Improved K-means Algorithm Using Initialization Technique Based on Edge-Mean Grid for Image Vector Quantizer Design

The K-means algorithm is famous as a traditional data clustering technique, which is also effective in the field of codebook design for Vector Quantization (VQ) based data compression. The main disadvantage of the conventional K-means algorithm is that the convergence speed and the quality of the generated codebook depend on the initial codebook. Many researchers have been devoted to improving the initial codebook, but their algorithms either do not make full use of the features of the training vectors or need high extra computational load. This paper presents an efficient initialization technique based on the edge-mean grid generated by an edge classifier as well as even grouping of sorted means. On the one hand, the training vectors are sorted in the descending order according to their mean values. On the other hand, the training vectors are classified into 8 categories based on the edge classifier, and the number of initial codewords generated from each category is proportional to the number of training vectors in this category. The mean-sorted training vectors in each category is evenly grouped into several groups with the number of groups equaling the number of codewords to be generated from the category. Thus, an edge-mean grid is obtained with many groups, and each group generates one initial codeword that is just the middle training vector in the group. Experimental results show that, compared with the random selection and the existing norm-sorted and meansorted grouping strategies, our initialization technique converges to a better codebook with a faster convergence speed.