Enhanced Hit Rate Simulated Annealing in Codebook Training

Abstract

Codebook generation and optimization in the Vector Quantization system is a high dimensional and nonlinear optimization process. Lloyd's algorithm aims at the finding of the least mean square between the training vectors and their centroids. The major problem of this method is ''local optimal state''. Therefore Simulated Annealing (SA) is applied in the optimization of the codebook in VQ system. In the existing SA scheme, the successful rate of each generation, especially in the end of the process, would be very low. In this paper, novel schemes (Local Region Searching and Least Frequently Hit) are suggested to enhanced the performance and arise the hit rate of the generation of the existing SA algorithms. From the experimental result, the hit rate is arise by over 100 times of existing method. Improvement is found in both the fidelity testing and perception testing on the decoded image.