A maximum likelihood approach to video error correction applied to H.264 decoding

摘要

Video error concealment has long been identified as the last line of defense against transmission errors. This is especially true for real time video communication systems where retransmissions are rarely used because of timing constraints. Since error handling is outside the scope of video coding standards, decoders may choose to simply ignore the corrupted packets, or attempt to decode their content. Video error correction is a viable alternative to deal with transmission errors when corrupted packets reach their destination. Until now, these approaches have received little considerations. This is mainly because the proposed methods either rely on specific coding tools or constraints, or require far too many computations compared to video error concealment techniques.ududIn this thesis, we propose a novel video error correction method based on maximum likelihood decoding. The method estimates the likeliest syntactically valid video slice content based on the erroneous video packets rather than discarding the content, and concealing the missing information. Such content is obtained by combining the likelihood of the candidate codewords with the bit modification likelihood associated to each candidate. We propose two solutions centered around our maximum likelihood decoding approach. First, we introduce a slice-level video error correction method. Furthermore, we show how to integrate the soft-output information shared by the channel decoder to evaluate the bit modification likelihood. We also show that it is possible to use our maximum likelihood decoding approach when soft-output information is not available. Then, we refine the solution at the syntax-element-level. The final solution we obtain can be used in real-time communication systems as it is computationally inexpensive compared to the slice-level solution, or the solutions proposed in the literature.ududOur final solution is then applied to the correction of videos conforming to the H.264 Baseline profile. We selected three 720x480 sequences, five 704x576 sequences, and one 720x576 sequence to run simulations. Each sequence was coded at a target bitrate of 1 Mbps, 1.2 Mbps, and 1.5 Mbps. All 27 sequences were then submitted to a noisy channel with a bit error rate ranging from 10−5 to 10−3. Our 5400 observations show a PSNR improvement of 1.69 dB over the video error concealment method implemented in the H.264 reference software. Furthermore, our results also indicate a 0.42 dB PSNR improvement over state-of-the-art error concealment STBMA+PDE.