The optimization of vector quantizers by minimizing variance

This paper presents the results of the optimization of a vector quantizer designed to quantize the coefficients from the transform coding of black and white images. Eight images were coded using 4×4 blocks of pixels which were in turn transformed using the discrete cosine transform. Three separate methods of optimization were investigated including the Linde-Buzo-Gray (1984) method, a simulated annealing method and a third method which minimizes the variance associated with each codebook value of the vector quantizer. Results indicate that the minimum variance method optimizes the codebook nearly as well as the simulated annealing method, but with much less processing time