Utilizing Soft Information in Image Decoding

1. Utilizing Soft Information in Image Decoding Hannes Persson Karlstad

2. Outline • Background • JPEG2000 • Soft information • Experiments • Conclusions & future work

3. Soft Information in Image Decoding • General purpose • Make end-points aware of channel conditions • Reduce retransmissions • Maximum a-posteriori (MAP) estimates • Soft decoding • Produced at the receiver • Assume bits are independent and equally likely • Transparent to the wireless system used (independent of modulation, coding) • Reliability measure of the received bits • Express the certainty of a bit’s value to be correct

4. Communication Layers • Modification limited to the receiver • Assumptions about underlying communication layers • Overlook bit-errors in user data • Support inter-layer communication to be able to forward soft information Application - JPEG2000 Transport - TCP-L Network - IP Data link Physical - log-MAP soft information

5. JPEG2000 Overview Forward wavelet transform Arithmetic encoder Compressed data Quantization Inverse wavelet transform Arithmetic decoder Dequantization Compressed data

6. Arithmetic Coder • Arithmetic coder applied to subsets of transformed data - code-blocks • Restricts the amount of data the arithmetic coder is exposed to • Bit-errors is contained in the code-block • No propagation to other code-blocks • Error resilient mechanisms in JPEG2000 • Encoded at the transmitter • Bit-error detection with the arithmetic decoder • Bit-error will stop the decoding of a code-block • Error concealment

7. Utilizing Soft Information • Bit-error detection with the arithmetic decoder • Soft information utilization attempts to correct bit-errors by iterative arithmetic decoding • Heuristic algorithm that swaps potential erroneous bits • Potential erroneous bits deduced from soft information • Goal is to correct all bit-errors in the code-block • At least make error concealment less frequent

8. Experimental Set-up • Image code-block size - 4x4, 16x16, 32x32, 64x64 • 16-QAM modulation • Signal to Noise power relation - 5-16 dB • Image transmitted 30 times over the noisy channel (lena image) • Received image decoded with 3 different decoders AWGN channel simulation

9. 6 dB

10. 8 dB

11. 8 dB

12. 8 dB Baseline 12.2452 dB Baseline 15.5012 dB Modified 17.7354 dB (no mechanisms used) (utilizes mechanisms) (utilizes mechanisms + soft info.)

13. 10 dB

14. 10 dB

15. 10 dB Baseline 18.2799 dB Baseline 19.335 dB Modified 28.8897 dB (no mechanisms used) (utilizes mechanisms) (utilizes mechanisms + soft info.)

16. 12 dB

17. Conclusion & Future • Soft information in image decoding • High gain in image quality possible (8-12dB) • Both PSNR and visually • Improvements of 30 percent observed • Impact of code-block size • Future • Enhance algorithm for finding erroneous bits • Layer interaction for the soft information exchange • Format on soft information • Combine with TCP-L

18. Soft information in multimedia • Multimedia standards already define error detection mechanisms • Motion JPEG2000 • MPEG-4 • RVLC • Soft information and iterative decoding • Soft information already exists • Implementation issue • Computational overhead