Invariant handwritten Chinese character recognition using fuzzy min-max neural networks

In this paper, an invariant recognition system using a fuzzy neural network to recognize handwritten Chinese characters on maps is proposed; characters can be in arbitrary location, scale and orientation. A normalization process is first used to normalize characters such that they are invariant to translation and scale. Simple rotation-invariant feature vectors called ring-data vectors are then extracted from thinned or non-thinned characters. Finally, a fuzzy min-max neural network is employed to classify the ring-data vectors by means of its strong ability of discriminating heavy-overlapped and ill-defined character classes. Several experiments with two kinds of character sets are carried out to analyze the influence factors of the proposed approach. The performances of the ring-data features and the fuzzy min-max neural network are compared with those of moment invariants and two traditional statistical classifiers, respectively. The ring-data features are found to be superior to the moment invariants, and also the fuzzy min-max neural network is found to be superior to the two classifiers. However, from the experimental results, we also see that the proposed approach is suitable to handle the translation, scale and rotation problem, but cannot solve the high-shape-variation problem of handwritten Chinese characters.