Kernel-Based Associative Memories,Clustering Algorithms And Their Applications

At present, support vector machine algorithms (SVMs) have achieved great success in many fields including pattern recognition, image processing, bioinformatics, etc. Accompanying with these successes, a newly developed research area named kernel-based learning theory and algorithms has rosen above horizon in the machine learning fields, following the neural networks upsurge in the end of last centrury. Now the research on kernel-based learning algorithms and their applications is one of the hot problems and mainstreams in machine learning works. Most of the present kernel-based learning algorithms put their focus on supervised learning problem such as classification. In this dissertation, however, another two important issues are focally investigated, which are associative memories and unsupervised clustering. Discrete associative memory models under kernel framework and kernel-based clustering algorithms are developed and applied to image compression, face recognition and medical image segmentation respectively. The main contributions of this dissertation are summerized as follows: Firstly, general discrete associative memory models under kernel framework are constructed. Concretly, one class of existing binary associative memeries is unified into a kernel-based binary associative memory model, and one class of multi-valued associative memeries is unified into a kernel-based multi-valued associative memory model. The conditions of stability for those associative memories are strictly discussed through defining corresponding energy functions. Finally, numerical experiments are performed to compare the storage capacity and error-correcting capability under different kernel functions for kernel-based multi-valued associative memory models. Secondly, two image compression algorithms using kernel-based associative memory are proposed. The first algorithm is based on kernel-based binary associative memory model and the second is based on kernel-based multi-valued associative memory. Experimental results on the benchmark images show that although the compression ratio of the first algorithm is not ideally high, it can implement progressive image compression and transmission. It can be also shown that without noises the second algorithm achives nearly the same compression performance as the vector quantization algorithm (VQ), and has apparent noise-dampping effects with dual (both channel and image) noises. Thirdly, a robust face recogntion algorithm using kernel-based associative memory is proposed. Experimental results on part of the FERET face image database show that the proposed algorithm still has relatively high recognition ratio after adding random noises up to 25% on face image, random discarding image blocks and occluding the face partly. Fourthly, several kernel-based algorithms including kernel-based fuzzy clustering in feature space named KFCM-â… , kernel-based fuzzy clustering in input space named KFCM-â…¡, kernel-based possibilistic clustering named KPCM, and online kernel-based algorithm called as ROC. Experimental results on artificial and benchmark datasets show that the proposed algorithms are robust, and suitable for clustering of incomplete or missing data and data corrupted by noises and outliers. Fifthly, two medical image segmentation algorithms using kernel-based clustering are proposed and implemented. Those two algorithms are designed by restricting the objective funtion of original kernel-based clusetering algorithm with membership constraints and spatial constraints respectively. Experimental results on BrainWeb simulated brain MRI images of McGill University in Canada and real brain MRI images show that both the proposed algorithms effectively solve the inherent problems in medical images including noises and intensity inhomogeneities, etc, and accordingly achive better image segmentation results.