Abstract
The Self-organizing Map(SOM)is the most widely used artificial neural network algorithm in the unsupervised learning category. A clustering system is a major application of SOM. Normally, a SOM uses neurons of fixed size. However, it is hard to detemine the specific size of neurons that match the purpose of a clustering system, such as facial image clustering In this study, I propose a method that assesses the variation in the size of neurons according to the diversity of data. This method uses the Euclidean distance as an index to decide to whether insert a new neuron or eliminate a neuron. Test images of 3 and 26 patterns were inputted during the self-orgrnization process into the SOM which initially had 20 neurons. As a result, 7 and 69 neurons were obtained respectively. The proposed method created a clustering system that varied with the number of clusters. I applied the proposed method to classify facial images. I chose the infants′ facial images of 2 and 5 month old infants as test data, which had difference in complication of facial expressions. As a result, the number of clusters changed according to the variety of the infants′ facial egressions. Thus the proposed method is useful to make a clustering system for facial image analysis.
