Fast entropy-constrained vector quantizer design

Vector quantization is the process of encoding vector data as an index to a dictionary-or codebook-of representative vectors. Entropy-constrained vector quantizers (ECVQ) explicitly control the entropy of the output, and are superior to simple nearest-neighbor vector quantizers in terms of rate-distortion performances. ECVQ codebook design based on empirical data involves an expensive training phase in which a Lagrangian cost measure has to be minimized over the set of codebook vectors. In this paper, we describe two new general elimination rules allowing significant acceleration in the codebook design process. These rules use features of the codebook vectors in order to discard most of them as fast as possible during the search. Experimental results are presented on image block data. They show that those new rules perform slightly better than the previously known methods.