An algorithm for designing vector quantizers

Abstract

Quantization plays an important role in most communication schemes, yet, though extensive results are available for scalar quantization, till recent years very little appeared in the literature concerning the design of vector quantizers. This work presents the first algorithm for designing uniform vector quantizers for various blocklengths (dimensions), which does not utilize a training sequence. The algorithm is motivated by a separation technique which has appeared in the literature and uses least dense covering lattices to generate the quantizers. The algorithm is used to generate quantizers for blocklength up to nine. These are compared with the best available quantizers in the literature and found to compare favorably. The quantizers are used to transform code an image. The results show the improvement to be obtained over scalar quantization. Coding strategies are also presented for both small and large codebooks.