Basic Image Compression

Concepts
Presenter：
Guan-Chen Pan
Research Advisor：
Jian-Jiun Ding , Ph. D.
Assistant professor

Digital Image and Signal Processing Lab

Graduate Institute of Communication Engineering
National Taiwan University

1
Outlines
 Introductions
 Basic concept of image compression
 Proposed method for arbitrary-shape
image segment compression
 Improvement of the boundary region by
morphology
 JPEG2000
 Triangular and trapezoid regions and
modified JPEG image compression

2
 Introduction
 Lossless or lossy(widely used)
B
G
R Chrominance
YCbCr Color 8×8
Downsampling
Image Transform FDCT
(4:2:2 or 4:2:0)

Zigzag &
Huffman
Run Length
Encoding
Quantizer Coding Bit-stream
Differential Huffman
Quantization Coding Encoding
Table

3
 Y   0.299 0.587 0.114   R  0 

YCbCr Cb    0.169 0.334 0.500  G   128

      
Cr   0.500 0.419 0.081  B  128

 Y：the luminance of the image which

represents the brightness
 Cb：the chrominance of the image which
represents the difference between
the gray and blue
 Cr：the chrominance of the image which
represents the difference between
the gray and red

4
Chrominance Subsampling
 The name of the format is not always
related to the subsampling ratio.
(a) 4 : 4 : 4 (b) 4 : 2 : 2 (c) 4 : 2 : 0 (d) 4 : 1 : 1
W W W W

H Y H Y H Y H Y

W W/2 W/2 W/4

H/2 Cb
H Cb H Cb H C
b

W W/2 W/2 W/4

H/2 Cr
H Cr H Cr H Cr

5
 Compression ratio (CR)：
n1
CR = ,
n2

where n1 is the data quantity of original image, and n2 is the

compressed data quantity

 Root mean square error (RMSE)：

σ𝑊−1 𝐻−1
𝑥=0 σ𝑦=0 [𝑓 𝑥, 𝑦 − 𝑓′(𝑥, 𝑦)]
2

R𝑀𝑆𝐸 = ,
𝑊𝐻

where H and W are the height and the width of the images
respectively

6
 Peak-to-signal ratio (PSNR)：
255
P𝑆𝑁𝑅 = 20 log10
𝑀𝑆𝐸

 We must keep in mind that the two

measurements are not related to the real
quality of the image.

7
Reduce the Correlation between
Pixels

 Transform coding
1. Coordinate rotation
2. Karhunen-Loeve transform
3. Discrete cosine transform
4. Discrete wavelet transform
 Predictive coding

8
Coordinate rotation
 Draw a line that has the mean square
error with all data for 𝐘 = 𝐀𝐗
Weight Transform
Original sequence X=(x0,x1) 200

180
Height Weight
160
65 170
140
75 188
120
60 150

x1
70 170 100

56 130 80

80 203 60

  68
68 160 40

50 110 20
40 80
0
0 10 20 30 40 50 60 70 80 90
x0
Height

9
 New Height y0 New Width y1
181.971 3.416
203.406 0.887
161.554 0.560
183.844 -1.220
141.512 -3.223
206.133 -2.999
173.823 -3.111
120.721 -5.152
89.159 -7.112

 do the inverse transform to get the data

and reduce the correlation
Discard the value of weight Inverse Transform
Height Weight Height Weight

181.97068384838812 0 68.16741797800859 168.72028006870275

202.40605916474945 0 75.82264431044632 187.66763012404562

161.55397378997284 0 60.51918377426525 149.79023613916877

183.8437168154677 0 68.86906847716196 170.45692599485025

141.51187032497342 0 XA Y-1

53.01127967035257 131.20752139486424

206.13345984096256 0 77.21895318005868 191.12361585053173

173.822665082968 0 65.11511642520563 161.165568622698

120.72055367314222 0 45.222715366778566 111.93014828010075

89.15897210197947 0 33.399538811586865 82.66675942272599

10
Karhunen-Loeve transform(KLT)

 The KLT is the optimal transform coding.

