Colour image segmentation using the self-organizing map and adaptive resonance theory

Abstract

We propose a new competitive-learning neural network model for colour image segmentation. The model, which is based on the adaptive resonance theory (ART) of Carpenter and Grossberg and on the self-organizing map (SOM) of Kohonen, overcomes the limitations of (i) the stability-plasticity trade-offs in neural architectures that employ ART; and (ii) the lack of on-line learning property in the SOM. In order to explore the generation of a growing feature map using ART and to motivate the main contribution, we first present a preliminary experimental model, SOMART, based on Fuzzy ART. Then we propose the new model, SmART, that utilizes a novel lateral control of plasticity to resolve the stability-plasticity problem. SmART has been experimentally found to perform well in RGB colour space, and is believed to be more coherent than Fuzzy ART.