Estimation-theoretic framework for scalability and packet loss resilience in predictive video coding.

摘要

This dissertation proposes an Estimation-Theoretic (ET) framework which enables scalability and packet loss resilience in predictive video coders, while maintaining high compression efficiency. The fundamental premise of ET framework is that optimal estimation, given all the available information, is the ultimate process that should underly compression. While this framework leads to several new solutions to long-standing problems in the predictive source coding, its most important application is in the field of standard DCT-based predictive video coding.; First, an ET approach is derived for enhancement-layer prediction in a generic scalable coder. The resulting ET prediction is shown to be optimal in the sense that it minimizes the mean squared prediction error given all the information available at the enhancement layer. The performance of the ET predictor is demonstrated on scalable DPCM coding of Markov sequences. The method is then adopted for predictive-DCT based scalable coding of video, where considerable performance gains are demonstrated. Further, ET prediction is used to improve the bit rate scalability of two-channel audio coding.; As further applications, the use of ET prediction for robust video compression is demonstrated in two settings. Multiple description video coding, an important tool for packet loss resilience, is shown to benefit from optimal prediction within the ET framework.; The second line of research uses the ET framework to address the problem of error propagation and packet loss resilience in predictive video coders. Here, the objective is to enable the design of efficient Joint Source-Channel Coding (JSCC) schemes which minimize the total decoder distortion, for the given rate and channel loss condition. Towards this goal, a Recursive Optimal per-Pixel Estimate (ROPE) of the expected total decoder distortion is derived.; The use of ROPE for parameter optimization in JSCC video coders leads to significant performance gains. This is illustrated using the important error resilience tool of macroblock (MB) coding mode selection. The resulting technique, referred to as ROPE-RD, achieves substantial PSNR gains over widely used RD and nonRD based mode-switching methods. The ROPE-RD algorithm is then extended to incorporate feedback information from the receiver. Simulation results show that ROPE-based mode selection substantially outperforms conventional prediction mode selection schemes.; Finally, the combination of the ET-prediction based error concealment method at the decoder, and ROPE-based mode selection at the encoder, is shown to further improve the packet loss resilience of scalable video coders. (Abstract shortened by UMI.)