Optimization of hybrid video coders

This thesis studies hybrid video coding with motion-compensated prediction and transform residual coding. It addresses three important issues in a hybrid video coder) namely motion compensation and estimation, residual field coding and rate control.;Consecutive frames in a video sequence are highly correlated. Their temporal relationship is represented by motion. Motion-compensated prediction using block matching is an effective tool to exploit this correlation. We show that there are two distinctive noise sources in motion compensation: motion uncertainty and quantization noise propagation. By analyzing the block motion estimate and its relationship to the motion field, a parametric solution is given for optimal component filters in overlapped block motion compensation, which reduces the motion uncertainty effects. We further propose a block-adaptive linear filtering framework that minimizes both motion uncertainty and quantization noise simultaneously. We also investigate the dual problem of motion compensation, i.e. motion estimation. A non-iterative motion estimation algorithm is proposed for overlapped motion compensation.;The residual field resulting from motion compensation is assumed to be independent of previous frames in current video coders. When this independence assumption is invalid, information leakage occurs. It is shown that there is in fact strong dependence between the residual frame and the previous frames. More specifically, the residual signal energy is related to the gradient magnitude which can be estimated from the previous decoded frame by using the motion estimates. This relationship is exploited to design efficient conditional video coders which significantly outperform current non-conditional video coders. It is also applied to the design of a pixel decimation-based fast motion estimation algorithm.;In practice, video coders operate in a closed loop with bandwidth constraint. A rate control unit is needed to ensure that the encoder and decoder buffers experience neither overflow nor underflow. We present such a rate control algorithm. It is primarily concerned with optimal bit allocation in video coding and is based on a parametric rate-distortion model. Bit allocation is performed in such a way that feedback control can be easily implemented to avoid buffer overflow and underflow, while optimal allocation is computed efficiently for the remaining bit budget after adjusting parameters each time the feedback information is received. The algorithm achieves visually optimal bit allocation by incorporating the characteristics of the human visual system. A perceptual classifier is used to exploit the noise sensitivity of each region and to assign bit budget accordingly.