Rule extraction from a trained neural network for image keywords extraction

This paper presents a rule extraction method from a trained neural network (NN), which is used for keywords extraction from images. In our approach, first, a bit map image in the RGB color space is transformed into that in the L*a*b* color space. Next, it clusters image pixels using the fuzzy c-means method and domains are extracted through a labeling process. Features, such as area of obtained domains, color information, and coordinates of the center of gravity, are then calculated, which are used as input attributes to NN. NN is then trained using such features. After NN learning, rule extraction is carried out using binarized output values in the hidden layer for each keyword. The rules extracted in this paper are If-then rules, which include logical functions. The methods of generating keywords using NN and the rules are presented and their comparative experiments are performed. Finally the validity of these methods was verified by means of computer simulations.