A new vector quantization algorithm based on simulated annealing
Zhenya He, Chenwu Wu, Jun Wang, Ce Zhu · 2002
This paper presents a new VQ technique called the SA-K algorithm which incorporates the simulated annealing mechanism into Kohonen's competitive learning to produce high quality codebooks. With a proper temperature schedule, the SA-K algorithm asymptotically becomes a descent competitive learning algorithm and both the centroid and the nearest neighbor conditions for optimality are satisfied, while the SA technique guarantees that the SA-K algorithm performs in a globally optimal manner. Experimental comparisons among the SA-K, Kohonen learning algorithm (KLA) and LBG algorithm for speech source data are given. The novel algorithm consistently shows the advantage over the KLA and LBG algorithm in the design of vector quantizers with different codebook sizes.>