K-mean Based Clustering and Context Quantization

In this thesis, we study the problems of K-means clustering and context quantization. The main task of K-means clustering is to partition the training patterns into k distinct groups or clusters that minimize the mean-square-error (MSE) objective function. But the main difficulty of conventional K-means clustering is that its classification performance is highly susceptible to the initialized solution or codebook. Hence the main goal of this research work is to investigate the effective K-means clustering algorithms to overcome this difficulty. An extensive task addressed by this thesis is to design a feasible context quantizer in circumventing the so-called context dilution problem, which is a specific form of K-means clustering problem. Publication P1 presents a genetic algorithm to tackle the k-center clustering problem by using randomized swapping to change one reference vector between the parent solutions in the crossover and then using a local repartition clustering procedure. The algorithm estimates the number of clusters automatically while optimizing the location of the clusters. It has been shown that the algorithm outperforms the other Kmeans algorithms reviewed in the publication.