An Introduction To Image Compression

1
An Introduction to
An Introduction to
Image Compression
Image Compression
Speaker: Wei-Yi Wei
Advisor: Jian-Jung Ding
Digital Image and Signal Processing Lab
GICE, National Taiwan University
NTU, GICE, MD531, DISP Lab An Introduction to Image Compression
Wei-Yi Wei 2
Outline
Outline
Image Compression Fundamentals
General Image Storage System
Color Space
Reduce correlation between pixels
Karhunen-Loeve Transform
Discrete Cosine Transform
Discrete Wavelet Transform
Differential Pulse Code Modulation
Differential Coding
Quantization and Source Coding
Huffman Coding
Arithmetic Coding
Run Length Coding
Lempel Ziv 77 Algorithm
Overview of Image Compression Algorithms
JPEG
JPEG 2000
Shape-Adaptive Image Compression
Wei-Yi Wei 3
Outline
Outline
Wei-Yi Wei 4
Wei-Yi Wei 5
Color Specification
Color Specification
Luminance
Received brightness of the light, which is proportional to the total
energy in the visible band.
Chrominance
Describe the perceived color tone of a light, which depends on the
wavelength composition of light
Chrominance is in turn characterized by two attributes
Hue
Specify the color tone, which depends on the peak wavelength of the light
Saturation
Describe how pure the color is, which depends on the spread or bandwidth of the
light spectrum
Wei-Yi Wei 6
YUV Color Space
YUV Color Space
In many applications, it is desirable to describe a color in terms
of its luminance and chrominance content separately, to enable
more efficient processing and transmission of color signals
One such coordinate is the YUV color space
Y is the components of luminance
Cb and Cr are the components of chrominance
The values in the YUV coordinate are related to the values in the RGB
coordinate by
0.299 0.587 0.114 0
0.169 0.334 0.500 128
0.500 0.419 0.081 128
Y R
Cb G
Cr B
_ _ _ _

+

, , , ,
Wei-Yi Wei 7
Spatial Sampling of Color Component
Spatial Sampling of Color Component
The three different chrominance downsampling format
Wei-Yi Wei 8
The Flow of Image Compression (1/2)
What is the so-called image compression coding?
To store the image into bit-stream as compact as possible and to display
the decoded image in the monitor as exact as possible
Flow of compression
The image file is converted into a series of binary data, which is called
the bit-stream
The decoder receives the encoded bit-stream and decodes it to reconstruct
the image
The total data quantity of the bit-stream is less than the total data quantity
of the original image
Wei-Yi Wei 9
Measure to evaluate the performance of image compression
Root Mean square error:
Peak signal to noise ratio:
Compression Ratio:
Where n1 is the data rate of original image and n2 is that of the encoded
bit-stream
The flow of encoding
Reduce the correlation between pixels
Quantization
Source Coding
[ ]
1 1
2
0 0
( , ) '( , )
W H
x y
f x y f x y
RMSE
WH

10
255
20log PSNR
MSE
1
2
n
Cr
n

Wei-Yi Wei 10
Outline
Outline
Wei-Yi Wei 11
Reduce the Correlation between Pixels
Reduce the Correlation between Pixels
Orthogonal Transform Coding
KLT (Karhunen-Loeve Transform)
Maximal Decorrelation Process
DCT (Discrete Cosine Transform)
JPEG is a DCT-based image compression standard, which is a lossy coding
method and may result in some loss of details and unrecoverable distortion.
Subband Coding
DWT (Discrete Wavelet Transform)
To divide the spectrum of an image into the lowpass and the highpass
components, DWT is a famous example.
JPEG 2000 is a 2-dimension DWT based image compression standard.
Predictive Coding
DPCM
To remove mutual redundancy between seccessive pixels and encode only the
new information
Wei-Yi Wei 12
Covariance
Covariance
The covariance between two random variables X and Y, with expected value
E[X]= and E[Y]= is defined as
If entries in the column vector X = [x
1
,x
2
,,x
N
]
T
are random variables, each
with finite variance, then the covariance matrix C is the matrix whose (i, j)
entry is the covariance
Y
cov( , ) [( )( )] [ ]
X Y X Y
X Y E X Y E XY
( )( ) ( , ) (discrete)
X Y
X Y
X Y f X Y
11 12 1
21 22 2
1 2
C [( )( ) ]
n
n
T
x x
n n nn
C C C
C C C
E X X
C C C

