On the initialization and training methods for Kohonen self-organizing feature maps in color image quantization

In this paper, we propose a new Gray-Color initialization method for use with the Kohonen's self-organizing feature maps in color image quantization. In our method, the neurons in the competitive layer are initialized in two distinct groups and the input pixels are categorized accordingly. By training the two groups of neurons separately, both the image intensity and color information are better managed for diverse classes of images when the number of neurons is sparse. Compared with the gray scale initialization, our method improves the mean square error of artificial images by 30% on average. The performance gain is achieved with no additional resource and little extra computational effort from the existing SOFM architecture.