End-to-end estimation and optimization techniques for error-resilient video coding and networking.

摘要

This dissertation investigates end-to-end estimation and optimization techniques for error-resilient video coding and networking.; The first main contribution is concerned with optimization of the motion compensated prediction (MCP) framework to support error resilience. The problem is considered at two levels: (i) encoder MCP optimization given the MCP mechanism adopted at the decoder, (ii) optimal re-design the entire MCP framework at both encoder and decoder. At the first level, end-to-end distortion estimation is employed to propose rate-distortion (RD) optimized motion estimation and prediction, along with their respective low complexity variants. At the second level, a novel overall prediction mechanism, called generalizes source-channel prediction, is derived. All the proposed schemes are standard compatible. When jointly implemented they achieve significant performance gains.; The second main contribution is concerned with solving open questions to expand the applicability of the recursive optimal per-pixel estimate (ROPE), with particular emphasis on central obstacles to implementation within the H.264 standard, including: efficient but low (computation and storage) complexity approximation of cross-correlation terms; and the important but long ignored issue of rounding error accumulation and propagation. Efficient solutions are proposed and demonstrated to recoup virtually all the ROPE gains potential within H.264. The capabilities of ROPE are extended to effectively estimate the distribution of the decoder reconstruction random variables, which enables ROPE application to a broad class of distortion measures.; Another focus of the dissertation is on performance optimization of point-to-point scalable video networking. A variant of ROPE is developed and embedded within an RD optimized mode selection scheme for SNR scalable video coding. The system allows better prediction of the current base layer frame from past enhancement layer frame data, while explicitly optimizing decoder drift management and adaptive bit rate allocation.; To achieve error resilience, within frame bit allocation, a novel source-channel constant distortion (SCCD) model is proposed. The proposed model achieves higher modelling accuracy than current competitors. Better quality consistency across frames, as well as higher video quality on average, are demonstrated by a system exploiting the model for rate control.