 We represent the autocorrelation matrix
of the output vector 𝐘=𝐀𝐗 as 𝐂𝐲
𝐂𝐲 = E[ 𝐘 − E 𝐘 𝐘 − E 𝐘 T ]
we can substitute 𝐘 with 𝐀𝐗, and assume E 𝐗 =0
T
𝐂𝐲 = E 𝐀𝐗 − E 𝐀𝐗 𝐀𝐗 − E 𝐀𝐗
⇒ 𝐂𝐲 = E[ 𝐀𝐗 𝐀𝐗 T ]
⇒ 𝐂𝐲 = 𝐀𝐂𝐗 𝐀T
11
 𝛌0 ⋯ 0
 𝐂𝐲 = ⋮ ⋱ ⋮
0 ⋯ 𝛌𝑁−1
 We can achieve optimal decorrelation
through diagonalization, which means that
𝐀 is composed of the eigenvectors of the
autocorrelation matrix.
 But it takes large amount of computation
to find eigenvectors and all the
eigenvectors are need to be stored.

12
Discrete cosine transform
 The DCT is an approximation of the KLT
and more widely used in image and video
compression.
 The DCT can concentrate more energy
in the low frequency bands than the DFT.

13
Discrete wavelet transform
 Wavelet transform is very similar to the
conventional Fourier transform, but it is
based on small waves, called wavelet,
which is composed of time varying and
limited duration waves.
 We use 2-D discrete wavelet transform in
image compression.

14
 Rows
Columns h ( m) ↓ 2 WD ( j , m, n)
h (n) ↓ 2
h ( m) ↓ 2 WV ( j , m, n)
W ( j  1, m, n)
h ( m) ↓ 2 WH ( j , m, n)
h ( n) ↓ 2
h ( m) ↓ 2 W ( j , m, n)

LL2 HL2
H
W ( j , m, n) W ( j , m, n)
 HL1
LH2 HH2
W ( j  1, m, n)

WV ( j , m, n) WD ( j , m, n) LH1 HH1

 We can see that 𝑊𝜙 (𝑗, 𝑚, 𝑛൯ is very

similar to the original image , so we can
use it to achieve image compression.
15
Predictive Coding
 Predictive coding means that we transmit
only the difference between the current
pixel and the previous pixel.
 The difference may be close to zero.
 However, the predictive coding algorithm
is more widely used in video.
 EX. Delta modulation (DM), Adaptive DM.
DPCM ,Adaptive DPCM (ADPCM)

16
Quantization

 Each DCT coefficient F(u,v) is divided by

the corresponding quantization matrix
Q(u,v) and rounded to the nearest integer.
𝐹(𝑢, 𝑣)
𝐹𝑞 𝑢, 𝑣 = 𝑅𝑜𝑢𝑛𝑑( )
𝑄(𝑢, 𝑣)

17
 Luminance quantization matrix
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

 Chrominance quantization matrix

 Removes the high frequencies

18
Entropy Coding Algorithms
1. Huffman Coding
◦ Difference Coding (DC)
◦ Zero Run Length Coding (AC)
2. Arithmetic Coding
3. Golomb Coding

19
Huffman Coding
 Huffman coding is the most popular
technique for removing coding
redundancy.
◦ Unique prefix property
◦ Instantaneous decoding property
◦ Optimality
 JPEG(fixed, not optimal)

20
Symbol Probability 1 2 3 4 Code
a2 0.4 0.4 0.4 0.4 0.6 1
a6 0.3 0.3 0.3 0.3 0.4 00
a1 0.1 0.1 0.2 0.3 011
a4 0.1 0.1 0.1 0100
a3 0.06 0.1 01010
a5 0.04 01011

21
Difference Coding
 For DC coefficients
 The DC coefficients is very close to its
neighbors and usually have much larger
value than AC coefficients.

Diffi = DCi -DCi−1

22
Zero Run Length Coding
 Encode each value which is not 0, than
add the number of consecutive zeroes in
front of it
 EOB (End of Block) = (0,0)
 Only 4-bit value
 [57,45,0,0,0,0,23,0,-30,-16,0,……,0]
⇒[(0,57)(0,45)(4,23)(1,-30)(0,16)EOB]
 “Eighteen zeroes, 3” ⇒(15,0) ; (2,3)
where (15,0) is 16 consecutive zeroes

