Mini-max algorithm for image adaptive vector quantisation

The paper presents a technique which employs the mini-max error criterion for image compression using adaptive vector quantisation. Vector quantisation is a promising method for low bit rate image compression. In vector quantisation (VQ), the image vectors are usually coded using a 'universal codebook' generated from a set of training images. The coding performance using this codebook is potentially poor for images outside the training set. A number of inter- and intra-image techniques have been proposed to adapt the codewords to the input image. However, these techniques do not guarantee the closest codewords to be within a prespecified bound of the input vectors. This can result in large errors which give rise to artifacts. We propose an intra-image adaptive technique which employs a criterion that minimises the maximum error. Here, the codebook is generated on the fly from the input vectors to be coded. A primary codebook of size two, four or eight is typically used to store the frequently used codewords. A larger secondary codebook is used to store the less frequently used codewords. Both the transmitter and receiver maintain identical codebooks and hence keep track of any changes without any overhead information. As it is a single-pass technique, real-time implementation is possible.