Chapter08 Image Compression

Image Compression
Data vs Information
Information = Matter ()
Data = The means by which information is conveyed
Image Compression
Reducing the amount of data required to represent

a digital image while keeping information as much as
possible
Relative Data Redundancy and Compression Ratio
Relative Data Redundancy
1
RD 1
CR
Compression Ratio
n1
CR
n2
Types of data redundancy

1. Coding redundancy
2. Interpixel redundancy
3. Psychovisual redundancy
Coding Redundancy
Different coding methods yield different amount of data needed to

represent the same information.
Example of Coding Redundancy :

Variable Length Coding vs. Fixed Length Coding
Lavg 3 bits/symbol Lavg 2.7 bits/symbol

(Images from Rafael C. Gonzalez and Richard E.
Wood, Digital Image Processing, 2nd Edition.
Variable Length Coding
Concept: assign the longest code word to the symbol with the least
probability of occurrence.

Interpixel Redundancy
Interpixel redundancy:
Parts of an image are
highly correlated.
In other words,we can

predict a given pixel
from its neighbor.
Run Length Coding
The gray scale image

of size 343x1024 pixels
Binary image
= 343x1024x1 = 351232 bits
Line No. 100
Run length coding
Line 100: (1,63) (0,87) (1,37) (0,5) (1,4) (0,556) (1,62) (0,210)
Total 12166 runs, each run use 11 bits Total = 133826 Bits
Psychovisual Redundancy
8-bit gray scale 4-bit gray scale 4-bit IGS
image image image
False
contours
The eye does not response with equal sensitivity to all visual
information. (Images from Rafael C. Gonzalez and Richard E.
Improved Gray Scale Quantization
Pixel Gray level Sum IGS Code

i-1 N/A 0000 0000 N/A
+
i 0110 1100 0110
0110 1100
i+1 1000 1011 1001 0111 1001
i+2 1000 0111 1000 1110 1000
i+3 1111 0100 1111 0100 1111
Algorithm
1. Add the least significant 4 bits of the previous value of Sum to
the 8-bit current pixel. If the most significant 4 bit of the pixel is
1111 then add 0000 instead. Keep the result in Sum
2. Keep only the most significant 4 bits of Sum for IGS code.
Fidelity Criteria: how good is the compression algorithm
-Objective Fidelity Criterion

- RMSE, PSNR
-Subjective Fidelity Criterion:

-Human Rating

Image Compression Models
Reduce data Increase noise
redundancy immunity
f ( x, y ) Source encoder Channel encoder
Noise Channel
f ( x, y ) Source decoder Channel decoder

Source Encoder and Decoder Models
Source encoder
f ( x, y ) Mapper Quantizer Symbol encoder
Reduce Reduce Reduce

interpixel psychovisual coding
redundancy redundancy redundancy
Source decoder
f ( x, y ) Inverse mapper Symbol decoder

Channel Encoder and Decoder
- Hamming code, Turbo code,

Information Theory
Measuring information
1
I ( E ) log log P( E )
P( E )
Entropy or Uncertainty: Average information per symbol
H P( a j ) log( P ( a j ))
j
Simple Information System
Binary Symmetric Channel

Source Destination
A = {a1, a2} ={0, 1} B = {b1,b2} ={0, 1}
z = [P(a1), P(a2)] v = [P(b1), P(b2)]
(1-Pe)
P(a1) 0 0 P(a1)(1-Pe)+(1-P(a1))Pe
Pe
Source Destination
Pe
1-P(a1) 1 1 (1-P(a1))(1-Pe)+P(a1)Pe
Pe= probability of error (1-Pe) (Images from Rafael C. Gonzalez and Richard E.
Source Destination
A = {a1, a2} ={0, 1} B = {b1,b2} ={0, 1}
z = [P(a1), P(a2)] v = [P(b1), P(b2)]
H(z) = - P(a1)log2P(a1) H(z|b1) = - P(a1|b1)log2P(a1|b1)
- P(a2)log2P(a2) - P(a2|b1)log2P(a2|b1)
H(z|b2) = - P(a1|b2)log2P(a1|b2)
- P(a2|b2)log2P(a2|b2)
H(z|v) = H(z|b1) + H(z|b2)
Mutual
information I(z,v)=H(z) - H(z|v)
Capacity C max I ( z, v )
z
Let pe = probability of error
p(a1 ) pbs
z
1 p ( a )
1 p
bs
1 pe pe pbs
v p
p e 1 pe bs
H ( z ) pbs log2 ( pbs ) (1 pbs ) log2 (1 p bs )
H ( z | v ) pbs (1 pe ) log2 ( pbs (1 pe )) (1 pbs ) pe log2 ((1 p bs ) pe )