1
1
1

1
1
]
L
L
M M M M
L
1 2 1 2 [ ] [ ( ) ( ) ( )]
T T
x n n E x E x E x K K
Wei-Yi Wei 13
The Orthogonal Transform
The Linear Transform
The forward transform
The inverse transform
If we want to obtain the inverse
transform, we need to compute the
inverse of the transform matrix
since
The forward transform
The inverse transform
If we want to obtain the inverse
transform, we need not to compute
the inverse of the transform matrix
since
T
y V x
y Ax
x Vy
x Vy
1
1
1
( ) ( )
V A
VA A A I
x Vy V Ax A A x x

( ) ( )
T T
T T
VV V V I
x Vy V V x V V x x

Wei-Yi Wei 14
Transform Coding (1/2)
Original sequence X=(x
0
,x
1
) Transformed Sequence Y=(y
0
,y
1
)
Height Weight
65 170
75 188
60 150
70 170
56 130
80 203
68 160
50 110
40 80
Height Weight
181.97068384838812 3.416170333863853
202.40605916474945 0.887250469682385
161.55397378997284 0.559957738379552
183.8437168154677 -1.219748938970085
141.51187032497342 -3.223438711671541
206.13345984096256 -2.999455616319722
173.822665082968 -3.111447163995635
120.72055367314222 -5.152467452589058
89.15897210197947 -7.11882670939854
The original sequence tends to cluster
around the line x
1
=2.5x
0
. We rotate the
sequence by the transform : Y AX
cos sin
Rotation Matrix :
sin cos
A

1
]
0 10 20 30 40 50 60 70 80 90
0
20
40
60
80
100
120
140
160
180
200
x0
x
1
Transform
68
o
Y AX
Wei-Yi Wei 15
Throw the value of weight y1
Height Weight
181.97068384838812 0
202.40605916474945 0
161.55397378997284 0
183.8437168154677 0
141.51187032497342 0
206.13345984096256 0
173.822665082968 0
120.72055367314222 0
89.15897210197947 0
^
1
X A Y
Inverse Transform
Height Weight
68.16741797800859 168.72028006870275
75.82264431044632 187.66763012404562
60.51918377426525 149.79023613916877
68.86906847716196 170.45692599485025
53.01127967035257 131.20752139486424
77.21895318005868 191.12361585053173
65.11511642520563 161.165568622698
45.222715366778566 111.93014828010075
33.399538811586865 82.66675942272599
Height Weight
65 170
75 188
60 150
70 170
56 130
80 203
68 160
50 110
40 80
Because the other element of the pair
contained very little information, we could
discard it without a significant effect on the
fidelity of the reconstructed sequence
^
^
0 cos sin 0
sin cos 0
1
x y
x

1
1 1
1

1 1
1
] ]
]
Wei-Yi Wei 16
KLT is the optimum transform coder that is defined as the one that minimizes
the mean square distortion of the reproduced data for a given number of total
bits
The KLT
X: The input vector with size N-by-1
A: The transform matrix with size N-by-N
Y: The transformed vector with size N-by-1, and each components v(k) are mutually uncorrelated
Cxixj: The covariance matrix of xi and xj
Cyiyj: The covariance matrix of yi and yj
The transform matrix A is composed of the eigenvectors of the autocorrelation matrix Cxixj, which
makes the output autocorrelation matrix Cyiyj be composed of the eigenvalues in

the diagonal direction. That is
0
1
1
0 0
0 0
0 0
0 0
T
yy xx
N
C AC A

1
1
1

1
1
]
L
M
M O
L
0, 1 1 ,... N
[( ( ))( ( )) ]
[ ] :
[( )( ) ] [( )]
[ ]
T
yy
T
T T T
T T T
xx
C E Y E Y Y E Y
E YY zero mean assumption
E Ax Ax E Axx A
AE xx A AC A

Wei-Yi Wei 17
Discrete Cosine Transform (1/2)
Forward DCT
Inverse DCT
1 1
0 0
2 (2 1) (2 1)
( , ) ( ) ( ) ( , )cos cos
2 2
for 0,..., 1 and 0,..., 1
1/ 2 for 0
where 8 and ( )
1 otherwise
N N
x y
x u y v
F u v C u C v f x y
N N N
u N v N
k
N C k

+ +
1 1
1 1
] ]

'
1 1
0 0
2 (2 1) (2 1)
( , ) ( ) ( ) ( , )cos cos
2 2
for 0,..., 1 and 0,..., 1 where 8
N N
u v
x u y v
f x y C u C v F u v
N N N
x N y N N

+ +
1 1
1 1
] ]

Why DCT is more appropriate for image compression than DFT?

