50 Adaptive Structure Based SOM for Color Reduction

This paper proposes a structure adapted self-organizing map neural network (SA-SOM) to achieve adaptive clustering. Depending on the distance between input and existing units, the number of output neurons of SA-SOM is changed from 1 to N based on the defined learning rules with a distance constraint. After all training vectors are used for training; the structure of the network is fixed. To further refine the cluster centers, the network is tuned locally. When the network converges, its number of output neurons is the number of clusters and the weights are the centers of clusters. The proposed method is used for solving the problem of color reduction, experimental results show that the proposed method can effectively avoid the overexpansion of network and decrease training time.