脱机手写体汉字识别研究
|
摘要
脱机手写体汉字识别是当前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,它的研究对汉字信息处理自动化及开拓新一代计算机的智能输入都有着重要的意义。手写体汉字识别是一个非常复杂的多模式识别问题,多年研究实践表明,单一方法的效果是有限的,各种方法有其自身的特点和优点,但也有其局限性。应用信息融合技术,采用多方法有机结合,走多特征融合、多方案集成的道路,是手写体汉字识别的一个发展趋势。
考虑到单个分类器不能从根本上有效地提高分类性能,需要依靠多分类器集成来解决问题,故本文在分析当前汉字识别最新发展技术的基础上,设计了一种基于多特征融合、多分类器集成的汉字识别策略,即三级串行分类器集成模型。在该模型中,将距离分类器和神经网络分类器串行集成,并与三种不同的特征提取方法结合构成识别系统,探讨了不同种系统设计方案的最优融合策略,最终选用如下集成模型:一级分类,采用基于汉字均匀外围特征的曼哈顿距离分类器进行粗分类;二级分类,一改传统均匀网格划分下提取汉字穿透特征的方法,提取基于弹性网格划分的汉字穿透特征,利用相似度进行细分类;三级分类,提取基于汉字弹性网格划分的四方向线素分解特征,综合前两级分类器的识别结果,选取较为流行的BP神经网络分类器对候选结果作确认分类。
本文研究对象为少量常用汉字,研究目标是探索非特定人低限制手写体汉字脱机识别的有效算法。实验选取了国标GB2312-80一级字库中的50个汉字,每个汉字采集了100个样本,共5000个汉字样本,并利用Matlab7.1工具箱对系统设计模型进行了初步仿真实验,结果表明该模型是有效的。
本文内容对汉字样张采集、预处理、粗分类、细分类和实验结果分析五大模块进行了较详细说明。汉字样张的预处理包括对待识汉字样张的二值化、平滑去噪、倾斜校正、汉字切分、汉字大小、位置归一化处理以及样本库的建立及存取操作。分类器设计中主要介绍了BP神经网络分类器的原理、网络结构设计及参数的选择,讨论了BP算法的缺陷及改进策略;最后,利用Matlab7.1神经网络工具箱编程实现了BP网的训练和仿真。
Off-line handwritten Chinese character recognition (HCCR) is the current hot spots of OCR technology research, and is also one of the most difficult task of the computer character recognition. The research on off-line handwritten Chinese character recognition is quite significant for the automatic processing of Chinese character information and the development of intelligent input of the new generation computer. Handwritten Chinese character recognition is a very complicated multi-pattern recognition issue, years of research shows that the effect of a single method is limited, and various methods have their own characteristics and advantages, but also have their limitations. Multi-feature fusion and integration of multiple schemes are considered to be a trend for the development of handwritten Chinese character recognition with the use of information fusion technology and the organic combination of multiple methods.
Since the individual classifier can not fundamentally improve the classification performance effectively, and the integration of multiple classifiers is required to solve the problem, a three-level serial classifier combination model was designed on the basis of analyzing the development of current technology for Chinese character recognition. This model was based on multi-feature fusion and multiple integrated classifiers. In this model, the distance classifiers and the neural network classifier which were serially integrated were combined with three different feature extraction methods to form a recognition system. The optimal integration strategies of different design schemes of systems were discussed and finally the integration model was obtained. On the first level, the characters are rudely classified by Manhattan distance classifier based on the peripheral feature of uniform meshes; On the second level, instead of using the traditional method of extracting characters stroke density feature on the basis of uniform meshes, the characters stroke density feature was extracted on the basis of elastic meshes partition, and the fine classification was also performed using similarity classifier; On the third level, four directional line element decomposition feature was extracted on the basis of elastic meshes partition, and the popular BP neural network classifier was selected to confirm the classification of the candidate results according to the former recognition results of the two classifiers.
In this paper a small number of commonly used Chinese characters were studied, and the research target is to explore the effective algorithm for recognition of off-line handwritten Chinese characters which is non-special and low limited. 50 Chinese characters are selected from first level library Chinese character of GB2312-80 in experiment. 100 samples of each Chinese character have been collected, and the total samples reached 5,000. Matlab7.1 toolbox was used to carry out a preliminary model simulation experiment, and the results show that the model is effective.
In this paper, five modules of Chinese sample collection, image preprocessing, rough classification, fine classification and experimental results analysis were detailed. The preprocessing of character image samples includes the establishment and access operation of Chinese characters to samples database and the processing of image binarization, smooth denoising, lean adjustment, segmentation, normalization of size and normalization of position. The principle of BP neural network classifiers, the network architecture and the choice of its parameters were introduced in the design of classifiers. The disadvantages of BP algorithm and the way of making improvement were then discussed, and finally the BP network training and simulation are realized by programming with the use of Matlab7.1 neural network toolbox.
