Fuzzy class learning vector quantizer in image compression

The Fuzzy Learning Vector Quantizer (FLVQ) improved on the Generalized LVQ by defining a fuzzy learning rate. FLVQ produced globally optimum codebook from the global minimization of the mean space of Euclidean distances, where distances were measured from the codevectors to the data. However, FLVQ produces centers with no regards to human visual sensitivity to edges. The vectors of edges and shades may be classified in one cluster represented by a centroid which looks more like a shade vector than an edge vector, if the number of shade vectors is large, leading to a degradation in image quality. We propose a Fuzzy Class Learning Vector Quantizer (FCLVQ) which classifies the image into three classes of shades, midranges, and edges and then applies FLVQ to each class separately with superior results. A 3 dB improvement over FLVQ was obtained for a 256/spl times/256 greyscale image with a typical codebook size of 75. The FCLVQ has also the advantage of being computationally much faster.