1. Chapter 7
漸進式影像傳輸

2. Outlines Introduction Bit-Plane Method (BPM)
Improved Bit-Plane Method (IBPM)
Dynamic Binary Threshold Method (DBTM)
Side Match Method (SMM)
Selective Progressive Image Transmission (SPITM)
Fast Reconstruction Method (FRM)
TSVQ Progressive Transmission Tree
Side-Match Reconstruction Method Using TSVQ (SMTSVQ)

3. 7.1 Introduction Progressive image transmission (PIT)
Transmit an image in many phases
Important feature is transmitted first
Approximate image can be reconstructed rapidly
Image quality is refined gradually

4. 7.1 Introduction (cont.) Why PIT? Goal: a new PIT method Low bandwidth
Tele-browsing image databases
To have at least browse quality image data ASAP
Goal: a new PIT method
Improve image quality of the bit-plane method
Low transmission and computational costs

5. 7.2 Bit-plane method (BPM)
Plane 7 (least significant bit)
Plane 0 (most significant bit)
Bit planes of an image (256 gray levels)

6. 7.2 BPM (cont.) A simple and intuitive PIT method [kunt 1978]
Each phase: 1 bit/pixel is sent
Phase i: transmit the i-th MSB of each pixel value
High quality picture: about 5 to 6 bits/pixel

7. Example 1: BPM How to recover the original image? Phase 0: 49 69 94
163 57
143 72
182
203
250 53
117
200
199
Phase 0:
How to recover the original image? 1 64
192

8. Example 1: BPM (cont.) Reason: Phase 1 0 0000 0000 1 0000 0001
… … 1 64
Receiver gets 00 01 10 11
192

9. Example 1: BPM (cont.) Reason: 32 How to recover the original image?… … …
… … … … … … … … … … 32
How to recover the original image? 32 96
160
224 96
160
224

10. 7.3 Improved Bit-Plane Method (IBPM)
Phase 0:
How to recover the original image? 1 71
193 傳送
與 {71, 193}
Reason:
0: {49, 69, 94, 57, 57, 72, 53, 117}  71
1: {163, 143, 182, 203, 250, 200, 199, 200}  193

11. How to recover the original image?
Phase 1
Receiver gets 1 00 01 10 11 傳送
與 {54, 163}
How to recover the original image? 54 88
163
210
Reason:
00: {49, 57, 57, 53}  54
01: {69, 94, 72, 117}  88
10: {163, 143, 182}  163
11: {203, 250, 200, 199, 200}  210

12. Compressions BPM Original Image IBPM Better 49 69 94 163 57 143 72 182
203
250 53
117
200
199 32 96
160 21
224
IBPM 54 88
163
210
200
Better

13. Experimental Results Experimental data
Test images: “Baboon”, “Boat”, “Girl”, “Lena”, and “Toys”
Image size: 256*256 pixels
Image color: 256 gray levels

14. Experimental Results (cont.)

15. Experimental Results (cont.)

16. Experimental Results (cont.)

17. IBMP is a new PIT method of spatial domain Simple and efficiency
Easy to implement and low computational complexity
High image quality
Improve the image quality of the BPM effectively for each phase

18. 7.4 Dynamic binary threshold method (DBTM)
Original image 50 69
134
139 54
200
235
179
168
210 80 90
Initial phase:
>= 128: (0,2), (0,3), (1,1), (1,2),
(1,3), (2,0), (2,1), (2,2),
(2,3), (3,0), (3,1)
< 128: (0,0), (0,1), (1,0), (3,2),
(3,3)

