Effect of neighbourhood size selection in SOM-based image feature extraction

Abstract

Feature extraction is one of the processes in pattern recognition systems. It is the procedure of selecting the best possible subset of input features from a high dimensional problem space and transforming them into a lower dimensional space. Mathematically, feature extraction can be perceived as dimensionality reduction problem. Kohonen's self-organizing map (SOM) is an unsupervised learning approach in which a group of neurons regularly adjusts their basic structure as a function of its experience and environment. Preliminary research reported that neighbourhood functions is one of the necessary parameters that influence the results. The neighbourhood size refers to the region covered by the activated neighbouring neurons in relation to the winner neuron in each iteration of learning. This study aims to evaluate the implementation of SOM in image feature extraction. Different neighbourhood size value was tested in the SOM learning process to observe the quality of the extracted feature. Gaussian neighbourhood function was applied to the SOM-based feature extraction in current study. This study also compared the feature extraction performance of different neighbourhood functions. Experimental results show that the resulting feature loses its details as the neighbourhood size increases. Gaussian neighbourhood function gives more significant extracted feature.