考虑到单个分类器不能从根本上有效地提高分类性能,需要依靠多分类器集成来解决问题,故本文在分析当前汉字识别最新发展技术的基础上,设计了一种基于多特征融合、多分类器集成的汉字识别策略,即三级串行分类器集成模型。在该模型中,将距离分类器和神经网络分类器串行集成,并与三种不同的特征提取方法结合构成识别系统,探讨了不同种系统设计方案的最优融合策略,最终选用如下集成模型:一级分类,采用基于汉字均匀外围特征的曼哈顿距离分类器进行粗分类;二级分类,一改传统均匀网格划分下提取汉字穿透特征的方法,提取基于弹性网格划分的汉字穿透特征,利用相似度进行细分类;三级分类,提取基于汉字弹性网格划分的四方向线素分解特征,综合前两级分类器的识别结果,选取较为流行的BP神经网络分类器对候选结果作确认分类。
本文研究对象为少量常用汉字,研究目标是探索非特定人低限制手写体汉字脱机识别的有效算法。实验选取了国标GB2312-80一级字库中的50个汉字,每个汉字采集了100个样本,共5000个汉字样本,并利用Matlab7.1工具箱对系统设计模型进行了初步仿真实验,结果表明该模型是有效的。
本文内容对汉字样张采集、预处理、粗分类、细分类和实验结果分析五大模块进行了较详细说明。汉字样张的预处理包括对待识汉字样张的二值化、平滑去噪、倾斜校正、汉字切分、汉字大小、位置归一化处理以及样本库的建立及存取操作。分类器设计中主要介绍了BP神经网络分类器的原理、网络结构设计及参数的选择,讨论了BP算法的缺陷及改进策略;最后,利用Matlab7.1神经网络工具箱编程实现了BP网的训练和仿真。
Off-line handwritten Chinese character recognition (HCCR) is the current hot spots of OCR technology research, and is also one of the most difficult task of the computer character recognition. The research on off-line handwritten Chinese character recognition is quite significant for the automatic processing of Chinese character information and the development of intelligent input of the new generation computer. Handwritten Chinese character recognition is a very complicated multi-pattern recognition issue, years of research shows that the effect of a single method is limited, and various methods have their own characteristics and advantages, but also have their limitations. Multi-feature fusion and integration of multiple schemes are considered to be a trend for the development of handwritten Chinese character recognition with the use of information fusion technology and the organic combination of multiple methods.
Since the individual classifier can not fundamentally improve the classification performance effectively, and the integration of multiple classifiers is required to solve the problem, a three-level serial classifier combination model was designed on the basis of analyzing the development of current technology for Chinese character recognition. This model was based on multi-feature fusion and multiple integrated classifiers. In this model, the distance classifiers and the neural network classifier which were serially integrated were combined with three different feature extraction methods to form a recognition system. The optimal integration strategies of different design schemes of systems were discussed and finally the integration model was obtained. On the first level, the characters are rudely classified by Manhattan distance classifier based on the peripheral feature of uniform meshes; On the second level, instead of using the traditional method of extracting characters stroke density feature on the basis of uniform meshes, the characters stroke density feature was extracted on the basis of elastic meshes partition, and the fine classification was also performed using similarity classifier; On the third level, four directional line element decomposition feature was extracted on the basis of elastic meshes partition, and the popular BP neural network classifier was selected to confirm the classification of the candidate results according to the former recognition results of the two classifiers.
In this paper a small number of commonly used Chinese characters were studied, and the research target is to explore the effective algorithm for recognition of off-line handwritten Chinese characters which is non-special and low limited. 50 Chinese characters are selected from first level library Chinese character of GB2312-80 in experiment. 100 samples of each Chinese character have been collected, and the total samples reached 5,000. Matlab7.1 toolbox was used to carry out a preliminary model simulation experiment, and the results show that the model is effective.
In this paper, five modules of Chinese sample collection, image preprocessing, rough classification, fine classification and experimental results analysis were detailed. The preprocessing of character image samples includes the establishment and access operation of Chinese characters to samples database and the processing of image binarization, smooth denoising, lean adjustment, segmentation, normalization of size and normalization of position. The principle of BP neural network classifiers, the network architecture and the choice of its parameters were introduced in the design of classifiers. The disadvantages of BP algorithm and the way of making improvement were then discussed, and finally the BP network training and simulation are realized by programming with the use of Matlab7.1 neural network toolbox.