19. Phase 1: 算S-list 之mean μS=68 算L-list 之mean μL=165 分裂 S-list 與L-list 為
可以看成:
S: (63)(0,0), (0,1), (1,0), (3,2), (3,3)
L: (191)(0,2), (0,3), (1,1), (1,2), (1,3), (2,0), (2,1),
(2,2), (2,3), (3,0), (3,1)
傳給 viewer
Phase 1:
算S-list 之mean μS=68
算L-list 之mean μL=165
分裂 S-list 與L-list 為
SS: (63)(0,0), (1,0)
SL: (68)(0,1), (3,2), (3,3)
LS: (165)(0,2), (0,3), (1,1), (1,3), (2,3), (3,0)
LL: (191)(1,2), (2,0), (2,1), (2,2), (3,1)
傳給viewer

20. Phase 2: 算SS-list 之mean μSS=52 算SL-list 之mean μSL=80
算LS-list 之mean μLS=137
算LL-list 之mean μLL=198
分裂 SS-list, SL-list, LS-list 與LL-list 為
(52)(0,0)
(63)(1,0)
(68)(0,1)
(80)(3,2), (3,3)
(137)(0,2), (1,3)
(165)(0,3), (1,1), (2,3), (3,0)
(191)(2,1), (2,2)
(198)(1,2), (2,0), (3,1)
傳給viewer

21. 7.5 Side Match Method (SMM)
Block
Sub-image

22. The transmitting order of the blocks in the sub-images
7.5 SMM (cont.)
F blocks S1
Sf+1
Sf*f-1
S2*f
Sf+2 S2
F blocks
S2*f+1 S3 4
S2*f-1 2
Sf*f
S3*f-2 Sf 5 3 1
The transmitting order of the blocks in the sub-images

23. The received indices after the first stage
7.5 SMM (cont.)
The received indices after the first stage

24. The received indices after the second stage
7.5 SMM (cont.)
The received indices after the second stage

25. The diagonal blocks (D) Three types of the unknown blocks
7.5 SMM (cont.) S1
Sf+1
Sf*f-1
S2*f
Sf+2 S2
S2*f+1 S3
S2*f-1
Sf*f
S3*f-2 Sf
The lower blocks (L)
The upper blocks (U)
The diagonal blocks (D)
Three types of the unknown blocks

26. The prediction method for the upper blocks
7.5 SMM (cont.)
The prediction method for the upper blocks

27. The prediction method for the upper blocks
7.5 SMM (cont.)
The prediction method for the upper blocks

28. Experimental Results

29. Experimental Results
The sequence of the successive approximation of the original image with f*f is 2*2 (1&2)

30. Experimental Results
The sequence of the successive approximation of the original image with f*f is 2*2 (3&4)

31. Experimental Results
The sequence of the successive approximation of the original image with f*f is 3*3 (1&2&3)

32. Experimental Results
The sequence of the successive approximation of the original image with f*f is 3*3 (4&5&6)

33. Experimental Results
The sequence of the successive approximation of the original image with f*f is 3*3 (7&8&9)

34. According to the experimental results, SMM is usually better than BPM and DBTM in transmission cost and image quality.
The number of the transmitted bits per stage is constant.

35. 7.6 Selective Progressive Image Transmission (SPITM)
VQ-based PIT method
A new PIT method in spatial domain
Suitable for the most common environment (without special hardware supported and low-speed channel)
Select VQ as the basic compression technique
The best candidate for low-speed PIT (the simple decoding and low bit-rate compression)

36. 7.6 Selective Progressive Image Transmission (SPITM)
Transmitter
Use VQ to compress the original image into a index map
Partition the index map into regions with different priorities
Progressively transmit indices of each region: diagonal sampling
The receiver
Collecting the indices of each region
Use VQ to decompress the indices: side match technique to predict the blocks of unknown indices
Reconstruct an approximate image

37. 7.6 Selective Progressive Image Transmission (SPITM)
The image partition and diagonal sampling
More complicated and majority parts: usually located in the center area
Pay more attention to the central regions than the surrounding regions
The blocks in a central region of an image will have higher transmission priority
The image can be selectively transmitted

38. The partition of the index-map of image M
index-map of an image M is partitioned into R1, R2 and R3
transmission priority: Pri(R1) > Pri(R2) > Pri(R3)

