Degree Name

Doctor of Philosophy


School of Electrical, Computer and Telecommunications Engineering


Lossless coding is widely applied in medical image compression because of its feasibility. This thesis offers two major contributions for lossless image compression; (i) the relationship between the minimum-mean-squared-error (MMSE) and the minimum-entropy-of-error (MEE) prediction in lossless image compression has been revealed, and (ii) novel methods of improving compression rates and operation using Shape-Vector Quantization (VQ) have been presented. These new schemes have a simpler implementation, more computational efficiency, and lower memory requirement than other lossless schemes have. The proposed schemes are capable of providing significant coding improvement over traditional predictive coders and adaptive predictive coders.

One major goal in any lossless image compression pursuit is to minimize the MEE. Realizing this goal is more valuable in terms of performance than minimizing the MMSE. Most predictive lossless coding techniques, however, are centered on the MMSE. The relationship between the MMSE and the MEE prediction and the limitation of linear prediction are the backbone of the Shape-VQ-based compression schemes introduced in this thesis. The concepts of the MMSE and the MEE are presented in detail and analyzed mathematically in the thesis. It is shown that one of the conditions for reaching minimum entropy using the MMSE is local stationarity, which makes adaptive coding feasible. This explains the reason behind the effectiveness of adaptive coding where the MMSE is pursuit rather than the MEE.

Predictive techniques are well accepted in lossless image coding. The main advantages of predictive technique over other coding techniques are the simplicity of its encoder and decoder, and its low computational complexity and overheads; the disadvantage is its comparatively poor decorrelation performance.

Several novel VQ-based compression techniques are presented in this thesis to improve the performance of predictive coders. The basis of these methods is that the indices of the vector-quantized version of image blocks represent their inter-block and inter-pixel correlation effectively. This fact is verified in this thesis where simulation results show the strength of Shape-VQ compared to other methods such as DCT in terms of the accuracy in classification. Methods introduced are all different adaptations of Shape-VQ-based adaptive lossless coding techniques. The significant features of these methods are their simplicity, speed, low memory requirement, and most importantly, their performance in decorrelation of redundancies among pixels.

Test results have shown that novel schemes presented in the thesis outperform other codecs in lossless image coding, especially with medical images such as mammograms. It is one of the major contributions of the thesis that traditionally hard-to-be-compressed mammograms, which have large amount of textures inside, can be encoded efficiently with high compression ratio and low computational and over-head requirement. This factor makes the proposed lossless image compression system be a feasible way for medical image especially mammogram compression and archiving.



Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.