Design optimization of code-excited neural vector quantizers
Zhicheng Wang, J.V. Hanson · 2005
The LBG algorithm is the most common and important algorithm of classical vector quantization (VQ) for speech or image signal compression. However, this algorithm has two major weaknesses. First, its encoding complexity grows exponentially with the product of coding rate and vector dimension and the storage requirement of the codebook increases linearly with the product. Secondly, it easily gets trapped in local minima of the distortion surface, resulting in a suboptimal vector quantizer. Neural vector quantizers have been developed to overcome the first problem. To solve the second problem, a class of randomized search algorithms such as simulated annealing and cauchy annealing have been applied to codebook designs. This paper presents a method to solve the two problems simultaneously with globally optimal code-excited neural vector quantizers (CENVQs), which applies annealing procedures to global optimization of CENVQs. Comparisons among the different vector quantizers are presented for several data sources.