Noise Reduction Prefiltering for Video Compression

1. Noise Reduction Prefiltering for Video Compression EE398, Stanford University March 16th 2006 Jim Deng Alex Giladi Fernando Gómez Pancorbo

2. Agenda • Project Background • Crawford’s Non Linear Filter • Test Setup • Experimental Results: Foreman • Experimental Results: Mother and Daughter • Future Work • Conclusions

3. Project Background • Practical video systems use a non-linear, so called “Crawford filter”, to improve both rate distortion performance and subjective quality for noisy input signals • The main goal of our project is to quantify the impact of Crawford filtering on the rate-distortion performance of an H.264 encoder • To that aim we have evaluated the impact of Crawford’s filter on several test video sequences • Based on our experiment results, we have come up with heuristic rules that link filter performance with the nature of the video and the noise

4. Crawford’s Non Linear Filter (1) • Crawford applies a non linear transfer function to the difference between co-located pixels in two consecutive frames. The output of the non linear characteristic is then added to the pixel located in the current frame. Intuitively, it “cuts off” the output when it exceeds a given threshold. • Reference • D.I. Crawford, "Spatio-temporal prefiltering for a video conference coder", Proc. Int. IEEE Conf. on electronic image processing, pp 236-242, 1982.

5. Crawford’s Non Linear Filter (2) • Crawford suggested the application of the filter to the static areas of the image • Two parameters control the non linear transfer function • Tau sets the threshold level • Gamma sets the sharpness of the cuttoff

6. Test Setup

7. Foreman (1)

8. Foreman (2)

9. Mother and Daughter (1)

10. Mother and Daughter (2)

11. Mother and Daughter (3)

12. Future Work • In the works • Filter along motion trajectory • Use sub-pel accuracy • Block adaptivity • Force minimum weight for source • Vary filter aggressiveness as a function of prediction error • Proposed • Optimize for forcing skipped blocks where possible • Implement as an in-loop filter

13. Conclusions (Objective) • The Crawford filter is effective in all cases we have tested in improving the PSNR of videos encoded with H.264. A gain of up to 7 dB has been achieved • The filter is more effective for Mother and Daughter (low motion) than for Foreman • With the tests done so far, we have determined the following heuristic rules for the optimum Crawford filter parameters when it is used together with our H.264 encoder, • Tau optimum is dependent on the variance of the noise. The higher the variance the higher the optimum tau. • Gamma optimum is located around 1 for both Gaussian and Uniform noise distributions. Additional tests should be done to determine whether this also holds for other noise distributions

14. Conclusions (Subjective) • We have noticed considerable improvement of quality in all optimum cases • Filtering with non optimal parameters impacts negatively the quality of the image, especially in moving areas • The filter significantly reduces the error propagations in interframe encoding of H.264

15. Q&A