Efficient searching for robust CNN templates with combined analytic and evolutionary methods

In this paper, we propose a method that combines the analytic method and a genetic algorithm (GA) for the design of robust templates for cellular neural networks (CNNs). The relationship of the template coefficients derived from the analytic method can serve as possible bounds for the solution space. A genetic algorithm then follows to search for robust templates in the reduced solution space. Due to the bound set by the analytic method, the number of the useless searches in the genetic algorithm can be dramatically reduced from more than 90% to about 30%. Two popular image processing methods: hole-filling and shadowing processes are presented to demonstrate the capability of the proposed method. The robust templates can be readily found in only a few, typically 2 to 5, generations.