Image coding employing vector quantisation

The work described in this thesis is concerned with the coding of digitised images employing vector quantisation (VQ). A new VQ-based coding system, named Directional Classified Gain-Shape Vector Quantisation (DCGSVQ), has been developed. It combines vector quantisation with transform coding tech-niques and exploits various properties of the human visual system (HVS) like frequency sensitivity, the masking effect, and orientation sensitivity, to produce reconstructed images with good subjective quality at low bit rates (0.48 bit per pixel). A content classifier, operating in the spatial domain, is employed to classify each image block of 8x8 pixels into one of several classes which represent various image patterns (edges in various directions, monotone areas, complex texture, etc.). Then a classified gain-shape vector quantiser is employed in the cosine domain to encode vectors of AC transform coefficients, while using either a scalar quantiser or a gain-shape vector quantiser to encode the DC coefficients. A new vector configuration strategy for defining AC vectors in the cosine domain has been proposed to better adapt the system to the local statistics of the image blocks. Accordingly, the AC coefficients are first weighted by an equivalent modulation transfer function (MTF) that represents the filtering characteristics of the HVS, and then they are grouped into directional vectors according to their direction in the cosine domain. An optional simple method for feature enhancement, based on inherent properties of the proposed strategy, has also been proposed enabling further image processing at the receiver. A new algorithm for designing the various DCGSVQ codebooks has been developed in two steps. First, a general-purpose new algorithm for classified VQ (CVQ) codebook design has been developed as an alternative to empirical methods proposed in the literature. The new algorithm provides a simple and systematic method for codebook design and reduces considerably the total num-ber of mathematical operations during codebook design. We have named this new algorithm Classified Nearest Neighbour Clustering (CNNC). A fast search algorithm has also been developed to reduce further computational efforts during codebook design. Secondly, a new optimisation criterion which is more suitable for shape code-book design has been developed and employed within the CNNC algorithm to design classified shape codebooks for the DCGSVQ. We have named this algo-rithm modified CNNC. The new algorithm designs the various shape codebooks simultaneously giving the designer full freedom to assign more importance to certain classes of vectors or to certain training vectors. The DCGSVQ system has been shown to outperform the full search VQ, the CVQ, and the transform coding CVQ (TC-CVQ) producing nicer coded images with better signal to noise ratio (SNR) figures at various bit rates. To improve further the perceived quality of coded images, a new postpro-cessing algorithm that can be applied at the decoder without increasing the bit rate has been developed. The proposed algorithm is based on various charac-teristics of the signal spectrum and the noise spectrum, and exploits various properties of the HVS. The proposed algorithm is a general-purpose algorithm that can be applied to block-coded images produced by various systems like VQ, transform coding (TC), and Block Truncation Coding (BTC). The algorithm is modular and can be applied in an adaptive way depending on the quality of the block-coded image. The last theme of this work has been the identification of useful fidelity criteria for image quality assessment. Quality predictors in the form of some subjectively weighted error measures were sought such that a smooth functional relationship exists between them and quality ratings made by human viewers. Quality predictors that incorporate simplified models of the HVS have been proposed and tested on a large set of VQ-coded images. Two such predictors have been shown to be better suited for image quality assessment than the commonly used mean square error (MSE) measure