A codebook design method for fast VQ search

Abstract

Vector quantization (VQ) is a multidimensional block quantizer methodology that can be very efficient with respect to approaching the rate-distortion bounds. Best performance is obtained for longer blocks. However, as the vectors become longer, the number of codebook vectors will generally increase, and the complexity of searching the codebook for the best codebook vector may soon become prohibitive. For best quantizer performance, the VQ must be trained for the source, but this usually prohibits the use of structured (or algebraic) codebooks that are fast to search. This paper presents a novel methodology for codebook design that combines the traditional training of codebooks by the well proven generalized Lloyd algorithm (GLA) with a structured codebook that can be searched efficiently. The concept is termed gradient search algorithm (GSA) since it is based upon a gradient in the error surface of the codebook pointing towards the optimum codebook vector choice.