(1 pbs )(1 pe ) log2 ((1 pbs )(1 pe )) pbs pe log2 ( p bs pe )
I ( z, v ) H bs ( pbs pe pbs pe ) H bs ( pe )
C 1 H bs ( pe )

Communication System Model
2 Cases to be considered: Noiseless and noisy

Noiseless Coding Theorem
Problem: How to code data as compact as possible?
Shannons first theorem: defines the minimum average code word

length per source that can be achieved.
Let source be {A, z} which is zero memory source with J symbols.

(zero memory = each outcome is independent from other outcomes)
then a set of source output of n element be
A {1 , 2 , 3 ,... J n }
Example:
A {0,1}
for n = 3,
A {000,001,010,011,100,101,110,111}
Noiseless Coding Theorem (cont.)
Probability of each j is
P ( j ) P ( a j1 ) P (a j 2 ) P (a jn )
Entropy of source :
jn
H ( z) P(i ) log( P(i )) nH ( z )
i 1
Each code word length l(i) can be
1 1
log l (i ) log 1
P(i ) P(i )
Then average code word length can be

Jn J n
J n
1 1

i 1
P ( i) log P( i)l (i ) P ( i) log
P(i ) i 1 i 1 P(i )
1
Noiseless Coding Theorem (cont.)
We get H ( z) Lavg
H ( z) 1
from H ( z) nH ( z )
then

Lavg 1
H ( z) H ( z)
n n
or

Lavg The minimum average code word
lim H ( z) length per source symbol cannot
n
n lower than the entropy.
Coding efficiency
H ( z)
n

Lavg
Extension Coding Example
H = 0.918
Lavg = 1
H = 1.83
Lavg = 1.89
0.918 1.83
1 1 0.918 2 0.97
1 1.89

Noisy Coding Theorem
Problem: How to code data as reliable as possible?
Example: Repeat each code 3 times:

Source data = {1,0,0,1,1}
Data to be sent = {111,000,000,111,111}
Shannons second theorem: the maximum rate of coded information is

R log
r
= code size
r = Block length
Rate Distortion Function for BSC

Error-Free Compression: Huffman Coding
Huffman coding: give the smallest possible number of

code symbols per source symbols.
Step 1: Source reduction

Error-Free Compression: Huffman Coding
Step 2: Code assignment procedure
The code is instantaneous uniquely decodable without referencing

succeeding symbols.

Near Optimal Variable Length Codes

Arithmetic Coding
Nonblock code: one-to-one correspondence between source symbols
And code words does not exist.
Concept: The entire sequences of source symbols is assigned a single
arithmetic code word in the form of a number in an interval of real
number between 0 and 1.

Arithmetic Coding Example
0.2x0.4 0.04+0.8x0.04 0.056+0.8x0.016
The number
between 0.0688
and 0.06752
can be used to
represent the
sequence
a1 a2 a3 a3 a4
0.2x0.2 0.04+0.4x0.04 0.056+0.4x0.016

LZW Coding
Lempel-Ziv-Welch coding : assign fixed length code words to
variable length sequences of source symbols.
24 Bits
9 Bits
LZW Coding Algorithm
0. Initialize a dictionary by all possible gray values (0-255)
1. Input current pixel
2. If the current pixel combined with previous pixels
form one of existing dictionary entries
Then
2.1 Move to the next pixel and repeat Step 1
Else
2.2 Output the dictionary location of the currently recognized
sequence (which is not include the current pixel)
2.3 Create a new dictionary entry by appending the currently

recognized sequence in 2.2 with the current pixel
2.4 Move to the next pixel and repeat Step 1

