG Image Compression
• Lossless & Lossy compression
• Data Redundancies
• Lossless Compression Methods
– Huffman Coding, Lossless Predictive Coding
• Lossy Compression Methods
– Block Transform Coding
• Learning Outcomes:
To understand image compression processing and the basic
operation
Introduction
• Image compression is a method to reduce the redundancies in
image representation in order to decrease data storage
requirements (Gonzalez and Woods, 2013).
• It is a technique used to compress an image without visually
reducing the quality of the image itself.
• Data vs. Information.
• The goal of these processes is to represent an image with the
same quality level, but in a more solid form.
Why the image must be compressed?
• The large storage requirement of multimedia data.
• The video or image files consume large amount of data and it
always required very high bandwidth networks in transmission as
well as communication costs.
Aims of Image Compression
• Reduce the data storage and maintain the visual image quality (Gonzalez,
Woods and Eddins, 2017).
• Increase the speed of transmission by using the repetition property of data.
• The goal of these processes is to represent an image with the same quality
level, but in a more solid form.
General JPEG Compression
General Image Compression Model
Images taken from Gonzalez & Woods, Digital Image Processing
Encoder
Decoder
Image Compression (Encoder)
Image Image Entropy Coding

Transformati Quantizatio AC Huffman
Table
on n
Zig-zag Huffman
Reordering Coding
Color AC
Scalar
component 8x8 DCT Uniform Bit-
s (Y, Cb, Cr) Transform Quantizatio stream
n
Huffman
Difference Coding
Encoding
DC
Quantizatio DC
n Table Huffman
Table
Block Transform Coding System
• An image is divided into 8x8 pixels of non-overlapping blocks.

• Each block is transformed by the transform function (e.g. DCT)
• Compression is achieved during the quantization of the transformed
coefficients.
DCT Transformation
DCT of an input image A is defined as follows
M 1 N 1
 (2m  1) p  (2n  1)q
B pq   p  q   Amn cos 2M
cos
2N
,
m0 n 0
where  1  1
 ,p0  ,q  0
 M N
p   q  
 2 ,p0  2
,q  0

 M  N
The inverse of DCT is given as follows:

M 1 N 1
 (2m  1) p  (2n  1)q
Apq    p  q Bmn cos 2M
cos
2N
,
m0 n 0
Quantization Process
The JPEG quantization are given as follows:
16 11 10 16 24 40 51 61  17 18 24 47 99 99 99 99 
12 55  18
 12 14 19 26 58 60
 21 26 66 99 99 99 99 
14 13 16 24 40 57 69 56   24 26 56 99 99 99 99 99 
   
14 17 22 29 51 87 80 62  47 66 99 99 99 99 99 99 
QCL  QCR 
18 22 37 56 68 109 103 77   99 99 99 99 99 99 99 99 
   
 24 35 55 64 81 104 113 92   99 99 99 99 99 99 99 99 
 49 64 78 87 103 121 120 101  99 99 99 99 99 99 99 99 
   
 72 92 95 98 112 100 103 99   99 99 99 99 99 99 99 99 
 c( i , j ) k 
u( i , j ) k  Round  
 Q( i , j )
 
Where c(i , j ) k is the frequency image signals at coordinates (i,j) in the

k block.
Differential Pulse Code Modulation
(DPCM) for DC Coefficients
The DC coefficients of each block are encoded with a DPCM instead of
block-based prediction. The DC coefficients are coded by DPCM as
follows:
Diffi  DCi  DCi 1

Where DCi is the current block and DCi 1 represents the previous
block of DC coefficient.
Differential Pulse Code Modulation of DC
Coefficients
Diffi-1 = Diffi =
Diff1 = DC1 DCi-1 - DCi-2 DCi - DCi-1
DC0 DC1 ... DCi-1 DCi
0 ...
Block 1 Block i-1 Block i
Then the difference is encoded using Huffman coding together with

the encoding of AC coefficients.
Zigzag Pattern for AC coefficients
The zigzag pattern is an important part of encoding process which gives
effect to the statistic of symbols used in entropy coding.
Zigzag order under regular of 8×8 discrete transform

