A new class of feature-orientated motion estimation for motion pictures

The key to high-performance video coding lies in an efficient reduction of the temporal redundancies. For this purpose, motion estimation and compensation techniques have been successfully applied. We have a discussion on the drawbacks of some fast algorithms and also problems that are related to the full-search algorithm in block motion estimation. Much attention has been given to fast algorithms using criteria such as the mean absolute difference (MAD). However problems arise from using both fast algorithms and even the full search algorithm. These problems cause poor motion-compensated prediction along some desirable feature, to which the human visual system is very sensitive. We propose a generalized class of algorithms to resolve the problems. In general, our algorithms are adaptive, which includes consideration of the characteristics of block motions for typical image sequences, and an intelligent classifier to separate blocks containing different features. The motion vectors of these blocks are computed using feature frames and a masking factor, so that the motion-compensated frames are tied more closely to physical features. Experimental results show that this approach gives a significant improvement in accuracy for motion-compensated frames and computational complexity, in comparison with traditional intensity-based block motion estimation methods. More importantly, it gives far better images under subjective tests as compared to all other algorithms, including the full search algorithm. Finally, we illustrate our approach by giving some details of specific sample algorithms.