LZW Coding Example
Dictionary Currently
Location Entry Input recognized Encoded
0 0 pixel Sequences Output (9 bits)
1 1 39 39
39 39 39
255 255 126 126 39
256 39-39 126 126 126
257 39-126 39 39 126
258 126-126 39 39-39
259 126-39 126 126 256
260 39-39-126 126 126-126
261 126-126-39 39 39 258
262 39-39-126-126 39 39-39
126 39-39-126
126 126 260
Bit-Plane Coding
Original image Binary image

Bit 7 compression
Bit 6 Binary image

compression

Bit 0 Binary image

compression
Bit plane
images
Example of binary image compression: Run length coding

Bit Planes
Bit 7 Bit 3
Bit 6 Bit 2
Original gray
scale image
Bit 5 Bit 1
Bit 4 Bit 0
(Images from Rafael C.
Gonzalez and Richard E.
Wood, Digital Image
Processing, 2nd Edition.
Gray-coded Bit Planes
Original
bit planes a Gray code:
7 g7
gi ai ai 1
for 0i6
and
a6 g6
g7 a7
ai= Original bit planes

a5 g5 = XOR
a4 g4 (Images from Rafael C.

Wood, Digital Image
Processing, 2nd Edition.
Gray-coded Bit Planes (cont.)
There are less 0-1 and 1-0

a3 g3 transitions in grayed code
bit planes.
Hence gray coded bit planes
are more efficient for coding.
a2 g2
a1 g1
a0 g0
Relative Address Coding (RAC)
Concept: Tracking binary transitions that begin and end eack black
and white run

Contour tracing and Coding
Represent each contour by a set of boundary points and directionals.

Error-Free Bit-Plane Coding

Lossless VS Lossy Coding
Source encoder
Lossless f ( x, y ) Mapper Symbol encoder
coding
Reduce Reduce
interpixel coding
redundancy redundancy
Source encoder
Lossy f ( x, y ) Mapper Quantizer Symbol encoder
coding
Reduce Reduce Reduce

interpixel psychovisual coding
redundancy redundancy redundancy
Transform Coding (for fixed resolution transforms)
Encoder
Input image Construct

Forward Symbol
(NxN) nxn Quantizer
transform encoder
subimages
Quantization process causes Compressed

The transform coding lossy image
Decoder
Construct
Decompressed Inverse Symbol
nxn
image transform decoder
subimages
Examples of transformations used for image compression: DFT and DCT

Transform Coding (for fixed resolution transforms)
3 Parameters that effect transform coding performance:
1. Type of transformation
2. Size of subimage
3. Quantization algorithm
2D Discrete Transformation
Forward transform:
N 1 N 1
T (u, v ) f ( x, y ) g ( x, y , u, v )
x 0 y 0
where g(x,y,u,v) = forward transformation kernel or basis function

T(u,v) is called the transform coefficient image.
Inverse transform:
N 1 N 1
f ( x, y ) T (u, v )h( x, y , u, v )
u 0 v 0
where h(x,y,u,v) = inverse transformation kernel or

inverse basis function
Transform Example: Walsh-Hadamard Basis Functions
m 1
1 bi ( x ) pi ( u )bi ( y ) pi ( v )
g ( x, y , u, v ) h(u, v, x, y ) ( 1) i 0
N
N = 2m
bk(z) = the kth bit of z
p0 (u ) bm1 (u )
p1 (u ) bm 1 (u ) bm 2 (u )
p2 (u ) bm 2 (u ) bm 3 (u )

pm 1 (u ) b1 (u ) b0 (u )
Advantage: simple, easy to implement
N=4 Disadvantage: not good packing ability
Transform Example: Discrete Cosine Basis Functions
(2 x 1)u ( 2 y 1)v
g ( x, y , u, v ) h(u, v, x, y ) (u ) (v ) cos cos
2N 2N
1
for u 0
(u ) N
2
for u 1, , N 1
N
DCT is one of the most frequently

