A perspective view of finite state binary residual VQ

A novel approach to vector quantization is examined in the context of image coding. The method can be characterized as a merger of finite state vector quantization (VQ) and a binary residual vector quantizer. Because the system is inherently predictive, the quality of representation easily exceeds that of memoryless VQ systems with comparable size vectors in the range from 4 to 64. The advantage offered is that the codebook storage requirement can be reduced dramatically without significantly impairing performance. This approach (called finite state binary residual VQ or FSBRVQ) is examined in the context of vector dimension. Still image coding based on FSBRVQ and BRVQ is studied using vectors that range in dimension from 16 to 1024. Some examples are given and some conclusions are drawn with respect to performance, memory, and VQ dimension tradeoffs