On the optimization of a Gabor filter-based feature extractor for handwritten character recognition

In order to use the Gabor feature extraction system in the real world, the problem of optimizing the Gabor parameters must be resolved. In the conventional optimization method, several candidates are prepared, the error rate for them is found through experimentation, and the parameters which give the lowest recognition error rates are taken to be optimal values. However, in the conventional method, setting the candidate values so that they are certain to include optimal values is a common problem. In this paper, the authors evaluate from the standpoint of the optimalness of the solutions and computational burden the optimization methods (two-stage conventional method, response surface method, and energy function method) that address the problems of the conventional method. (C) 2003 Wiley Periodicals, Inc.