Improvement to color quantization based on self-organizing maps for computer graphics partition

A novel approach to improve color quantization based on self-organizing maps for computer graphics partition is proposed. According to the attributes of the color images, the output neurons of the self-organizing maps were divided into gray-level group and color group which were initialized separately. The neurons in the gray-level group and the color group were also trained separately in training process. At the same time, the scheme of growing, pruning and merging was utilized to adjust adaptively the structure of network. Experimental results showed that the convergence rate and quantization precision of the neural network were drastically improved. The requirements of the sequential image segmentation and recognition could be satisfied.