The DCT can concentrate the energy of the transformed signal in low frequency,
whereas the DFT can not
For image compression, the DCT can reduce the blocking effect than the DFT
Wei-Yi Wei 18
The 8-by-8 DCT basis
u
v
Wei-Yi Wei 19
Discrete Wavelet Transform (1/2)
Subband Coding
The spectrum of the input data is decomposed into a set of bandlimitted
components, which is called subbands
Ideally, the subbands can be assembled back to reconstruct the original spectrum
without any error
The input signal will be filtered into lowpass and highpass components
through analysis filters
The human perception system has different sensitivity to different frequency
band
The human eyes are less sensitive to high frequency-band color components
The human ears is less sensitive to the low-frequency band less than 0.01 Hz and
high-frequency band larger than 20 KHz
Wei-Yi Wei 20

( , ) ( ) ( ) x y x y
( , ) ( ) ( )
H
x y x y
( , ) ( ) ( )
V
x y y x
( , ) ( ) ( )
D
x y x y
2D scaling function
2D wavelet function
( ), ( ) x y
( ), ( ) x y
1D scaling function
1D wavelet function
L H
LL
HL
LH
HH
1D DWT applied alternatively to vertical and horizontal direction line by line
The LL band is recursively decomposed, first vertically, and then horizontally
Wei-Yi Wei 21
DPCM (1/3)
DPCM (1/3)
DPCM CODEC
[ ] [ ] [ ] u n u u e n +
[ ] [ ] ( [ ] [ ]) ( [ ] [ ]) ( ) [ ] [ ] u n u u u n e u u n e u e n e u q n + +
Quantizer
Predictor
With Delay
Communication
Channel
Predictor
With Delay
[ ] u n [ ] e n [ ] e n [ ] e n
[ ] u n
[ ] u n
[ ] u n
[ ] u n
There are two components to design in a DPCM system
The predictor
The predictor output
The quantizer
A-law quantizer
-law quantizer
0 1 1 2 2
... n n S a S a S a S + + +
Wei-Yi Wei 22
DPCM (2/3)
DPCM (2/3)
0 1 1 2 2 ,
0 0 0
2 2
0 0 0 1 1 2 2
0 1 1 2 2
0 1 1
... (The predictor output)
(The predictition error)
[( ) ] [( ( ... )) ]
2 [( ( ... )) ] 0, for i=0, 1, 2, ,n
[( (
n n
n n
i i
n n i
S a S a S a S
e S S
E S S E S a S a S a S
a a
E S a S a S a S S
E S a S
+ + +
+ + +

+ + +
+
L
2 2
0 0
0 0
0 1 1 2 2 1 1 2 2
1
2
0 1 2
... )) ] 0
[( ) ] 0, for i=0, 1, 2, ,n
[ ] [ ]
[ ]
[ ... ]
[ ] [ ]
[
n n i
i
i i
ij i j
i i i n n i i i n ni
i i i ni
n
a S a S S
E S S S
E S S E S S
R E S S
R E a S S a S S a S S a R a R a R
a
a
R R R R
a
+ +

+ + + + + +
1
1
1
1
1
]
L
L
L
M
1
1 2 0 ] [ ] [ ] i i i ni i a R R R R
L
Design of Linear Predictor
Wei-Yi Wei 23
DPCM (3/3)
DPCM (3/3)
When S
0
comprises these optimized coefficients, a
i
, then the mean square
error signal is
2 2
0 0 0 0 0 0 0 0
0 0 0 0 0 0
0 0 0
2 2
0 0 0 0 0 0
00 1 01 2 02 0
2
[( ) ] [( ) ( ) ]
[( ) ] [( ) ]
But [( ) ] 0 (orthogonal principle)
[( ) ] [( ) ] [ ]
( ... )
: The varia
e
e
n n
e
E S S E S S S S S S
E S S S E S S S
E S S S
E S S S E S E S S
R a R a R a R

