Self-organized Color Image Quantization for Color Image Data Compression

This paper presents a neural network approach to image color quantization and hence image data compression. Self-organizing feature maps form a basis for general vector quantization and this is applied to the tristimulus color values of image pixels. For image telecommunication systems such as videoconferencing, it Is desirable to constrain the encoder to a single pass of the image using the normal raster scan but this conflicts with the training requirements of a self-organized network. By appropriate choice of codebook size, this limitation can be turned into an advantage. The network performs a mix of vector quantization and run length coding, thus compressing the image data in two ways.