DCT as choice of Transform
Original Image
Compressed Image
JPEG image compression produces visible artifact image, less
computational complexity and it has a good energy compactness properties.
Other Compression Method
• Symbol-based Coding – JBIG2 format compression for binary
images.
• LZW (Lempel-Ziv-Welch) Coding – integrated into GIF, TIFF,
PDF.
• Arithmetic Coding.
• Golomb Coding.
• Lossy Predictive Coding – DPCM (Differential pulse code
modulation).
• Wavelet Coding – Haar, Daubechies wavelets.
• JPEG-2000
Compression Ratio (CR)
• Compression is measured in compression ratio denoted by:
Uncompressed File Size

Compressio n Ratio   CR
Compressed File Size
Example:
The original image is 256×256 pixels, 8-bits per pixel grayscale.
The file is 65536 bytes (64 kb) in size.
After compression the image file is 6554 bytes.
The compression ratio is 65536
compressio n Ratio   10
6554
Relative Data Redundancy ()
The relative data redundancy, R can be determined as:
1
RD  1
CR
1
RD  1   0.9
10
This indicates that 90% of its data is redundant. The higher the
RD value, the more data is redundant and will be compressed. If
RD=0, (no redundant data)
Redundancy
Type of Redundancy
There are three types of redundancy:
• Coding redundancy
• Interpixel redundancy
• Psychovisual redundancy
Coding Redundancy
A code is a symbol (letters, numbers, bits) used to represent a body of
information (Gonzalez and Woods, 2016). Each information is
assigned a sequence of code symbols, called a code word. The number
of symbols in each code word is its length.
Symbols with higher appearing probabilities are assigned with codes

of less amount of data.
Symbol: codes that human represent information

E.g., Chinese characters, English letters, …
Coding: transform symbols into more efficient codes
E.g., A: 00, B: 01, C: 10, …
Coding Redundancy: Example
Average bit length (Lavg1)=3 bits, Average bit length (Lavg2)=2.7

bits, Compression ratio= 3/2.7=1.11, Relative Redundancy=0.099
Inter-pixel Redundancy
 Also called spatial redundancy, geometric redundancy, inter-frame

redundancy.
 Results from structural or geometric relationships between objects and the

image.
 Adjacent pixels are usually highly correlated (pixel similar or very close
to neighboring pixels), thus information is unnecessarily replicated in the
representations.
 Neighboring pixels in a natural image are highly correlated. In natural

images, local area usually contains pixels of same or similar gray level.
Inter-pixel Redundancy: Example
Pixels of line 128 of Tiffany:
Pixels of line 128 of wheel:
Inter-pixel Redundancy: Example
• Pixels of line 128 of wheel:
n1 represent data information and n2 is amounts of data to represent

the same information.
Similar segments (graylevel, Coding each segment with 16 bits:
number): (77, 1), (206, 30), CR=(256x8) / (7x16)=18.3
(121, 88), (77,18), (121, 88), RD=1-1/18.3=0.95
(206, 30), (77, 1)
Psychovisual Redundancy
• For human visual perception, certain information has less relative
importance. E.g,: appropriate quantization of gray levels does not
impact its visual quality, e.g.,: Tiffany
Can be eliminated without significant quality loss.

Use Lossy compression or Irreversible compression.
“Quantization”.
Information Theory
Basic Information Theory
• Measuring amount of information: I(p) = log 1/p = – log p

– Average amount of information: entropy.
– Theoretically, the lowest average number of bits required to
represent one symbol.
H ( S )  i pi log pi
Example:
Symbol Probability
A 1/4
B 1/4 In average, minimum 2 bits required.
C 1/4
D 1/4 H (S )   (1 / 4) log 2 (1 / 4)  2
Symbol Probability
A ½
B 1/8
In average, minimum 1.75 bits required.
C 1/8
D 1/4
H (S )  (1 / 2) log 2 (1 / 2)  2(1 / 8) log 2 (1 / 8)  (1 / 4) log 2 (1 / 4)  1.75

Compression Method
There are two types of compression methods:

