Abstract:

This paper presents an improved codebook design algorithm of vector quantization which combines Hadamard-transform and K-means theory,according to the Chen's vector quantization algorithm which has two weaknesses of high randomness for the formation of initial codebook and low encoding quality（PSNR）.The proposed algorithm uses the statistical features of classification average method of training vectors to generate initial codebook,and then improves the frequency of calculation the cell's centroid.Whenever a training vector is classified into lumen,the corresponding cell's centroid will be calculated to instead of the original code word.The proposed algorithm combines the advantages of LBG algorithm.The adjusted code word represents the characteristics of the lumen,accelerates the convergence speed of codebook,and improves the performance of codebook.Compared with the Chen's algorithm,the encoding quality can be improved by 0.5 dB,even when the iteration is low,the PSNR is improved by 0.9 dB.

Key words: vector quantization, codebook design, Hadamard-transform