Tuning Baseline JPEG Encoding for Individual Images

1
Tuning Baseline JPEG Encoding for Individual
Images

• Chien-An Lai
• Changchuan Yu
• EE398A Final Project
• Mar 15th 2007

2
Outline

• Motivation
• Baseline JPEG
• Optimization Algorithm with baseline JPEG
  compliance
• Results
• Conclusion

3
Motivation

• Flexibility of baseline JPEG Q and H tables
• ? what gain could we get from changing the
  tables?
• A 97 algorithm revisited in 07
• Joint implementation of different optimizing
  techniques
• R-D optimized quantizer selection
• R-D optimal thresholding
• Optimal Huffman encoding
• Compare the best baseline JPEG to
  state-of-the-art JPEG2000

4
Baseline JPEG

• Block-Diagram

DCT (88)
Quantizer Q matrix (88)
Entropy Coding (Huffman)
101101
Bit sequence
88 blocks
5
Modified baseline JPEG

• Block-Diagram

DCT (88)
Quantizer
Entropy Coding
Threshold T matrix
101101
Bit sequence
Q
T
H
88 blocks
Q matrix Selection
Optimal Thresholding
Huffman table Customization
Image-adaptive optimization
6
Algorithm Overview
DCT (88)
Quantizer
Entropy Coding
Threshold T matrix
101101
Bit sequence
Q
T
H
88 blocks
Q matrix Selection
Optimal Thresholding
Huffman table Customization
Loop til converge
Image-adaptive optimization
7
Algorithm - Q

• For n164
• For q1255
• End
• End
• A fast algorithm is needed, or its equivalent to
  jpeg encoding 25564 pictures for one single Q
  search!

8
Algorithm - Thresholding
find which k is optimal for minimal cost and
backtracking from k to find the T
coefficients. This is equivalently the optimal
value for 263 thresholding combinations.
9
Algorithm Huffman Table

• Optimization under the constraint of baseline
  JPEG format
• Reordering of Huffman table according to the real
  statistics

table... 0 1 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 2 2 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 3 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 4
1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 5 5 1 1 0 1
0 0 0 0 0 0 0 0 0 0 0 0 0 6 6 1 1 1 0 0 0 0
0 0 0 0 0 0 0 0 0 0 7 7 1 1 1 1 0 0 0 0 0 0
0 0 0 0 0 0 0 8 10 1 1 1 1 1 1 0 1 1 0 0 0 0
0 0 0 0 9 16 1 1 1 1 1 1 1 1 1 0 0 0 0 0 1 0
0 10 16 1 1 1 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1
4 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 2 6 1 1
1 0 0 1 0 0 0 0 0 0 0 0 0 0 1 3 7 1 1 1 1 0
0 1 0 0 0 0 0 0 0 0 0 1 4 9 1 1 1 1 1 0 1 1 0
0 0 0 0 0 0 0 1 5 11 1 1 1 1 1 1 1 0 1 1 0 0 0
0 0 0 1 6 16 1 1 1 1 1 1 1 1 1 0 0 0 0 1 0 0
Re-ordering
2-D symbol
run
cat
code length
10
How we did the testing

• Matlab modeling of baseline JPEG
• An accurate model, with everything but headers
• Slow!
• Baseline JPEG rate-psnr using scaling of Q table
• JPEG2000 rate-psnr using rate-constrained Jasper
  coder
• Mostly run on Scien Lab machines

http//www.ece.uvic.ca/mdadams/jasper/
11
Results1 bridge512x512
12
Results1
32.5dB, 2.0 bits/pixel
32.3dB, 1.4 bits/pixel
13
Results1
Q optimized 11 29 33 31 41
33 43 41 37 33 31 37 39
41 41 57 33 33 37 37 41
41 45 49 35 35 37 37 43
43 53 85 35 41 41 41 43
57 75 111 39 41 43 45 57
83 95 101 37 47 51 53 75
77 93 81 47 47 57 65 97
95 83 87
Q baseline 15 10 10 15 23
38 49 58 11 11 13 18 25
55 57 52 13 12 15 23 38
54 66 53 13 16 21 28 49
83 76 59 17 21 35 53 65
104 98 73 23 33 52 61 77
99 108 88 47 61 74 83 98
115 114 96 69 88 90 93 107
95 98 94
14
Results2 512x512
15
Results2 Halfdome top
32.8dB, 0.8 bits/pixel
32.9dB, 1.0 bits/pixel
16
Results2 Halfdome top
27.9dB, 0.8 bits/pixel
26.5dB, 0.8 bits/pixel
17
Results2
Q optimized 46 19 19 17 19
15 21 23 23 19 19 19 19
21 21 27 20 19 17 19 19
21 25 27 19 19 17 21 21
21 19 27 15 19 17 19 19
21 21 23 19 17 19 17 21
19 21 25 17 21 21 19 21
25 23 35 21 19 21 21 19
23 25 59
Q baseline 8 6 5 8 12 21
26 31 6 6 7 10 13 30
31 28 7 7 8 12 21 29
36 29 7 9 11 15 26 45
41 32 9 11 19 29 35 56
53 40 12 18 28 33 42 54
58 47 25 33 40 45 53 62
62 52 37 47 49 51 58 52
53 51
18
Result scaling effects
256x256, halfdome
512x512, halfdome
19
Result it does not always work
peppers512x512
20
Conclusion

• Optimization of Q and H tables can result in
  significant gain, compared to recommended tables
• Most gain from Q optimization
• Huffman optimization contributes some performance
  gain
• Optimal Thresholding produces little gain
• At a cost of thousands times of computation
  effort
• 1 hours of searching in matlab, on Scien Lab
  machines
• No guarantee for global optimization
• JPEG2000 is a far superior winner, hands down
• Encoding takes only mini-seconds
• Precise control of bit rate

21
References

• 1 G. K. Wallace, "The JPEG still picture
  compression standard," Communications of the ACM,
   vol. 34, no. 4, April 1991.
• 2 M. Crouse and K. Ramchandran, "Joint
  thresholding and quantizer selection for
  transform image coding entropy-constrained
  analysis and applications to baseline JPEG", IEEE
  Transactions on Image Processing, vol. 6, no. 2,
  February 1997.
• 3 K. Ramchandran and M. Vetterli,
  Rate-distortion optimal fast thresholding with
  complete JPEG/MPEG decoder compatibility, IEEE
  Trans. Image Processing, vol. 3, pp. 700704,
  Sept. 1994.