+ + +
00
nce of the difference signal
: The variance of the original signal R
The variance of the error signal is less than the variance of the
original signal.
Wei-Yi Wei 24
Differential Coding - JPEG (1/2)
Transform Coefficients
DC coefficient
AC coefficients
Because there is usually strong correlation between the DC coefficients of adjacent
88 blocks, the quantized DC coefficient is encoded as the difference from the DC
term of the previous block
The other 63 entries are the AC components. They are treated separately from the
DC coefficients in the entropy coding process
0 1 5 6 14 15 27 28
2 4 7 13 16 26 29 42
3 8 12 17 25 30 41 43
9 11 18 24 31 40 44 53
10 19 23 32 39 45 52 54
20 22 33 38 46 51 55 60
21 34 37 47 50 56 59 61
35 36 48 49 57 58 62 63
ZigZag Scan
[6
]
Wei-Yi Wei 25
We set DC
0
= 0.
DC of the current block DC
i
will be equal to DC
i-1
+ Diff
i
.
Therefore, in the JPEG file, the first coefficient is actually the
difference of DCs. Then the difference is encoded with Huffman coding
algorithm together with the encoding of AC coefficients
Differential Coding :
Wei-Yi Wei 26
Outline
Outline
Wei-Yi Wei 27
Quantization
The objective of quantization is to reduce the precision and to achieve
higher compression ratio
Lossy operation, which will result in loss of precision and unrecoverable
distortion
Source Coding
To achieve less average length of bits per pixel of the image.
Assigns short descriptions to the more frequent outcomes and long
descriptions to the less frequent outcomes
Entropy Coding Methods
Huffman Coding
Arithmetic Coding
Run Length Coding
Dictionary Codes
Lempel-Ziv77
Lempel-Ziv 78
Wei-Yi Wei 28
Source Coding
Source Coding
Source Encoder
Sequence of source
symbols u
i
Sequence of code symbols a
i
1 2 { } M U u u u K
1 2 { } M P p p p K
1 2 { } n A a a a K
Source alphabet Code alphabet
meaasge
( )
( )
( ) ( ) ( )
1 2
( )
i
1 1 1
, , ,
where , k=1,2, ,N and i = 1, ,M
: codeword
: length of the codeword
Average length of codeword : ( ) ( )
i
i
i i i
i
Ni
i
k
i
M M M
i i i i i
i i i
u X X X
X A
X
N
N p X N p u N p N

L
L L
Wei-Yi Wei 29
Huffman Coding (1/2)
The code construction process has a complexity of O(Nlog
2
N)
Huffman codes satisfy the prefix-condition
Uniquely decodable: no codeword is a prefix of another codeword
Huffman Coding Algorithm
(1) Order the symbols according to the probabilities
Alphabet set: S
1
, S
2
,, S
N

