Source representation for improved channel code performance in Wyner-Ziv video coding

Abstract

The Wyner-Ziv video coding paradigm is a new coding paradigm which exploits most of the source correlation at the decoder. This differs from the traditional predictive video coding schemes, where the source correlations are exploited solely at the encoder. Hence, the new paradigm can enable the implementation of low complexity encoders suitable for various applications such as endoscopy capsules and low-power surveillance systems. Slepian-Wolf (SW) and Wyner-Ziv (WZ) theorems prove that when the complexities, of exploring the source statistics, are shifted from the encoder to the decoder, the coding efficiency should not be affect. Hence, under certain conditions, the coding performance of WZ video coding schemes can theoretically be made arbitrarily close to that of conventional schemes where the sources are jointly encoded and decoded. However, the Rate-Distortion (R-D) performance of practical Wyner-Ziv video coding architectures are still far from the best performance attained with predictive video coding architectures like the H264/AVC or High Efficiency Video Coding (HEVC) video coding scheme. This Thesis investigates several methods to improve the performance of Slepian-Wolf coding, in terms of coding efficiency and reduced decoding delays. It is noticed that the traditional Slepian-Wolf coding approaches encode the bits within the same bit-plane level randomly using Low-Density Parity-Check Accumulate (LDPCA) codes and this leads to a sub-optimal performance. The reliability of the bits can be predicted and used to ensure that the bit nodes receiving low reliability bit-predictions are given better protection. A novel LDPCA code construction, targeted to consider the coding problem at hand, is thus proposed. Furthermore, this work also analyses the performance of the traditional LDPCA codes at different entropy points and studies the best way to distribute the correlation noise amongst bit-planes to improve coding efficiency. This is achieved by accumulating the most unreliable bits within the first decoded bit-planes and correcting the remaining bit-planes, having few bit-errors, using 8-bit or 16-bit Cyclic Redundancy Check (CRC) codes. The careful arrangement of discrepancies amongst bit-planes is used together with the arrangement of bits within each bit-plane and the new LDPCA codes, to obtain performance gains of up to 23 % during Slepian-Wolf coding. In context of Slepian-Wolf coding, the Thesis also addresses a comprehensive analysis of the mismatch present within regions of low motion. This is used to develop a scheme where the quantisation module alters between the floor and the round operator, at different pixel or coefficient locations. The operator which is more likely to avoid the Slepian-Wolf codec from correcting mismatch caused by small variation in light intensity is then chosen to improve R-D performance by up to 0.52 dB. Finally, the long decoding times required for Slepian-Wolf decoding are reduced by considering a new indexing scheme and histogram equalisation technique. For parallel WZ video coding architectures, these techniques ensure that the Slepian-Wolf decoders running on different cores of a multi-core processor can finish decoding at the same time, aiding parallel decoding and reducing decoding times by up to 32 %, with minimal affect on the R-D performance. The obtained reduction in rates and decoding delays helps bridge the gap in performance compared to the traditional video coding systems and pave the pathway for applications based on WZ video coding paradigm.