(A) study on low rate vector quantization for image coding

Abstract

This thesis is concerned with the applications of low-rate vector quantization (VQ) to image coding. The basic goal is to improve the performance of VQ systems at low rates with less computational complexity and memory requirement, utilizing mean-based coding algorithms. To this end, firstly a practical mean-based source model is newly presented, and then conventional VQ algorithms are modified based upon the model of mean variation. However, the modifications are maintained simple such that the structure of the conventional full-search VQ fully utilized. The contents are summarized in four topics as follows. First, in order to model sources with nonstationarity, a composite autoregressive(CAR) model is presented, which is an autoregressive process with additive mean process of periodic variation. It has been shown that the CAR model can well represent real image sources with local mean variations. Rate distortion functions(RDF) are derived from parameters of a test image based on the model. It has been shown that the RDF bound is tighter than that based on the conventional AR model, since the CAR model incorporates the nonstationarity of local means. Second, a symmetrical ordering method is presented and shown to be able to reduce block effects which are common for any block-based image coding methods operated at low rates. Since the method is based on a simple modification of vector component order, one needs to change only the vector order or memory access address when implementing the method into a VQ system. Third, utilizing the block mean based on the CAR model as a selection criterion of a smaller codebook (window) which is sliding over a super codebook, a forward sliding search VQ(SSVQ-I) algorithm is proposed. It is shown to reduce the search complexity to 1/8 - 1/16 with negligible performance degradations. The algorithm is compared with classified VQ and tree search VQ of comparable complexity. Fourth, adopting the prediction of block mean into t...