Probabilities: P
1
, P
2
,, P
N

The symbols are arranged so that P
1
P
2
P
N
(2) Apply a contraction process to the two symbols with the smallest
probabilities. Replace the last two symbols S
N
and S
N-1
to form a new
symbol H
N-1
that has the probabilities P
1
+P
2
.
The new set of symbols has N-1 members: S
1
, S
2
,, S
N-2
, H
N-1
(3) Repeat the step 2 until the final set has only one member.
(4) The codeword for each symbol S
i
is obtained by traversing the binary
tree from its root to the leaf node corresponding to S
i
Wei-Yi Wei 30
Codeword
length
2
2
3
3
3
Codeword X
01
10
11
000
001
1
2
3
4
5
Probabilit
y
0.25
0.25
0.2
0.15
0.15
0.25
0.25
0.2
0.45 0.55 1 0.3
0.3
0.25
0.45
0
1
1
00
01
00
01
10
11
01
10
11
000
001
Wei-Yi Wei 31
Arithmetic Coding (1/4)
Shannon-Fano-Elias Coding
We take X={1,2,,m}, p(x)>0 for all x.
Modified cumulative distribution function
Assume we round off to , which is denoted by
The codeword of symbol x has l(x) bits
Codeword is the binary value of with l(x) bits
1
( ) ( )
2
a x
F P a P x
<
+
( ) ( ) l x F x
1
] ( ) F x
( ) l x
x x P(x) P(x) F(x) F(x) in binary in binary l(x) l(x) codeword codeword
1 1 0.25 0.25 0.25 0.25 0.125 0.125 0.001 0.001 3 3 001 001
2 2 0.25 0.25 0.50 0.50 0.375 0.375 0.011 0.011 3 3 011 011
3 3 0.20 0.20 0.70 0.70 0.600 0.600 0.1001 0.1001 4 4 1001 1001
4 4 0.15 0.15 0.85 0.85 0.775 0.775 0.1100011 0.1100011 4 4 1100 1100
5 5 0.15 0.15 1.00 1.00 0.925 0.925 0.1110110 0.1110110 4 4 1110 1110
( ) F x ( ) F x
1
log 1
( ) p x
1
+
1
1
( ) F x
Wei-Yi Wei 32
Arithmetic Coding: a direct extension of Shannon-Fano-Elias coding
calculate the probability mass function p(x
n
) and the cumulative distribution
function F(x
n
) for the souece sequence x
n
Lossless compression technique
Treate multiple symbols as a single data unit
Arithmetic Coding Algorithm
Input symbol is l
Previous
low
is the lower bound for the old interval
Previous
high
is the upper bound for the old interval
Range is Previous
high
- Previous
low
Let Previous
low
= 0, Previous
high
= 1, Range = Previous
high
Previous
low
=1
WHILE (input symbol != EOF)
get input symbol l
Range = Previous
high
- Previous
low
New Previous
low
= Previous
low
+ Range* interval
low
of l
New Previous
high
= Previous
low
+ Range* interval
high
of l
END
Wei-Yi Wei 33
Symbol Symbol Probability Probability Sub-interval Sub-interval
k k 0.05 0.05 [0.00,0.05) [0.00,0.05)
l l 0.2 0.2 [0.05,0.25) [0.05,0.25)
u u 0.1 0.1 [0.20,0.35) [0.20,0.35)
w w 0.05 0.05 [0.35,0.40) [0.35,0.40)
e e 0.3 0.3 [0.40,0.70) [0.40,0.70)
r r 0.2 0.2 [0.70,0.90) [0.70,0.90)
? ? 0.2 0.2 [0.90,1.00) [0.90,1.00)
Input String : l l u u r e ?
l l u u r e ?
?
r
e
w
u
l
k
0
1.00
0.25
0.35
0.40
0.70
0.90
0.05
0
1
?
r
e
w
u
l
k
0.25
0.05
?
r
e
w
u
l
k
0.10
0.06
?
r
e
w
u
l
k
0.074
0.070
?
r
e
w
u
l
k
0.0714
0.0710
?
r
e
w
u
l
k
0.07136
0.07128
?
r
e
w
u
l
k
0.071336
0.07132
?
r
e
w
u
l
k
0.0713360
0.0713336
Codeword : 0001001001000011
0.0713348389
=2
-4
+2
-7
+2
-10
+2
-15
+2
-16
Wei-Yi Wei 34
Symbol Symbol Probability Probability Huffman codeword Huffman codeword
k k 0.05 0.05 10101 10101
l l 0.2 0.2 01 01
u u 0.1 0.1 100 100
w w 0.05 0.05 10100 10100
e e 0.3 0.3 11 11
r r 0.2 0.2 00 00
? ? 0.2 0.2 1011 1011
Input String : l l u u r e ?
Huffman Coding
Codeword : 01,01,100,100,00,11,1101
18 bits
Arithmetic coding yields better compression because it encodes a message as
a whole new symbol instead of separable symbols
Most of the computations in arithmetic coding use floating-point arithmetic.
However, most hardware can only support finite precision
While a new symbol is coded, the precision required to present the range grows
There is a potential for overflow and underflow
If the fractional values are not scaled appropriately, the error of encoding occurs
Arithmetic Coding
Codeword : 0001001001000011
16 bits
Wei-Yi Wei 35
Zero-Run-Length Coding-JPEG (1/2)
57, 45, 0, 0, 0, 0, 23, 0, -30, -16, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, ..., 0
(0,57) ; (0,45) ; (4,23) ; (1,-30) ; (0,-16) ; (2,1) ; EOB
(0,57) ; (0,45) ; (4,23) ; (1,-30) ; (0,-16) ; (2,1) ; (0,0)
(0,6,111001);(0,6,101101);(4,5,10111);(1,5,00001);(0,4,0111);(2,1,1);(0,0)
1111000 1111001 , 111000 101101 , 1111111110011000 10111 ,
11111110110 00001 , 1011 0111 , 11100 1 , 1010
The notation (L,F)
L zeros in front of the nonzero value F
EOB (End of Block)
A special coded value means that the rest elements are all zeros
If the last element of the vector is not zero, then the EOB marker will not be
added
An Example:
1.
2.
3.
4.
5.
Wei-Yi Wei 36
run/category run/category code length code length code word code word
0/0 (EOB) 0/0 (EOB) 4 4 1010 1010
15/0 (ZRL) 15/0 (ZRL) 11 11 11111111001 11111111001
0/1 0/1 2 2 00 00
... ...
0/6 0/6 7 7 1111000 1111000
... ...
0/10 0/10 16 16 1111111110000011 1111111110000011
1/1 1/1 4 4 1100 1100
1/2 1/2 5 5 11011 11011
... ...
1/10 1/10 16 16 1111111110001000 1111111110001000
