基于模糊神经网络的印刷体汉字容错识别方法

类似于人脑的工作方式，模糊系统善于表达人的经验性知识，能处理难以依靠数学模型清晰表达的模糊性信息，在对汉字模式的辨识中具有较强的抗噪性。文中利用4层神经网络方法实现的一种模糊分类器，弥补了单纯的神经网络和单纯的模糊系统各自的不足，实现了模糊系统中规则的自学习性和自优化性，使系统具有很好的鲁棒性能。利用建立在一组网络上的表决机制，不仅降低了网络的复杂性，而且实现了对任意类数分类器的构造。实验结果表明，文中提出的方法很好地解决了对印刷体汉字进行客错识别问题。

On Achieving Better Fault-Tolerant Capability for Recognizing Heavily Stained Printed Chinese Characters with a Four-Layer Fuzzy Neural Network

How to achieve better fault-tolerant capability for recognizing written words remains, to our best knowledge, to be a rather difficult problem. It is even more difficult for Chinese characters, especially heavily stained Chinese characters. In the case of pattern recognition, including printed Chinese character recognition, fuzzy systems naturally possess high resolution sensitivity and very strong anti-noise ability. Sections 2 and 3 describe in detail how to employ a four layer fuzzy neural network for classifying Chinese character patterns by which fuzzy rules can be learned automatically. Section 4 describes in some detail a vote mechanism, which relies on a group of fuzzy networks, for solving the following difficulty: a large number of chasses for different Chinese character patterns always makes them hard to be classified accurately. Section 5 gives experimental results. Fig.2(a) gives 50 clearly legible printed Chinese characters; these 50 legible printed Chinese characters do not need our new method for getting correctly recognized. Fig.2(b) shows the same 50 characters with noise added intentionally to get them heavily stained; it is not unusual for printed Chinese characters to be heavily stained; these stained 50 printed Chinese characters do need our new method to be correctly recognized. The usual methods can recognize only 25~39 of these 50 heavily stained characters. Fig.3(b) shows 2 heavily stained and very complicated printed Chinese characters and Fig.4(b) shows 3 heavily stained and also quite complicated Chinese characters. The heavily stained 2 characters in Fig.3(b) have the additional difficulty of being very similar even when legible. Our new method can still recognize correctly the 2 characters in Fig.3(b). The same can be said about Fig.4(b).