23
 run/category code length code word
0/0 (EOB) 4 1010
Category Values Bits for the value
15/0 (ZRL) 11 11111111001
1 -1,1 0,1
0/1 2 00
2 -3,-2,2,3 00,01,10,11
...
3 -7,-6,-5,-4,4,5,6,7 000,001,010,011,100,101,110,111
0/6 7 1111000
4 -15,...,-8,8,...,15 0000,...,0111,1000,...,1111
...
5 -31,...,-16,16,...31 00000,...,01111,10000,...,11111
0/10 16 1111111110000011
6 -63,...,-32,32,...63 000000,...,011111,100000,...,111111
1/1 4 1100
7 -127,...,-64,64,...,127 0000000,...,0111111,1000000,...,1111111
1/2 5 11011
8 -255,..,-128,128,..,255 ...
...
9 -511,..,-256,256,..,511 ...
1/10 16 1111111110001000
10 -1023,..,-512,512,..,1023 ...
2/1 5 11100
11 -2047,...,-1024,1024,...,2047 ...
...
4/5 16 1111111110011000
...
15/10 16 1111111111111110

24
Arithmetic Coding
 Arithmetic coding is another coding
method widely used in image and video
compression, and its performance is
better than Huffman coding.
 We treat the whole input data as a single
symbol and find the corresponding
codeword for it.
 Huffman, probability very close to 1.0,
1
log 2
𝑝𝑖

25
 Arithmetic Coding Algorithm
Input symbol is S
Previouslow is the lower bound for the old interval
Previoushigh is the upper bound for the old interval
Range = Previoushigh - Previouslow
Let
Previouslow= 0, Previoushigh = 1
Range = Previoushigh – Previouslow =1
WHILE (input symbol != EOF)
get input symbol S
Range = Previoushigh - Previouslow
New Previouslow = Previouslow + Range × intervallow of S
New Previoushigh = Previouslow + Range × intervalhigh of S
END

26
 Symbol Probability Sub-interval

k 0.05 [0.00,0.05)
l 0.2 [0.05,0.25)
u 0.1 [0.25,0.35)
w 0.05 [0.35,0.40)
e 0.3 [0.40,0.70)

r 0.2 [0.70,0.90)
? 0.1 [0.90,1.00)

 First l Input String : l l u u r e ? 0.0713348389

1. 𝑃𝐿 = 0 Codeword : 0001001001000011 =2-4+2-7+2-10+2-15+2-16
2. 𝑃𝐻 = 1
l l u u r e ?
3. 𝑅=1
1 0.25 0.10 0.074 0.0714 0.07136 0.071336 0.0713360
4. 𝑁𝑃𝐿 = 0 + 1 × 1.00
0.05 = 0.05 0.90
? ? ? ? ? ? ? ?

5. 𝑁𝑃𝐻 = 0 + 1 × r r r r r r r r
0.25 = 0.25 0.70

e e e e e e e e
 Second l 0.40
w w w w w w w w
1. 𝑃𝐿 = 0.05 0.35
u u u u u u u u
2. 𝑃𝐻 = 0.25 0.25
l l l l l l l l
3. 𝑅 = 0.2
0.05
4. 𝑁𝑃𝐿 = 0.05 + k k k k k k k k
0
0.2 × 0.05 = 0.06 0 0.05 0.06 0.070 0.0710 0.07128 0.07132 0.0713336
5. 𝑁𝑃𝐻 = 0.05 +
0.2 × 0.25 = 0.1
27
 Symbol Probability Sub-interval

k 0.05 [0.00,0.05)
l 0.2 [0.05,0.25)

w
u 0.1
0.05
[0.20,0.35)
[0.35,0.40)  0.071334 ⇒ L
e 0.3 [0.40,0.70)
r 0.2 [0.70,0.90)
? 0.1 [0.90,1.00)

 For interval 0.05~0.25

Symbol Probability Sub-interval

k 0.05 [0.05,0.06)
l 0.2 [0.06,0.1)

0.071334 ⇒ L
u 0.1 [0.1,0.12)
w
e
0.05
0.3
[0.12,0.13)
[0.13,0.19)

r 0.2 [0.19,0.23)
? 0.1 [0.23,0.25)

28
Golomb Coding
 Golomb coding is a special case of the
Huffman coding.
 Optimal for the data with a geometric
distribution.
 Prob y = a = (1 − p)pa
 No table

29
 log(𝑝+1)
1. First, determine m from p , 𝑚 = −
log(𝑝)
2. a = q×m + r
3. Convert q into the prefix. The prefix is
composed of q “1” bits followed by a “0” bit.
4. Convert r into the suffix using the binary code.
Threshold parameter (m) = 2^ log 2 𝑚  m.
◦ If r < (m), the length of the suffix is log 2 𝑚 bits.
◦ If r ≥ (m), we update r into r +(m) and encode it
into a log 2 𝑚 -length suffix.

