Vector quantization of images using neural networks and simulated annealing

Abstract

Vector quantization (VQ) has already been established as a very powerful data compression technique. Specification of the 'codebook', which contains the best possible collection of 'codewords', effectively representing the variety of source vectors to be encoded is one of the most critical requirements of VQ systems, and belongs, for most applications, to the class of hard optimization problems. A number of new approaches to codebook generation methods using neural networks (NN) and simulated annealing (SA) are presented and compared. The authors discuss the competitive learning algorithm (CL) and Kohonen's self-organizing feature maps (KSFM). The algorithms are examined using a new training rule and comparisons with the standard rule is included. A new solution to the problem of determining the 'closest' neural unit is also proposed. The second group of methods considered are all based on simulated annealing (SA). A number of improvements to and alternative constructions of the classical 'single path' simulated annealing algorithm are presented to address the problem of suboptimality of VQ codebook generation and provide methods by which solutions closer to the optimum are obtainable for similar computational effort.<>