Bayesian VQ image filtering design with fast adaption competitive neural networks

Abstract

Vector Quantization (VQ) is a well known technique for signal compression and codification. In this paper we propose the filtering of images based on the codebooks obtained from Vector Quantization design algorithms under a Bayesian framework. The Bayesian VQ filter consists in the substitution of the image pixel by the central pixel of the codevector that encodes the pixel and its neighborhood. This process can be interpreted as a Maximum A Posteriori restoration based on the codebook estimated from the image. We apply the VQ filter to noise removal in images from micromagnetic resonance. We compare our approach with the more conventional approach of applying VQ compression as a noise removal filter. Some visual results show the improvement introduced by our approach.