引文
[1]高彦宇,杨扬,脱机手写体汉字识别研究综述[J],计算机工程与应用,(2004)7,74-77
[2]丁晓青,汉字识别研究的回顾[J],电子学报,(2002)9,1364-1368
[3]张世辉,孔令富,汉字识别及现状分析[J],燕山大学学报,(2003)27,367-369
[4]荆涛,王仲,光学字符识别技术与展望[J],计算机工程,(2003)29,1~2
[5]李鑫,慧晓威,张全贵,脱机汉字识别技术研究的方法及发展趋势[J],计算机时代,(2005)1,1-2
[6]朱小燕,史一凡,马少平,手写体字符识别研究[J],模式识别与人工智能,(2000)13,174-180
[7]万红梅,金连文,结合距离分类器的神经网络手写体汉字识别[J],计算机工程与应用,(2004)11,55-56
[8]金连文,梁宇杰,一种新的距离分类方法及其应用[J],计算机工程,(1999)25,30-31
[9]居琰,汪同庆,彭建等,特征融合用于手写体汉字识别研究[J],电子科技大学学报,(2002)31,229-233
[10]杨键,杨静宇,一种组合特征抽取的新方法[J],计算机学报,(2002)5,570~575
[11]何华,刘冠荣,脱机手写汉字机器识别方法研究[J],自动化博览,(2000)6,43-44
[12]封筠,李巍,李莉蓉,脱机手写体汉字识别技术研究的几点思考[J],电脑学习,(2003)3,4~5
[13]郭戈,闰继宏,蒋红梅等,基于结构特征的汉字识别[J],甘肃工业大学学报,(2003)29.81-85
[14]马少平,基于模糊方向线素特征的手写体汉字识别[J],清华大学学报,(1997)37,42--45
[15]吴天雷,马少平,基于重叠动态网格和模糊隶属度的手写汉字特征抽取[J],电子学报,(2004)32,86-89
[16]孙利民,基于子块特征及其相关模糊特征的手写体汉字识别方法[J],通信学报,(1999)20,81-84
[17]李国宏,施鹏飞,手写体汉字笔画特征点的完整性分析[J],计算机工程,(2006)32,14-16
[18]李国宏,施鹏飞,基于笔划方向特征和非对称分布的手写体汉字识别[J],上海交通 大学学报,(2005)39,1988~1992
[19]金连文,彭秀兰,尹俊勋,一种手写体汉字特征提取新方法-小波变换及弹性网格技术的应用[J],中国图像图形学报,(1998)7,549~552
[20]马少平,夏莹,朱小燕,基于模糊方向线素特征的手写体汉字识别[J],清华大学学报(自然科学版),(1997)3,42-45
[21]张睿,丁晓青,方驰,脱机手写汉字识别的最优采样特征新方法[J],中国图像图形学报,(2002)7,176-180
[22]Jin L W,Handwritten Chinese Character Recognition with Directional Decomposition Cellular Features[J],Journal of Circuits System and Computers,(1998)8,517~524
[23]高学,金连文,尹俊,一种基于笔画密度的弹性网格特征提取方法,模式识别与人工智能,(2002)15,351-354
[24]金连文,徐秉铮,手写体汉字识别中的一种新的特征提取方法-弹性网格方向分解特征[J],电路与系统学报,(1997)2,7~12
[25]钟国华,金连文,手写体汉字扇形弹性网格特征提取的新方法[J],计算机工程,(2002)28,61~62
[26]吴天雷,马少平,基于重叠动态网格和模糊隶属度的手写汉字特征抽取[J],电子学报,(2002)32,186~190
[27]陈友斌,丁晓青,一种手写汉字特征抽取新方法[J],信号处理,(1998)14,117~122
[28]居琰,汪同庆,刘建胜等,基于集成RBF神经网络的小类别手写体汉字识别系统[J],(2002)23,100~102
[29]居琰,汪同庆,刘建胜等,有限集手写体汉字特征提取及分类器设计[J],重庆大学学报(自然科学版),(2002)25,96-129
[30]金连文,覃剑钊,手写汉字识别弹性网格Gabor特征提取方法研究[J],计算机应用研究,(2003)12,163~165
[31]李玉静,扬扬,基于矩和Gabor变换的手写体汉字识别方法[J],信息技术,(2003)12,44-46
[32]王学文,丁晓青,基]~Gabor变换的高鲁棒汉字识别新方法[J],电子学报,(2002)9,1317-1322
[33]田学东,郭宝兰,基于Gabor变换的汉字字体识别研究[J],计算机工程,(2002)20,89~91
[34]李晓辉,翟娟娟,历史文献识别中图像预处理方法的研究[J],微机发展,(2004)14,30~31
[35]N.Otsu.Traps System,A threshold selection method from gray-level histograms[J],IEEE Trans System Man arid Cybernetics,(1979)9,62-66
