A gain-shape vector quantizer for image coding

The design of a gain-shape vector quantizer (GSVQ) for image coding that gives significantly improved image quality over basic VQ is presented. The quantizer is a spatial domain quantizer belonging to the class of product code VQ called mean-residual vector quantizers (MRVQ). In MRVQ the basic vector space of image blocks is transformed into a space of vector residuals by removing the block sample mean from the vectors. This results in a better utilization of code vectors for encoding shade blocks because they are mapped into small regions near the origin where they can be encoded efficiently. There still remains a problem, however, in finding an optimal partition for the rest of the space to encode the remaining visually sennitive edge/texture blocks so that distortion is minimized. GSVQ handles this problem nicely by a normalization process that maps each block into a gain part representing the vector´s magnitude that controls the luminance level of the block and a shape part representing the spatial distribution of the vector elements. Among the implementation features, an k-d tree nearest neighbor clustering process is used to generate the initial codebook for the VQ codebook optimization procedure and is shown to yield significant improvements in performance.