Parallelization of VQ codebook generation by two algorithms: parallel LBG and aggressive PNN [image compression applications]

Abstract

Summary form only given. We evaluate two parallel algorithms for the codebook generation of the VQ compression: parallel LBG and aggressive PNN. Parallel LBG is based on the LBG algorithm with the K-mean method. The cost of both latter algorithms mainly consists of: a) the computation part; b) the communication part; and c) the update part. Aggressive PNN is a parallelized version of the PNN (pairwise nearest neighbor) algorithm, whose cost mainly consists of: a) the computation part; b) the communication part; and c) the merge part. We measured the speedups and elapsed times of both algorithms on a PC cluster system. When the quality of images compressed by both algorithms is the same, the number of training vectors required by the aggressive PNN is much less than that by the parallel LBG, and the aggressive PNN is superior in terms of the elapsed time.