Evolutionary design of edge detector using rule-changing Cellular automata

A new design method for Cellular automata (CA) rules are described. We have already proposed a method for designing the transition rules of two-dimensional 256-state CA for grayscale image denoising. The gene expression programming was employed as the learning algorithm, in which the chromosome encodes the transition rule as the expression. The CA designed by the method ran faster than previous methods. In this paper, an improved method for designing the CA based edge detector is proposed. The ground truth for training CA is generated by the Canny edge detector, from which two objective functions are calculated. Both objective functions are optimized by a multi-objective evolutionary algorithm. The rule-changing CA is used to improve the performance. The experimental results showed that rule-changing CA designed by the proposed method have higher performance for edge detection than the ordinary CA.