混合模式识别系统研究

讨论基于多种分类方法的模块组合实现的混合模式识别系统，它不同于利用多分类器输出结果表决的集成系统．提出两个系统：一个面向印刷体汉字文本识别，另一个面向自由手写体数字识别．利用多种特征和多种分类方法的组合、部分识别信息控制混淆字判别策略以及提出的动态模板库匹配后处理方法，使系统的性能与传统单一分类器系统比较，获得明显改善．实验表明：多方法多策略混合是解决复杂和增强系统鲁棒性的一条途径

The hybrid pattern recognition systems based on multiple classification models, which are different from the integrated system by voting outputs of multiple classifiers, are addressed in this paper. One of them is the printed chinese character recognition system, and another the unconstrained handwritten digit recognition system. By means of multiple features and multi  stage classification process, discrimination for ambiguous characters, and the novel post  processing stage employing dynamic reference mask matching technique, the performance of both recognition systems is substantially improved as compared with  single classifier system. Experimental results show that hybrid methodology is a promising solution to complex pattern recognition and robustness of a pattern recognition system.

国家自然科学基金