An improved initialization algorithm for use with the K-means algorithm for code book generation

Abstract

A novel algorithm for codebook initialization for use with the K-means algorithm for codebook generation is presented. This algorithm was shown to result in better codebooks at considerably reduced generation time compared to Lloyd's algorithm and to other codebook initialization algorithms. The proposed algorithm generated codebooks with smaller maxima and smaller standard deviations of the codebook vector quantization (VQ) distortions. These outcomes were consistently observed with substantial margins for all cases considered.<>