【作者】 鲍胜利；

【导师】 沈予洪；

【作者基本信息】 四川大学 ， 机械设计与理论， 2002， 硕士

【摘要】 手写体(脱机)汉字识别是当前OCR技术研究的热点之一，也是计算机字符识别中最为困难的一个课题。本文在分析当前汉字识别最新发展技术的基础上，提出了一种基于多算法集成和神经网络的汉字识别系统。在此系统中，我们针对神经网络在小类别模式识别中的成功应用，先采用基于汉字粗外围特征的距离分类器作为粗分类，以将待识汉字集分成若干个小的汉字集合，然后用基于汉字弹性网格像素概率分布特征的BP神经网络分类器作为细分类，以实现汉字识别的目的。 本文首先对系统中汉字输入、预处理、粗分类、细分类和后处理五大模块进行了较详细的说明，特别是对神经网络分类器，不仅讨论了其原理、特征提取、BP算法实现和网络结构及参数选择，还探讨了BP算法的缺陷问题并提出了改进方法。 然后在Matlab神经网络工具箱的基础上，探讨了BP网络在Matlab环境中的实现，并给出了BP网络建立、训练和仿真过程的编程方法。 最后，我们在Matlab中对10类汉字100个不同样本进行了初步仿真实验，取得了识别率为95％的良好效果，表明将神经网络引入到手写体汉字识别的研究是比较成功和可行的。更多还原

【Abstract】 Handwritten (off-line) Chinese character recognition (HCCR) has become favorable area in OCR by now,and is the most difficult project in computer character recognition too. After analyzing the currently up-to-date techniques for Chinese character recognition,in this paper,we propose a Chinese character recognition system based on multi-algorithm integration and neural network. According to the successful application to pattern recognition of small category for neural network,in this system,we use a distance classifier based on gross periphery feature for rough classification in order to classify the total Chinese character set to some small sets,and then a BP network classifier based on the probability distribution of pixels with elastic meshing is used for fine recognition.In this paper,first,the five modules in the system are explained in detail including the input of Chinese character,preprocessing,rough classification,fine classification and post-processing. Especially as to the neural network classifier,we not only discuss the fundamental principle of BP network,feature extraction,the realization of BP network,the selection of network structure and parameters,but also discuss its drawbacks and its improved solutions.Secondly,based on the neural network toolbox,a convenient realization on MATLAB is discussed for BP neural network,and the programming methods are presented about how to create a network,train a network and simulate a network.Lastly,by use of MAILAB,preparatory experiment on total 100 samples of 10 categories of handwritten Chinese characters produced the result of recognition rate of 95%,showing that using neural network for HCCR is feasible and promising. It will be of great importance to direct the establishment of practical HCCR system.更多还原