Adaptive vector quantization on image sequence coding

This dissertation focuses on the application of vector quantization (VQ) toward image compression. Since images are subjected to human observation, image compression does not require an exact reconstruction but only reconstruction similar to the original image. In the last decade, numerous variations of vector quantization have been studied in the domain of speech and image compression. Pruned tree-structured vector quantization (PTSVQ) has been developed to take advantage of variable rate coding, which assigns a higher bit rate to active regions in the image and assigns a lower bit rate to less active regions in the image. Adaptive tree search vector quantization (ATSVQ) has also been developed to take advantage of variable rate coding. For a given tree-structured codebook, a given image and a fixed bit rate, the optimization equations for the bit allocation vector quantization (BTVQ) are formulated. Then, the bit allocation problem is solved by using the Lagrangian multiplier method. The performance of these three algorithms is compared. For mean-removed VQ, the bit allocation algorithm has the best performance on the X binary or hexadecimal tree-structured codebook.; Because the codebook used in VQ determines the coding performance, the BTVQ is chosen to evaluate the effects of codebook changing and codevector updating methods which adapt the codebook to local statistics. In changing the entire codebook, it is found that multistage tree-structured VQ (MSTSVQ) has the best performance. A codevector updating method applied on the multistage 2-level tree-structured codebook is developed, which updates a portion of the codebook according to the encoded distortions in multistage VQ (MSVQ). The codevector updating algorithm is a two-pass algorithm.; For sequence images, VQ alone does not compress images well. A better method to reduce the redundancy in time domain is to use the motion compensation technique. The motion compensation method uses the previous frame (image) to predict the current frame (image). The residual (motion compensated image) is then vector quantized. Two motion compensation techniques are proposed. One is a modified two-dimensional logorithm algorithm. The other is a motion vectors prediction algorithm. A coding state of using motion compensation only is also incorporated in the bit allocation vector quantization algorithm. Finally, the codevector updating algorithm is operated in the motion-compensated images.