Sample Weighting When Training Self-organizing Maps for Image Compression

| Image compression is an essential task for image storage and transmission applications. Vector quantiza-tion is often used when high compression rates are needed. Self-Organizing Map (SOM) algorithm can be used to generate codebooks for vector quantization. Previously it has been demonstrated that using the special property of the SOM algorithm that the codebook entries are ordered one can use prediction coding of codewords to make the compression more eective. In this paper it is shown that training the SOM algorithm by using diierent weighting for sample blocks having diierent statistical characteristics one can further increase the compression eeciency. Image compression is an essential task for image storage and transmission 1]. Lately the image compression using Vector Quantization (VQ) techniques has received large interest 2]. VQ methods ooer good performance when high compression rates are needed. In VQ approaches adjacent pixels are taken as a single block, which is mapped into a-nite set of codewords. In decoding stage the codewords are replaced by corresponding model vectors. The set of codewords and the associated model vectors together is called a codebook. In VQ the correlation which exists between adjacent pixels in a block is taken into account, and with a comparatively small codebook one achieves a small quantization error in reconstructed image. The main problem in vector quantization is to nd a codebook which minimizes the quantization mean error. Many design algorithms have been proposed. One of the best known is the Linde-Buzo-Gray (LBG) algorithm 3], which iteratively searches clusters in the training data. The cluster centers are used as the codebook model vectors, while the