39. R1, R2 and R3 further divided into sub-maps of size
d1*d1, d2*d2, and d3*d3 indices respectively
Diagonal Sampling 1
d+1
d*d-1
d*d d
3*d-2
2*d-1 2
2*d
d+2 3
2*d+1
d-1
b1,1
b1,2
b1,d
b2,1
b2,2
b2,d
bd*1
bd*2
bd,d
(a) A sub-image of size d*d blocks
(b) The diagonal sampling order
A sub-image and the order of diagonal sampling

40. The basic structure of selective PIT method for image M

41. Experimental Results Test the performance of SPITM Compare with
Bit-plane method (BPM)
Dynamic binary threshold method (DBPM)
Side-match method (SMM)
Test image: “Lena” of size 512*512 pixels (256 gray levels)

42. The partition of index map of the test image
VQ codebook (LBG): 256 vectors (16 dimensions)
Partition the index map into three regions R1, R2, and R3
Divide to sub-maps of size 2*2, 3*3, and 4*4
(0, 0)
(16, 16)
(40, 40)
(88, 88)
(112, 112)
(128, 128)
The partition of index map of the test image

43. Experiments
R(k): bit-rate
T(k): accumulated bit-rate

44. Experiments (cont.)

46. An image is partitioned into several regions
The region with highest priority will be transmitted first
SPITM is capable of selective transmission
Easy to implement and low transmission cost

47. 7.7 Fast Reconstruction Method (FRM)
FRM is an improvement of SPITM
FRM contains three technique VQ
Central Sampling Technique (CST)
Pixels Copy Technique (PCT)

48. 7.7 FRM (cont.) To select the indices to send
Central sampling technique (CST)
(a) A sub-image of 3 * 3 blocks
(b) A sub-image of 4 * 4 blocks
(c) A sub-image of 5 * 5 blocks

49. 7.7 FRM (cont.) To predict the unknown block
Pixels copy technique (PCT)

50. The codebook search of VQ is time-consuming
Pixels copy technique will make block effect

51. 7.8 Tree-structure Vector Quantization (TSVQ)
The most commonly used method so far for reducing searching time of VQ is TSVQ.
TSVQ is amenable to PIT due to its built-in feature of successive approximation

52. 7.8 TSVQ-progressive transmission tree
Sender
Receiver 1
2: Phase 1
5: Phase 2
10: Phase 3
21: Phase 4

53. TSVQ-Sender
Image vector
Indices

54. Image vector

55. Drawback of TSVQ
It is ineffective for image quality to send only one bit of each index in each phase.

56. 7.9 Side-Match reconstruction method using TSVQ (SMTSVQ)
SMTSVQ is an improvement of TSVQ
SMTSVQ contains three technique VQ
Progressive Transmission Tree (TSVQ)
Side Match

57. SMTSVQ-Sender Assume the depth of the codebook tree is n.
There are n phases in the whole process.
We break the process into two parts.
Part one: phase 1 to phase n/2
Part two: phase (n/2+1) to phase n
Part one
Part two
Transmit two bits of the index of each block in each phase

58. SMTSVQ-Sender (assume n=8)
Phase 5
Phase 1
Phase 6
Phase 2
Phase 7
Phase 3
Phase 8
Phase 4

59. SMTSVQ-Receiver
Transmitted blocks: TSVQ
Untransmitted blocks: SMVQ

60. SMTSVQ-Sender (phase one)
Image vector
Indices
Codebook size = = 31

61. Image vector

62. Experimental Results

63. Experimental Results

64. Experimental Results (phase 1)

65. Experimental Results (phase 2)

66. Experimental Results (phase 3)

67. Experimental Results (phase 4)

68. According to experimental results
SMTSVQ = SM + TSVQ
According to experimental results
Its image quality is better than that of SPITM, FRM and TSVQ in each transmission phase,
Its reconstruction time is much less than that of SPITM
SMTSVQ is therefore efficient and effective