A refinement algorithm for vector quantization codebook design

A new algorithm for vector quantization codebook design is presented, which iteratively calls the classical generalized Lloyd algorithm (GLA) and keeps track of the best codebook so far. In our proposed Refine_ GLA the GLA runs are not independent trials, instead, a fraction of the initial codevectors are taken from the best codebook, while the rest are taken randomly from the training set. As the algorithm progresses that fraction is reduced so that the search in the codevector space gets more and more localised. Experiments to compress a grey scale image under different compression factors and transmission error rates have been conducted and compared against GLA in a fair basis (regarding computational load). The overall distortion reduction of our Refine_ GLA over GLA is finally reported and commented