N=4 used transform for image compression.
For example, DCT is used in JPG files.
Advantage: good packing ability,
N=4 modulate computational complexity
Transform Coding Examples
Fourier Error
RMS Error = 1.28
Hadamard
Original image
512x512 pixels
Subimage size:
8x8 pixels = 64 pixels RMS Error = 0.86
DCT
Quatization by truncating
50% of coefficients (only
32 max cofficients are kept.)
(Images from Rafael C.
Wood, Digital Image
Processing, 2nd Edition. RMS Error = 0.68
DCT vs DFT Coding
DFT
coefficients
have abrupt
changes at
boundaries
of blocks
1 Block
Advantage of DCT over DFT is that the DCT coefficients are

more continuous at boundaries of blocks.
Subimage Size and Transform Coding Performance
This experiment:
Quatization is made by
truncating 75% of
transform coefficients
DCT is the best
Size 8x8 is enough

Subimage Size and Transform Coding Performance
Reconstructed DCT Coefficients

by using 25%
of coefficients
(CR = 4:1)
with 8x8 sub-
images
Zoomed detail
Zoomed detail Subimage size:
Original 2x2 pixels
Zoomed detail Zoomed detail

Subimage size: Subimage size:
4x4 pixels 8x8 pixels

Quantization Process: Bit Allocation
To assign different numbers of bits to represent transform

coefficients based on importance of each coefficient:
- More importance coefficeitns assign a large number of bits
- Less importance coefficients assign a small number of bits

or not assign at all
2 Popular bit allocation methods

1. Zonal coding : allocate bits based on the basis of
maximum variance, using fixed mask for all subimages
2. Threshold coding : allocate bits based on maximum

magnitudes of coefficients
Example: Results with Different Bit Allocation Methods
Reconstructed Reconstructed
by using 12.5% by using 12.5%
of coefficients of coefficients
(8 coefficients (8 coefficients
with largest with largest
magnitude are variance are
used) used)
Threshold coding Zonal coding

Error Error
Zoom details

Zonal Coding Example
Zonal mask Zonal bit allocation

Threshold Coding Example
Threshold mask Thresholded coefficient

ordering

Thresholding Coding Quantization
3 Popular Thresholding Methods
Method 1: Global thresholding : Use a single global threshold
value for all subimages
Method 2: N-largest coding: Keep only N largest coefficients
Method 3: Normalized thresholding: each subimage is normalized by a
normalization matrix before rounding
Bit allocation T ( u, v )
T (u, v ) round
Z ( u , v )
Restoration before decompressing
~
T (u, v ) T (u, v ) Z (u, v )
Example of
Normalization
Matrix Z(u,v) (Images from Rafael C. Gonzalez and Richard E.
DCT Coding Example
(CR = 38:1) (CR = 67:1)
Method:
- Normalized
Error image Thresholding,
RMS Error = 3.42 - Subimage size:
8x8 pixels
Blocking
Zoom details Artifact at
Subimage
boundaries
Wavelet Transform Coding: Multiresolution approach
Encoder
Input image Wavelet Symbol

(NxN) Quantizer
transform encoder
Compressed
image
Decoder
Inverse
Decompressed Symbol
wavelet
image decoder
transform
Unlike DFT and DCT, Wavelet transform is a multiresolution transform.

What is a Wavelet Transform
One up on a time, human uses a line to represent a number.
For example
= 25
With this numerical system, we need a lot of space to represent
a number 1,000,000.
Then, after an Arabic number system is invented, life is much easier.
We can represent a number by a digit number:
X,XXX,XXX
An Arabic number is one The 1st digit = 1x

kind of multiresolution
Representation. The 2nd digit = 10x
The 3rd digit = 100x

Like a number, any signal can also be represented by a multiresolution
data structure, the wavelet transform.
What is a Wavelet Transform
Wavelet transform has its background from multiresolution

analysis and subband coding.
Other important background:

- Nyquist theorem: The minimun sampling rate needed for sampling
a signal without loss of information is twice the maximum frequency
of the signal.
-We can perform frequency shift by multiplying a complex sinusiodal

signal in time domain.
f ( x, y )e j 2 ( u0 x v0 y ) F (u u0 , v v0 )
Wavelet History: Image Pyramid
If we smooth and then down sample an image repeatedly, we will
get a pyramidal image:
Coarser, decrease resolution
Finer, increase resolution
Pyramidal structured image (Images from Rafael C. Gonzalez and Richard E.

