A multilayer address vector quantization technique

A new vector quantization scheme for encoding monochrome and RGB color images is presented. The technique, called multilayer address vector quantization, is based upon using vector quantization in a multilayer fashion. In the first layer, a standard vector quantization is performed on each block of the image. In the second layer, the resulting vector quantization addresses of the neighboring blocks are merged together to form a new vector which is said to be address vector quantized if this address combination vector can be found in an address codebook. This process of merging four nodes (neighboring addresses) is repeated recursively in a bottom-up hierarchical manner in the subsequent higher layers using the corresponding address codebook in each layer. The address codebook in each layer consists of the most probable address combinations. A three-layer coding system is simulated where the codebook in the first layer is a standard VQ codebook, and the codebooks in the other two layers are address codebooks with codevectors consisting of the addresses of the previous layer codebook. Simulation results demonstrate good quality reproduction at bit rates above 0.23 b/pixel for monochrome images and 0.35 b/pixel for RGB color images. >