Hybrid learning methods for vector quantization and its to image compression

Abstract

Neural networks for vector quantization such as K-means, neural-gas (NG) network and Kohonen's self-organizing map (SOM) is proposed. K-means, which is a "hard-max" approach, converges very fast, but it easily falls into local minima. On the other hand, the NG and SOM methods, which are "soft-max" approaches, are good at the global search ability. Though NG and SOM exhibit better performance in coming close to the optimum than K-means, they converge slower than K-means. In order to the drawbacks that exist when K-means, NG and SOM are used individually, we have developed hybrid methods such as NG-K and SOM-K. This paper investigates the effectiveness of NG- K and SOM-K in an image compression application. Our simulation results show that NG-K and SOM-K have good scalability to the number of weight vectors.