Image Pyramid and Multiscale Decomposition
Down
Image Image
Smooth Sampling
NxN N/2xN/2
By 2
Question: What Up
Information is Sampling
loss after down By 2
Sampling?
Answer: Loss Interpolate

Information is
Prediction + A prediction
error image: Predicted
Error
Image
(loss details) -
NxN
NxN
Image Pyramid and Multiscale Decomposition (cont.)
Hence we can decompose an image using the following process
Approxi- Approxi-
Smooth and Smooth and
Image -mation -mation
down sampling down sampling .
NxN Image Image
by 2 by 2
N/2xN/2 N/4xN/4
Up Up
sampling sampling
by 2 and by 2 and
interpolate interpolate
- -
+ +
Prediction Prediction
Error Error
NxN N/2xN/2
Image Pyramid and Multiscale Decomposition (cont.)
Approximation image N/8xN/8
Prediction error
Original Image N/4xN/4
=
NxN Prediction
error
N/2xN2
Prediction error
(residue)
NxN
Multiresolution Representation
Multiresolution Decomposition Process
Note that this process is not a wavelet decomposition process !

Example of Pyramid Images
Approximation
Images (using
Gaussian
Smoothing)
Prediction
residues

Subband Coding
Subband decomposition process
h0(n) Approximation
LPF Down
a(n)
Sampling
N/2 points
by 2
x(n)
N points
h1(n) Detail
Down
HPF Freq. d(n)
Sampling
shift by N/2 points
by 2
N/2
All information of x(n) is completely

preserved in a(n) and d(n).
Subband Coding (cont.)
Subband reconstruction process
g0(n)
Up
a(n)
Sampling
N/2 points Interpolation
by 2
x(n)

N points
g1(n)
Up
d(n) Freq.
Sampling
N/2 points Interpolation shift by
by 2
N/2
Subband Coding (cont.)

2D Subband Coding

Example of 2D Subband Coding
Vertical detail:
Approximation: filtering in x-
filtering in both direction using
x and y directions h0(n) and in y-
using h0(n) direction using
h1(n)
Horizontal detail: Diagonal detail:

filtering in x- filtering in both
direction using x and y directions
h1(n) and in y- using h1(n)
direction using
h0(n)

1D Discrete Wavelet Transformation
h(n) 2 d1(n) N/2 points

x(n)
N points
h(n) 2
h (n) 2 d2(n) N/4 points
h (n) 2
Note that the number
of points of x(n) and
wavelet coefficients h (n) 2 d3(n) N/8 points
are equal.
h (n) 2 a3(n) N/8 points
(n) = a wavelet function
(n) = a scaling function Wavelet coefficients
(N points)
d = diagonal detail
Original image h = horizontal detail
NxN v = vertical detail
a = approximation
d1 h1
v1 a1
d2 h2
Level 1 Level 3
v2 a2
d3 h3
Level 2 v3 a3
2D Discrete Wavelet Transformation (cont.)
a3h3 h2
v3d3 h1
Original image v2 d2
NxN
v1 d1
Wavelet coefficients
NxN
d = diagonal detail: filtering in both x and y directions using h (n)

h = horizontal detail: filtering in x-direction using h (n) and in y
direction using h (n)
v = vertical detail: filtering in x-direction using h (n) and in y
direction using h (n)
a = approximation: filtering in both x and y directions using h (n)
Example of 2D Wavelet Transformation
Original
Image
Original image
Example of 2D Wavelet Transformation (cont.)
LL1 HL1
LH1 HH1
The first level wavelet decomposition

LL2 HL2
HL1
LH2 HH2
LH1 HH1
The second level wavelet decomposition

LL3 HL3
HL2
LH3 HH3
HL1
LH2 HH2
LH1 HH1
The third level wavelet decomposition

Example of 2D Wavelet Transformation

Examples: Types of Wavelet Transform
Haar Daubechies
wavelets wavelets
Biorthogonal
Symlets wavelets

Wavelet Transform Coding for Image Compression
Encoder
Input image Wavelet Symbol

(NxN) Quantizer
transform encoder
Compressed
image
Decoder
Inverse
Decompressed Symbol
wavelet
image decoder
transform
Unlike DFT and DCT, Wavelet transform is a multiresolution transform.

