Improved Initial-codebooks for LBG Algorithm

摘要

In this paper, we, starting from space distribution of trained vectors, generate better initial codebooks so that LEG algorithm with our initial codebooks can optimize the performance of vectors quantization. More precisely, we obtain N_1 cuts by applying theory of fuzzy sets as a cell partition of trained vectors space. By controlling A -cut level dynamically, the number of cuts and size of corresponding cell are determined. Then N_1 centers are found out as representative vectors (N_1 should be greater than the size N of initial codebooks) and its probability distribution is calculated. In a suitable manner, N_1 representative vectors are merged into N vectors which serve as initial codebooks. The tests show that the value of SNR increases by 0.5dB or so.