2/1 2/1 5 5 11100 11100
... ...
4/5 4/5 16 16 1111111110011000 1111111110011000
... ...
15/10 15/10 16 16 1111111111111110 1111111111111110
Huffman table of Luminance AC coefficients
Wei-Yi Wei 37
Dictionary Codes
Dictionary Codes
Dictionary based data compression algorithms are based on the
idea of substituting a repeated pattern with a shorter token
Dictionary codes are compression codes that dynamically
construct their own coding and decoding tables on the fly by
looking at the data stream itself
It is not necessary for us to know the symbol probabilities
beforehand. These codes take advantage of the fact that, quite
often, certain strings of symbols are frequently repeated and
these strings can be assigned code words that represent the
entire string of symbols
Two series
Lempel-Ziv 77: LZ77, LZSS, LZBW
Lempel-Ziv 78: LZ78, LZW, LZMW
Wei-Yi Wei 38
Lempel Ziv 77 Algorithm (1/4)
Search Buffer: It contains a portion of
the recently encoded sequence.
Look-Ahead Buffer: It contains the
next portion of the sequence to be
encoded.
Once the longest match has been
found, the encoder encodes it with a
triple <Cp, Cl, Cs>
Cp :the offset or position of the longest
match from the lookahead buffer
Cl :the length of the longest matching
string
Cs :the codeword corresponding to the
symbol in the look-ahead buffer that
follows the match
LZ77 Compression Algorithm
searches the search buffer for the longest
match
If (longest match is found and all the
characters are compared)
Output <Cp, Cl, Cs>
Shift window Cl characters
ELSE
Output <0, 0, Cs>
Shift window 1 character
END
Wei-Yi Wei 39
Advantages of LZ77
Do not require to know the probabilities of the symbols beforehand
A particular class of dictionary codes, are known to asymptotically
approach to the source entropy for long messages. That is, the longer the
size of the sliding window, the better the performance of data
compression
The receiver does not require prior knowledge of the coding table
constructed by the transmitter
Disadvantage of LZ77
A straightforward implementation would require up to (N-F)*F character
comparisons per fragment produced. If the size of the sliding window
becomes very large, then the complexity of comparison is very large
Wei-Yi Wei 40
L Z 7 7 T y p e I s O l d e s t L Z 7 7 I s v x
Sliding Window of N characters
Already Encoded
N-F characters
Lookahead buffer
F characters
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 7 8
L Z 7 7 L Z 7 7
Codeword = (15,4,I) for LZ77I
I
y p e I s O l d e s t L Z 7 7 I s v x O I d e s
Already Encoded
N-F characters
Lookahead buffer
F characters
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 7 8
s s s
Codeword = (6,1,v) for sv
v
Shift 5 characters
Wei-Yi Wei 41
e I s O l d e s t L Z 7 7 I s v x O l d e s t Z
Already Encoded
N-F characters
Lookahead buffer
F characters
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 1 2 3 4 5 6 7 8
x
Codeword = (0,0,x) for x
Shift 2 characters
Wei-Yi Wei 42
The LZ78 algorithm parsed a
string into phrases, where each
phrase is the shortest phrase
not seen so far
The multi-character patterns
are of the form: C
0
C
1
. . . C
n-1
C
n
.
The prefix of a pattern consists
of all the pattern characters
except the last: C
0
C
1
. . . C
n-1
This algorithm can be viewed
as building a dictionary in the
form of a tree, where the nodes
corresponding to phrases seen
so far
Step 1: In the parsing context, search
the longest previously parsed phrase P
matching the next encoded substring.
Step 2: Identify this phrase P by its
index L in a list of phrases, and place
the index on the code string. Go to the
innovative context.
Step 3: In the innovative context,
concatenate next character C to the code
string, and form a new parsed phrase P
C.
Step 4: Add phrase P C to the end of
the list of parsed phrases as (L,C)
Return to the Step 1.
Wei-Yi Wei 43
Advantages
Asymptotically, the average length of the codeword per source symbol is
not greater than the entropy rate of the information source
The encoder does not know the probabilities of the source symbol
beforehand
Disadvantage
If the size of the input goes to infinity, most texts are considerably shorter
than the entropy of the source. However, due to the limitation of memory
in modern computer, the resource of memory would be exhausted before
compression become optimal. This is the bottleneck of LZ78 needs to be
overcame
Wei-Yi Wei 44
Input String: ABBABBABBBAABABAA
0
1 2
3 4 5
6 7 8
A B
B A B
A A A
Parsed String: A, B, BA, BB, AB, BBA, ABA, BAA
Output Codes: (0,A), (0,B), (2,A), (2,B), (1,B), (4,A), (5,A), (3,A)
1 A (0,A)
2 B (0,B)
3 BA (2,A)
4 BB (2,B)
5 AB (1,B)
6 BBA (4,A)
7 ABA (5,A)
8 BAA (3,A)
Index Phrases (L,C)
Wei-Yi Wei 45
Outline
Outline
Wei-Yi Wei 46
JPEG
JPEG
The JPEG Encoder
The JPEG Decoder
Wei-Yi Wei 47
Quantization in JPEG
Quantization in JPEG
( , )
( , )
( , )
Quantization
F u v
F u v round
Q u v
_

