Classified vector quantization using variance classifier and maximum likelihood clustering

Abstract

A classified vector image quantizer is proposed here. The algorithm employs a one-feature variance classifier. This classifier has good properties as it sorts the classes by its entropy contents. Then every class is clustered using the mixture maximum likelihood criterion instead of the Euclidean distance. This shows that the number of clusters required to represent any class can be determined. It also provides better clustering by emphasizing on fitting a model of a mixture of Gaussian distributions by the data. Impressive results regarding the image quality and bit rates are obtained when the algorithm is applied to image compression.