An Evolutionary Approach for the Generation of Diversiform Characters Using a Handwriting Model

In pattern recognition, a large number of diversiform characters is necessary to train / test a handwritten character recognition system. However, it is not easy to collect a large number of natural samples. The artificial diversification of characters has been suggested as one means of collecting a variety of characters[1]. In this paper, we show that a handwriting model can be applied to the diversification of characters. The characters diversified by the model can be used as a database of character images for training / testing purposes. Wada & Kawato’s handwriting model [2] is based on an optimal principle and the feature space of the characters includes sets of via-points extracted from actual handwritten characters. The handwriting model can be used to generate a variety of characters by changing viapoint information . In this paper, we propose a method for generating a large variety of characters by changing via-point information based on a genetic algorithm and we show that the accuracy of a handwritten character recognition system that uses the characters generated by the proposed method as the training data, is equivalent to that of a system composed by using natural data.