,
16 11 10 16 24 40 51 61
12 12 14 19 26 58 60 55
14 13 16 24 40 57 69 56
14 17 22 29 51 87 80 62
18 22 37 56 68 109 103 77
24 35 55 64 81 104 113 92
49 64 78 87 103 121 120 101
72 92 95 98 112 100 103 99
Y Q
_

,
17 18 24 47 99 99 99 99
18 21 26 66 99 99 99 99
24 26 56 99 99 99 99 99
47 66 99 99 99 99 99 99
99 99 99 99 99 99 99 99
99 99 99 99 99 99 99 99
99 99 99 99 99 99 99 99
99 99 99 99 99 99 99 99
C Q
_

,
( , ) ( , ) ( , ) deQ Quantization F u v F u v Q u v
Quantization
Dequantization
Quantization is the step where we actually throw away data.
Luminance and Chrominance Quantization Table
lower numbers in the upper left direction
large numbers in the lower right direction
The performance is close to the optimal condition
Wei-Yi Wei 48
JPEG 2000
JPEG 2000
The JPEG 2000 Encoder
The JPEG 2000 Decoder
Wei-Yi Wei 49
Quantization in JPEG 2000
Quantization in JPEG 2000
Quantization coefficients
a
b
(u,v) : the wavelet coefficients of subband b
Quantization step size
R
b
: the nominal dynamic range of subband b
b
: number of bits alloted to the exponent of the subbands coefficients
b
: number of bits allotted to the mantissa of the subbands coefficients
Reversible wavelets
Uniform deadzone scalar quantization with a step size of b =1 must be
used
Irreversible wavelets
The step size is specified in terms of an exponent
b
, 0
b
2
5
, and a
mantissa
b
, 0
b
2
11
[ ]
| ( , ) |
( , ) ( , )
b
b u
b
a u v
q u v sign a u v floor
1