30
 Example
◦ p = 0.93, m = 10, a = 19
 q = 1, r = 9
 Prefix = “10”
 Threshold parameter (m) = 2^ log 2 𝑚  m = 6
 r >threshold
⇒ r = r + threshold = 9 + 6 = 15
⇒ encode 15 into a log 2 𝑚 -length suffix,
log 2 𝑚 =4
⇒ Suffix = “1111”
 Code = “101111”

31
 Decode “101111”
Encoding of quotient part Encoding of remainder part
q output bits r offset binary output bits
0 0 0 0 0000 000
1 10 1 1 0001 001
2 110 2 2 0010 010
3 1110 3 3 0011 011
4 11110 4 4 0100 100
5 111110 5 5 0101 101
6 1111110 6 12 1100 1100
: : 7 13 1101 1101
N <N repetitions of 1>0 8 14 1110 1110
9 15 1111 1111

 q = 1, r = 9
⇒ a = 10*1+9 = 19
32
 However, Golomb coding can just achieve
optimal coding efficiency when the data is
geometrically distributed.
 To solve this problem, there is an Adaptive
Golomb Code.
 Prob y = a = (1 − p)pa
⇒ Prob y = a = (1 − p(x))p(x)a
Without Flexibility
codeword and
table adaptation
Huffman NO GOOD
Golomb YES MIDDLE
Adaptive YES GOOD
Golomb
33
Proposed Method for Arbitrary-
Shape Image Segment Compression
 An arbitrary-shape image segment f and
its shape matrix.
75 96 0 0 0 0 1 1 0 0
105 98 99 101 73 85 66 60 1 1 1 1 1 1 1 1
100 97 89 94 87 64 55 0 1 1 1 1 1 1 1
84 94 90 81 71 66 0 0 1 1 1 1 1 1
93 86 94 81 70 0 0 1 1 1 1 1 0
86 86 81 72 0 0 0 1 1 1 1 0
98 97 78 0 0 0 1 1 1 0 0
105 104 0 0 0 1 1 0 0 0

34
 Standard 8x8 DCT bases with the shape
of f
0 1 2 3 4 5 6 7

35
 The 37 arbitrary-shape orthonormal
DCT bases by Gram-Schmidt process
1 2 3 4 5 6 7 8 9 10

11 12 13 14 15 16 17 18 19 20

21 22 23 24 25 26 27 28 29 30

31 32 33 34 35 36 37

W 1 H 1
F (k )   f ( x, y) 'x, y (k ) for k  1,2,..., M
x 0 y 0

36
Quantization
 Q(k )  Qa k  Qc , for k  1,2,..., M

 F (k ) 
 Fq (k )  Round   , where k  1,2,..., M
 Q(k ) 
45

40

35

30

25

20

15

10

5
0 100 200 300 400 500 600

37
Improvement of the Boundary
Region by Morphology

38
JPEG2000
 JPEG 2000 is a new standard and it can
achieve better performance in image
compression.
 Advantages
◦ Efficient lossy and lossless compression
◦ Superior image quality
◦ Additional features such as spatial scalability
and region of interest.
 Complexity
39
  JPEG 2000 encoder
Rate
Control

Original Coded
Image Forward Image
Forward Tier-1 Tier-2
Component Quantization
2D DWT Encoder Encoder
Transform

 Embedded Block Coding with Optimized

Truncation(EBCOT) : Tier-1+Tier-2
 JPEG 2000 decoder

Coded Reconstructed
Image Inverse Image
Tier-2 Tier-1 Inverse
Dequantization Component
Decoder Decoder 2D DWT
Transform

40
41
Irreversible component transform
(ICT)
 Irreversible and real-to-real
V0  x, y    0.299 0.587 0.114  U 0  x, y  
    U x , y 
V
 1  x , y  
  0.5 0.41869 0.08131 1 
V2  x, y    0.16875 0.33126 0.5  U 2  x, y  

 𝑈0 , 𝑈1 and 𝑈2 are just like 𝑅, 𝐺 and

𝐵 ,respectively.
 𝑉0 , 𝑉1 and 𝑉2 are just like 𝑌, 𝐶𝑏 and
𝐶𝑟 ,respectively. Y   0.299 0.587 0.114   R  0 
Cb    0.169 0.334 0.500  G   128
      
