One-pass vector quantizer design by sequential pruning of the training data

A one-pass vector quantizer design algorithm is presented. The algorithm sequentially selects a subset of the training vectors (in a high density area of the training data space) and computes a VQ code vector. Next, a sphere is constructed about the code vector whose radius is determined such that the encoding error for points within the sphere is acceptable. Finally, the data within the sphere is then pruned (deleted) from the training data set. This procedure continues on the remainder of the training set until the desired number of code vectors are located. This one-pass VQ design algorithm is compared with several benchmark results for uncorrelated Gaussian, correlated Gaussian, and Laplace sources; its performance is seen to be as good or better than the benchmark results. Further, the one-pass algorithm needs only slightly more computation than a single iteration of the LBG algorithm