The Optimized Context Modeling And Its Application On Microbial Genome Sequence Compression And Image Compression

The Context modeling technology is widely used in information compression.To enhance the compression efficiency,the Context model which has high performance should be constructed for coding.When the estimation of each conditional probability distribution contained in the Context model is obtained by counting those symbols encoded previously,one phenomenon referred to as“Context Dilution” will reduce the accuracy of the estimation.This is introduced since the number of conditional probability distributions in the model is too large or the number of samples for the training of the model is insufficient.Context quantization is suggested to tackle this problem by merging some of the conditional probability distributions and the size of the Context model is thus reduced.However,the previous Context quantizers based on the Context tree modeling is suitable only for those sources with a small symbol alphabet and the Context quantizer designed to minimize the adaptive codelength can only be applied to binary sources.Therefore,it is necessary to design new optimized Context quantizers for aryI-sources such that the number of quantization levels and their corresponding quantization partitions are optimized at the same time.On the other hand,another Context modeling approach which can efficiently make use of the correlation among source symbols without increasing the size of the Context model is the weighting of Context models.However,there is no criterion to evaluate the performance of Context weighting at present.That means the weighting of the Context models should be executed during the whole coding process once the weights are given,no matter which probability distributions in the models are participated in weighting.By far,the weights are determined either empirically or on the basis of the previous coding performance of the respective models.It implies that weights are not optimized.Our work in this dissertation aims to tackle the aforementioned problems in Context quantization and Context weighting by theoretical analyses and algorithmimprovements.Those improved methods will be applied to microbial genome sequence compression and image compression.The contributions are given as follows:For Context quantization,the minimum description length is suggested as the optimization objective to design the Context quantizer for aryI-sources.The increment of the description length is defined as the measure for evaluating the similarity among those counting vectors which are to be merged.Its properties are discussed in details and it is applied in adaptive clustering algorithms in order to implement the Context quantization.Meanwhile,the Context quantization based on the quantization of the values of the Context in an order-1 Context model is explored.By using dynamic programming,the order-1 Context model for aryI-sources can be globally optimized to improve the coding efficiency.For Context weighting,it is found that Context weighting is equivalent to the weighting of the description lengths of those counting vectors participated in weighting.Then the weighting cost is defined to evaluate the performance of Context weighting before the weighting operation is executed.Furthermore,the similarity between these counting vectors can also be evaluated by the weighting cost.This implies that the weighting operation can be optional during the coding process to avoid the weighting of the probability distributions which will lead to a high weighting cost.Meanwhile,the weights can be optimized when the weighting cost is set as the minimization objective.The searching algorithm MOA is suggested to implement the weights optimization.Furthermore,for those stationary sources,the weights optimization algorithm based on the least square algorithm are also discussed.Then,two mapping methods are used to transform genome sequences into numerical sources for the compression of microbial genome sequences.Those Context weighting techniques proposed are also employed to encode bases in the genome sequences.For image compression,the Context quantizer designed based on the minimum description length is applied instead of the fine-tuned empirical Context quantizer to enhance the codingefficiency.For the compression of the fingerprint images,the optimized Context weighting algorithms are used to improve the coding results.The experimental results indicate that the proposed Context quantizer can achieve slightly better results than other Context quantizers.Meanwhile,the optional Context weighing and the weights optimizing algorithms can lead to better results for either microbial genome sequences or fingerprint images.