Cr   0.500 0.419 0.081  B  128
42
Reversible component transform
(RCT)
 Reversible and integer-to-integer
U 0  x, y   2U1  x, y   U 2  x, y  
V0  x, y    
 4 

V1  x, y   U 2  x, y   U1  x, y 

V2  x, y   U0  x, y   U1  x, y 

43
 Tiling DWT in each tile

Image
Component

44
 Rows
Columns h ( m) ↓ 2 WD ( j , m, n)
h (n) ↓ 2
h ( m) ↓ 2 WV ( j , m, n)
W ( j  1, m, n)
h ( m) ↓ 2 WH ( j , m, n)
h ( n) ↓ 2
h ( m) ↓ 2 W ( j , m, n)

 Irreversible , Daubechies 9/7 filter

Analysis Filter Coefficients Synthesis Filter Coefficients
n Lowpass Filter Highpass Filter Lowpass Filter Highpass Filter
0 0.602949018236 1.115087052456 1.115087052456 0.6029490182363
±
0.266864118442 -0.059127176311 0.591271763114 -0.2668641184428
1
±
-0.078223266528 -0.057543526228 -0.057543526228 -0.0782232665289
2
±
-0.0168641184428 0.091271763114 -0.0912717631142 0.0168641184428
3
±
0.026748757410 0.0267487574108
4

45
46
 Tier-1 Encoder
Quantized
DWT Context
coefficients
Fractional
Bit-plane Arithmetic
Bit-plane
Conversion Decision Coding
Coding

Tier-1 Encoder

 Each Fractional Bit-plane coding will

generate the Context (CX) and the Decision
(D), which are used for arithmetic coding.
◦ zero coding
◦ sign coding
◦ magnitude refinement coding
◦ run length coding
47
Bit-plane Conversion
 Converts the quantized wavelet
coefficients into several bit-planes
 First bit-plane is the sign plane
 The other planes are the magnitude plane,
from MSB to LSB N
Sign
M
MSB
0 MSB
Insignificant coding
0 order
Significant
First 1 appear 1
The bit to be coded
0
LSB
0

1 LSB

48
17 22 33 48 64 80 96 112
22 28 38 52 67 81 96 112
33 38 48 62 75 86 100 116  17 = 000100012
48 52 62 70 83 96 110 125
64 67 75 83 96 108 118 132  160 = 101000002
80 81 86 96 108 117 128 142
96 96 100 110 118 128 140 150
112 112 116 125 132 142 150 160

49
Stripe and Scan Order

50
 d v d
Zero Coding h D h
d v d
 D : current encode data, binary : 0 or 1
 h :0~2 v :0~2 d :0~4

51
 v

Sign Coding h D h
v
 h, v: neighborhood sign
status
◦ -1: one or both negative
◦ 0: both insignificant or
both significant but
opposite sign
◦ -1: one or both positive
 D : ෡𝑋 ⊗ 𝑋,
⊗ = XOR

52
Magnitude Refinement Coding
 σ′[x,y] is initialized to 0, and it will
become 1 after the first time of the
magnitude refinement coding is met at
[x,y]

53
Run-Length Coding
 For four zeros : (CX,D) is (0,0)
 Else is (0,1), and use 2 uniform(CX=18)
to record the 1’s position
◦ (0110)
◦ The first nonzero position is (01)2
⇒(0,1), (18,0), (18,1)

54
 D
(0,1) Arithmetic Compressed
encoder data
CX
(total 19)

55
Why Called Fractional？

56
Tier-2 Encoder
 Rate/Distortion optimized truncation

57
Triangular and Trapezoid Regions and
Modified JPEG Image Compression
 Divide an image into 3 parts:
1. Lower frequency regions
2. Traditional image blocks and
3. The arbitrarily-shaped image blocks

Trapezoid and triangular

regions
8X8 SADCT image blocks

Traditional 8X8 image

blocks

58
1 1 1 1 1 1 1 1 1 0  1 sections
0 1 1 1 1 1 1 1 1 1 1 sections
0 1 1 1 1 1 1 1 1 0

 1 sections
 Zone 1
0 1 1 1 1 1 1 1 0 0  1 sections
0 0 1 1 0 1 1 1 1 0  2 sections
0 0 1 0 0 1 1 1 0 0  2 sections  Zone 2
0 0 0 0 0 0 1 1 0 0  1 sections
0 0 0 0 0 0 1 1 0 0  1 sections  Zone 3

Zone 1 1 zone

Zone 2-1 Zone 2-2 2 zones

Zone 3
1 zone

59
 α -distance

 α-distance < threshold

60
 Corner too close

 Trapezoid inside the zone

61
(M-1)th row
(M-2)th row
.
.
.
.
N = 10
.
.

