A new algorithm for vector quantization

Summary form only given. Given a training set TS of n k-dimensional input vectors TS=(x₁,x₂,...,x_n ,) considered representative of the signal statistics, the task of designing the optimal vector quantizer consists of finding a set of clusters P={p1,p2,...,ph} containing n_p1, n_p2,..., n_ph vectors each in a way to minimize a distortion functional. In the article the mean square error with respect to the training set TS is considered. The experimental results presented have been obtained on three 512×512 pixels monochromatic (8 bit/pel) images. The proposed method and the well known generalized Lloyd algorithm are iterated until a stable quantizer is obtained. The computational complexity of the two methods is comparable