Wavelet Transform Coding Example
(CR = 38:1) (CR = 67:1)
Error Image Error Image

RMS Error = 2.29 RMS Error = 2.96
Zoom details
No blocking
Artifact
Wavelet Transform Coding Example (cont.)
(CR = 108:1) (CR = 167:1)
Error image Error image

Zoom details

Wavelet Transform Coding vs. DCT Coding
Wavelet DCT 8x8
(CR = 67:1) (CR = 67:1)
Error image Error image

Zoom details

Type of Wavelet Transform and Performance

No. of Wavelet Transform Level and Performance
Threshold Level and Performance

Table 8.14 (Cont)

Table 8.19 (Cont)

Lossless Predictive Coding Model

Lossless Predictive Coding Example

Lossy Predictive Coding Model

Delta Modulation

Linear Prediction Techniques: Examples

Quantization Function

Lloyd-Max Quantizers

Lossy DCPM

DCPM Result Images

Error Images of DCPM


Chapter08 Image Compression

Uploaded by

Document Information

Original Title

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

Chapter08 Image Compression

Uploaded by

Copyright:

Available Formats

Image Compression

Data = The means by which information is conveyed

Reducing the amount of data required to represent

Types of data redundancy

Different coding methods yield different amount of data needed to

Example of Coding Redundancy :

Lavg 3 bits/symbol Lavg 2.7 bits/symbol

(Images from Rafael C. Gonzalez and Richard E.

In other words,we can

The gray scale image

Line No. 100

Run length coding

Pixel Gray level Sum IGS Code

-Objective Fidelity Criterion

-Subjective Fidelity Criterion:

(Images from Rafael C. Gonzalez and Richard E.

f ( x, y ) Source decoder Channel decoder

Reduce Reduce Reduce

f ( x, y ) Inverse mapper Symbol decoder

- Hamming code, Turbo code,

Entropy or Uncertainty: Average information per symbol

Binary Symmetric Channel

H ( z ) pbs log2 ( pbs ) (1 pbs ) log2 (1 p bs )

H ( z | v ) pbs (1 pe ) log2 ( pbs (1 pe )) (1 pbs ) pe log2 ((1 p bs ) pe )

(Images from Rafael C. Gonzalez and Richard E.

2 Cases to be considered: Noiseless and noisy

(Images from Rafael C. Gonzalez and Richard E.

Shannons first theorem: defines the minimum average code word

Let source be {A, z} which is zero memory source with J symbols.

Each code word length l(i) can be

Then average code word length can be

(Images from Rafael C. Gonzalez and Richard E.

Example: Repeat each code 3 times:

Data to be sent = {111,000,000,111,111}

Shannons second theorem: the maximum rate of coded information is

(Images from Rafael C. Gonzalez and Richard E.

Huffman coding: give the smallest possible number of

Step 1: Source reduction

(Images from Rafael C. Gonzalez and Richard E.

Step 2: Code assignment procedure

The code is instantaneous uniquely decodable without referencing

(Images from Rafael C. Gonzalez and Richard E.

(Images from Rafael C. Gonzalez and Richard E.

(Images from Rafael C. Gonzalez and Richard E.

0.2x0.2 0.04+0.4x0.04 0.056+0.4x0.016

(Images from Rafael C. Gonzalez and Richard E.

2.3 Create a new dictionary entry by appending the currently

2.4 Move to the next pixel and repeat Step 1

Original image Binary image

Bit 6 Binary image

Bit 0 Binary image

Example of binary image compression: Run length coding

ai= Original bit planes

a4 g4 (Images from Rafael C.

There are less 0-1 and 1-0

(Images from Rafael C. Gonzalez and Richard E.

(Images from Rafael C. Gonzalez and Richard E.

(Images from Rafael C. Gonzalez and Richard E.

Reduce Reduce Reduce

Input image Construct

Quantization process causes Compressed

Examples of transformations used for image compression: DFT and DCT

3 Parameters that effect transform coding performance: