【作者】 唐方坤；

【导师】 龚敏；

【作者基本信息】 四川大学 ， 凝聚态物理， 2004， 硕士

【摘要】 脱机手写体汉字识别一直是一个很复杂的模式识别问题。汉字字符集所具有的数量大，结构复杂，字形变化大等特点，使得脱机手写汉字识别成为字符识别领域最大的难题之一，但同时，脱机手写汉字识别是一个非常重要的课题之一，它的成功决定着无限制机器汉字自动识别的实现，是智能化计算机在中国真正普及的前提。本文主要在汉字字符的特征提取和自动识别方面做了如下的工作。 论文首先就手写汉字识别的研究现状进行了综合论述，介绍了汉字识别的发展阶段，对各种方法进行了较为详细的介绍，对其优缺点进行了论述。 论文提出了一种新的特征提取的方法。目前汉字特征提取主要分为两类：基于结构的特征提取和基于统计的特征提取方法。前者理论上而言更为准确，能体现汉字的基本特征，但是在实现问题上要准确提取出汉字结构是非常困难的；后者实现简单，也能体现汉字的一些宏观特征，但是却失去了汉字的主要结构特征，不足以描述汉字。本文提取了一种统计和结构相结合的特征提取方法，首先用基于方向码，多个分解算子相结合的方法将汉字分解为四个方向上的分量子图像，同时对汉字字符图像进行模糊网格的构造，然后以网格为单位，分别统计图像四个子图像的网格特征。最后得到特征向量。 论文将“以高维空间几何分析方法为工具”，“以空间复杂几何形体最佳覆盖为目的”的仿生模式识别原理用于手写汉字识别，并详细描述了用多权值神经元网络来具体实现汉字识别系统的过程。并将识别过程和效果同目前流行的SVM方法进行了分析对比。实验说明，由于仿生模式识别理论是以每一类事物的“认识”为目的，因此它在属于超多类模式识别的手写汉字识别方面有无可比拟的优越性。更多还原

【Abstract】 Off-line Handwritten Chinese Characters Recognition(OHCCR) is always a complex pattern recognition problem. Due to the Chinese characters’ large vocabulary, complex structure, large variations of shapes and fonts, OHCCR becomes one of the most difficult problems in character recognition field. In the other hand, OHCCR is a very important task, and it’ s breakthrough will result in the realization of non-restricted computer Chinese character auto-recognition, and is the precondition that intelligent computer can be prevalent in China. This thesis mainly finishes the following work in the field of off-line handwritten Chinese characters’ feature extracting and auto-recognition.In this thesis, we first present a comprehensive and critical survey of Chinese characters recognition, including it’ s development phases, various methods in the process of pre-processing, feature extracting and recognition, at the same time, we point out the current research emphasizes and difficulties of OHCCR.Secondly we present a novel feature-extracting method of Chinese characters. In the present, the methods of Chinese character feature-extracting mainly classify two classes: based on structure and based on statistic. The former is more precise in theory and can embodythe main feature of Chinese characters, but it is very difficult to extract the structure of Chinese characters exactly. The latter can realize easier relatively and it also can embody some macro feature of characters, but it loses the main structure feature of characters. In this thesie, we present a method combining structure and statistic feature. First divide the character image into four sub-images according the direction code and combining multi dividing operators, at the same time, constitute the fuzzy network of character image, then to each grid, account the four sub-images’ s feature respectively and get the feature vector.Thirdly, we use a novel theory- Biomimetic Pattern Recognition, which use high-dimension space Geometry analysis as tools and regard The high-dimensional complex geometrical shape optimal covering as the goad, to the Chinese characters recognition, and despict how to use ANN with multi-weights vector to actualize recognition in detail. We compare the recognition performance of BPR with the traditional pattern recognition methods, such as SVM. Experiment results show that BPR is superior when it is used in the large classes pattern recognition, like Chinese character recogniton.更多还原