1
]
11
2 1
2
R
b
b b
b
_
+

,
Wei-Yi Wei 50
Bitplane Scanning
Bitplane Scanning
The decimal DWT coefficients can be converted into signed binary format, so
the DWT coefficients are decomposed into many 1-bit planes.
In one 1-bit-plane
Significant
A bit is called significant after the first bit 1 is met from MSB to LSB
Insignificant
The bits 0 before the first bit 1 are insignificant
Wei-Yi Wei 51
Scanning Sequence
Scanning Sequence
The scanning order of the bit-plane
Sample
Each element of the bit-plane is called a sample
Column stripe
Four vertical samples can form one column stripe
Full stripe
The 16 column stripes in the horizontal direction can form a full stripe
Wei-Yi Wei 52
The Context Window
The Context Window
Current sample
The curr is the sample which is to be coded
The other 8 samples are its neighbor samples
The diagonal sample
The samples in the diagonal direction
The vertical samples
The samples in the vertical direction
The horizontal samples
The samples in the horizontal direction
Wei-Yi Wei 53
Arithmetic Coder in JPEG 2000
Arithmetic Coder in JPEG 2000
The decision and context data generated from context formation is coded in
the arithmetic encoder
The arithmetic encoder used by JPEG 2000 standard is a binary coder
More Possible Symbol (MPS): If the value of input is 1
Less Possible Symbol (LPS): If the value of input is 0
Wei-Yi Wei 54
Rate Distortion Optimization
Rate Distortion Optimization
For meeting a target bit-rate or transmission time, the packaging
process imposes a particular organization of coding pass data in
the output code-stream
The rate-control assures that the desired number of bytes used by
the code-stream while assuring the highest image quality
possible
Wei-Yi Wei 55
Both the JPEG and JPEG 2000 image compression standard can achieve great
compression ratio, however, both of them do not take advantage of the local
characteristics of the given image effectively
Instead of taking the whole image as an object and utilizing transform coding,
quantization, and entropy coding to encode this object, the SAIC algorithm segments
the whole image into several objects, and each object has its own local characteristic
and color
Because of the high correlation of the color values in each image segment, the SAIC
can achieve better compression ratio and quality than conventional image compression
algorithm
Wei-Yi Wei 56
Reference
Reference
[1] R. C. Gonzalea and R. E. Woods, "Digital Image Processing", 2nd Ed., Prentice
Hall, 2004.
[2] Liu Chien-Chih, Hang Hsueh-Ming, "Acceleration and Implementation of JPEG
2000 Encoder on TI DSP platform" Image Processing, 2007. ICIP 2007. IEEE
International Conference on, Vo1. 3, pp. III-329-339, 2005.
[3] ISO/IEC 15444-1:2000(E), "Information technology-JPEG 2000 image coding
system-Part 1: Core coding system", 2000.
[4] Jian-Jiun Ding and Jiun-De Huang, "Image Compression by Segmentation and
Boundary Description", Masters Thesis, National Taiwan University, Taipei, 2007.
[5] Jian-Jiun Ding and Tzu-Heng Lee, "Shape-Adaptive Image Compression",
Masters Thesis, National Taiwan University, Taipei, 2008.
[6] G. K. Wallace, "The JPEG Still Picture Compression Standard", Communications
of the ACM, Vol. 34, Issue 4, pp.30-44, 1991.
[7] JPEG 2000 ( ) IC
2003.8 .
[8]
2004.
[9] Subhasis Saha, "Image Compression - from DCT to Wavelets : A Review",
available in http://www.acm.org/crossroads/xrds6-3/sahaimgcoding.html
Wei-Yi Wei 57
Thank You
Q & A
Wei-Yi Wei 58
Question 1
Question 1
In Slide DPCM(3), whats the orthogonal principle?
0 0 0
0 0
0 0 0
0 0 1 1 2 2
1 0 0 1 2 0 0 2 0 0
Orthogonal principle : [( ) ] 0
Proof:
Given [( ) ] 0
[( ) ]
[( )( ... )]
[( ) ] [( ) ] [( ) ]
0 0 0
0
i
n n
n n
E S S S
E S S S
E S S S
E S S a S a S a S
a E S S S a E S S S a E S S S

+ + +
+ +
+ + +
L
L

An Introduction To Image Compression

Uploaded by

Copyright:

Available Formats

An Introduction To Image Compression

Uploaded by

Document Information

Original Description:

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

An Introduction To Image Compression

Uploaded by

Copyright:

Available Formats

1

NTU, GICE, MD531, DISP Lab An Introduction to Image Compression

Why DCT is more appropriate for image compression than DFT?

... (The predictor output)

(The predictition error)

But [( ) ] 0 (orthogonal principle)

You might also like