1st row
0th row

 N = K(m) + K(M-1-m)

62
(a) 1. Construct the
rectangular region and
m = M-1
m = M-2
.
.
.
.
.
.
obtain the
m= 1
m= 0
orthonormal DCT
n=0 1 2
basis Cp,q [m, n]
(b)

Region A Region B
2. Select the DCT basis
Cp,q [m, n] that
Rotation by 180 ∘

Region B
Region A
satisfies p+q=even
Rectangular Region

3. Bp,q [m, n]=

2Cp,q [m, n]

63
Reference:
1. J.D Huang "Image Compression by Segmentation and Boundary Description, " 2008.
2. G. Roberts, "Machine Perception of Three-Dimensional Solids," in Optical and Electro-
Optical Information Processing, J. T. T. e. al., Ed. Cambridge, MA: MIT Press, 1965, pp. 159-
197.
3. J. Canny, "A Computational Approach to Edge Detection," IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 8, pp. 679-698, Nov. 1986.
4. D. Comaniciu and P. Meer, "Mean Shift: A Robust Approach toward Feature Space
Analysis, " IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, pp. 603-619, 2002.
5. J.J Ding, P.Y Lin, S.C Pei, and Y.H Wang,"The Two-Dimensional Orthogonal DCT
Expansion in Triangular and Trapezoid Regions and Modified JPEG Image Compression,
",VCIP2010
6. J.J Ding, S.C Pei, W.Y Wei, H.H Chen, and T.H Lee, "Adaptive Golomb Code for Joint
Geometrically Distributed Data and Its Application in Image Coding", APSIPA 2010
7. W.Y Wei, "Image Compression", available in http://disp.ee.ntu.edu.tw/tutorial.php
8. K. R. Rao and P.Yip, Discrete Cosine Transform, Algorithms, Advantage, Applications, New York:
Academic, 1990.
9. S.S. Agaian, Hadamard Matrices and Their Applications, New York, Springer-Verlag, 1985.
10. H. F. Harmuth, Transmission of information by orthogonal functions, Springer, New York,
1970.

64
11. R. Koenen, Editor, “Overview of the MPEG-4 Standard,” ISO/IEC JTC/SC29/WG21,
MPEG-99-N2925, March 1999, Seoul, South Korea.
12. T. Sikora, “MPEG-4 very low bit rate video,” IEEE International Symposium on Circuits and
Systems, ISCAS ’97, vol. 2, pp. 1440-1443, 1997.
13. T. Sikora and B. Makai, “Shape-adaptive DCT for generic coding of video,” IEEE Trans.
Circuits Syst.Video Technol., vol. 5, pp. 59-62, Feb. 1995.
14. W.K. Ng and Z. Lin, “A New Shape-Adaptive DCT for Coding of Arbitrarily Shaped
Image Segments,” ICASSP, vol. 4, pp. 2115-2118, 2000.
15. S. C. Pei, J. J. Ding, P.Y. Lin and T. H. H. Lee, “Two-dimensional orthogonal DCT expansion
in triangular and trapezoid regions,” Computer Vision, Graphics, and Image Processing, Sitou,
Taiwan, Aug. 2009.
16. D. A. Huffman, "A method for the construction of minimum-redundancy codes,"
Proceedings of the IRE, vol. 40, no. 9, pp. 1098-1101, 1952.
17. S. W. Golomb, "Run length encodings," IEEE Trans. Inf.Theory, vol. 12, pp. 399-401, 1966.
18. R. Gallager and D.V.Voorhis, "Optimal source codes for geometrically distributed
integer alphabets," IEEE Trans. Information Theory, vol. 21, pp. 228–230, March 1975.
19. R. F. Rice, "Some practical universal noiseless coding techniques–part I," Tech. Rep. JPL-
79-22, Jet Propulsion Laboratory, Pasadena, CA, March 1979.
20. G. Seroussi and M. J. Weinberger, "On adaptive strategies for an extended family of
Golomb-type codes," Proc. DCC’97, pp. 131-140, 1997.
21. C. J. Lian “JPEG2000 “, DSP/IC design lab, GIEE, ntu

65