Feature based compression of vector quantized codebooks and data for optimal image compression

摘要

The Linde-Buzo-Gray (LBG) algorithm is a procedure in vector quantization. An attempt is made to compress the LBG codebook by simple feature-based clustering. The centroids of the resulting clusters form a reduced LBG codebook. The resulting reduced LBG codebook has almost identical SNR as the equivalent size original LBG codebook. Several other experimental results are presented which show that the above clustering technique may be applied directly to the original images in order to find an initial codebook for LBG codebook training, or to quickly generate an alternative codebook with comparable quality.