Lossy image compression
Lossless image compression
Lossy Compression Methods
Lossy image compression methods are required to achieve high
compression ratios for complex images.
An image reconstructed Lossy compression can be performed in
both spatial or transform domains.
The process of quantization-dequantization introduces loss in the
reconstructed image and is inherently responsible for the “lossy”
nature of the compression scheme.
The quantized transform coefficient is computed by
 B pq 
T pq  Round  
Q 
 pq 
Where B pq is the frequency image signals at coordinates (i,j) in the
k block.
Lossy Compression Methods
During dequantization, approximate DCT coefficients are obtained by
multiplying the corresponding quantization threshold with the
quantized coefficient as follows:
B pq'  Tpq  Q pq
Lossless Compression Methods
Lossless compression methods are needed in some digital imaging
applications, such as: medical images, x-ray images, etc.
Lossless compression techniques: run-length coding, Huffman coding,

lossless predictive coding etc.
Run-Length Encoding of AC
Coefficients
The frequency image signals after quantization process consist of
many zeros. Then, a special condition as known end-of-block
(EOB) is applied to get an efficient in the entropy code.
A symbol of EOB indicates that the rest of the coefficients in the
block are zero. Next, run-length encoding exploits the repeating
frequency image signals as the symbols in the sequence a set of the
AC coefficients.
The output of run-length encoding represents a sequence value with
the consecutive repetition as symbols and the length of occurrence
of the symbols
Huffman Coding
■Huffmancoding is a famous method that uses variable-length
codes (VLC) tables for compressing data (Jayaraman,
Veerakumar, and Esakkirajan, 2017).
■Given a set of data symbols and their probabilities, the method

creates a set of variable-length codeword with the shortest
average length and assigns them to the symbols.
Huffman Coding (cont..)
Example I
Example I- Cont..
Example I- Cont..
Example II
Example II
Example II- Cont..
Binary coderwork tree representation

Example II- Cont..
Note: Log2 x = Log10 . Log2 (10) = 3.3219 Log10 x
Summary of Huffman Code
 Achieve minimal redundancy subject to the constraint that the
source symbols be coded one at a time.
 Sorting symbols in descending probabilities is the key in the step of
source reduction.
 The codeword assignment is not unique. Exchange the labeling of
“0” and “1” at any node of binary codeword tree would produce
another solution that equally works well.
 Only works for a source with finite number of symbols (otherwise,
it does not know where to start).
References
• R.C. Gonzalez and R.E. Woods, 2016. Digital Image Processing,
Pearson Education India; Third edition.
• A.K. Jain, 2015. Fundamentals of Digital Image Processing, Pearson
Education India; First edition.
• R.C. Gonzalez, R.E. Woods and S.L. Eddins, 2017. Digital Image
Processing Using MATLAB. McGraw Hill Education; 2 edition.
• S. Jayaraman, T. Veerakumar, S. Esakkirajan, 2017.Digital Image
Processing, McGraw Hill Education; 1 edition.

6 Image Compression

Uploaded by

Copyright:

Available Formats

6 Image Compression

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

6 Image Compression

Uploaded by

Copyright:

Available Formats

IMAGE PROCESSING

• Increase the speed of transmission by using the repetition property of data.

Image Image Entropy Coding

• An image is divided into 8x8 pixels of non-overlapping blocks.

The inverse of DCT is given as follows:

Where c(i , j ) k is the frequency image signals at coordinates (i,j) in the

Diffi  DCi  DCi 1

DC0 DC1 ... DCi-1 DCi

Block 1 Block i-1 Block i

Then the difference is encoded using Huffman coding together with

Zigzag order under regular of 8×8 discrete transform

Uncompressed File Size

Symbols with higher appearing probabilities are assigned with codes

Symbol: codes that human represent information

Average bit length (Lavg1)=3 bits, Average bit length (Lavg2)=2.7

 Also called spatial redundancy, geometric redundancy, inter-frame

 Results from structural or geometric relationships between objects and the

 Neighboring pixels in a natural image are highly correlated. In natural

Pixels of line 128 of wheel:

n1 represent data information and n2 is amounts of data to represent

Can be eliminated without significant quality loss.

• Measuring amount of information: I(p) = log 1/p = – log p

H (S )  (1 / 2) log 2 (1 / 2)  2(1 / 8) log 2 (1 / 8)  (1 / 4) log 2 (1 / 4)  1.75

There are two types of compression methods:

Lossless compression techniques: run-length coding, Huffman coding,

■Given a set of data symbols and their probabilities, the method

Binary coderwork tree representation

Note: Log2 x = Log10 . Log2 (10) = 3.3219 Log10 x

You might also like