[36]杨淑莹,VC++图像处理程序设计[M],北京,清华大学出版社,(2005)1,56-107
[37]张世辉,汉字图像预处理算法的研究及实现[J],微机发展,(2003)13,53~58
[38]Lianwen Jin,Gang Wei,Handwritten Chinese Character Recognition with Directional Decomposition Cellular Features[J],Journal of Circuit System and Computer,(1998)8,517~524
[39]徐丽娜,神经网络控制[M],哈尔滨,哈尔滨工业大学出版社,(1998)12,126~127
[40]Kimura F,Shridhar M,Handwritten numerical recognition based on multiple algorithms[J],Pattern Recognition,(1991)24,969-983
[41]韩力群,人工神经网络教程[M],北京,北京邮电大学出版社,(2006)12,22-80
[2]丁晓青,汉字识别研究的回顾[J],电子学报,(2002)9,1364-1368
[3]张世辉,孔令富,汉字识别及现状分析[J],燕山大学学报,(2003)27,367-369
[4]荆涛,王仲,光学字符识别技术与展望[J],计算机工程,(2003)29,1~2
[5]李鑫,慧晓威,张全贵,脱机汉字识别技术研究的方法及发展趋势[J],计算机时代,(2005)1,1-2
[6]朱小燕,史一凡,马少平,手写体字符识别研究[J],模式识别与人工智能,(2000)13,174-180
[7]万红梅,金连文,结合距离分类器的神经网络手写体汉字识别[J],计算机工程与应用,(2004)11,55-56
[8]金连文,梁宇杰,一种新的距离分类方法及其应用[J],计算机工程,(1999)25,30-31
[9]居琰,汪同庆,彭建等,特征融合用于手写体汉字识别研究[J],电子科技大学学报,(2002)31,229-233
[10]杨键,杨静宇,一种组合特征抽取的新方法[J],计算机学报,(2002)5,570~575
[11]何华,刘冠荣,脱机手写汉字机器识别方法研究[J],自动化博览,(2000)6,43-44
[12]封筠,李巍,李莉蓉,脱机手写体汉字识别技术研究的几点思考[J],电脑学习,(2003)3,4~5
[13]郭戈,闰继宏,蒋红梅等,基于结构特征的汉字识别[J],甘肃工业大学学报,(2003)29.81-85
[14]马少平,基于模糊方向线素特征的手写体汉字识别[J],清华大学学报,(1997)37,42--45
[15]吴天雷,马少平,基于重叠动态网格和模糊隶属度的手写汉字特征抽取[J],电子学报,(2004)32,86-89
[16]孙利民,基于子块特征及其相关模糊特征的手写体汉字识别方法[J],通信学报,(1999)20,81-84
[17]李国宏,施鹏飞,手写体汉字笔画特征点的完整性分析[J],计算机工程,(2006)32,14-16
[18]李国宏,施鹏飞,基于笔划方向特征和非对称分布的手写体汉字识别[J],上海交通 大学学报,(2005)39,1988~1992
[19]金连文,彭秀兰,尹俊勋,一种手写体汉字特征提取新方法-小波变换及弹性网格技术的应用[J],中国图像图形学报,(1998)7,549~552
[20]马少平,夏莹,朱小燕,基于模糊方向线素特征的手写体汉字识别[J],清华大学学报(自然科学版),(1997)3,42-45
[21]张睿,丁晓青,方驰,脱机手写汉字识别的最优采样特征新方法[J],中国图像图形学报,(2002)7,176-180
[22]Jin L W,Handwritten Chinese Character Recognition with Directional Decomposition Cellular Features[J],Journal of Circuits System and Computers,(1998)8,517~524
[23]高学,金连文,尹俊,一种基于笔画密度的弹性网格特征提取方法,模式识别与人工智能,(2002)15,351-354
[24]金连文,徐秉铮,手写体汉字识别中的一种新的特征提取方法-弹性网格方向分解特征[J],电路与系统学报,(1997)2,7~12
[25]钟国华,金连文,手写体汉字扇形弹性网格特征提取的新方法[J],计算机工程,(2002)28,61~62
[26]吴天雷,马少平,基于重叠动态网格和模糊隶属度的手写汉字特征抽取[J],电子学报,(2002)32,186~190
[27]陈友斌,丁晓青,一种手写汉字特征抽取新方法[J],信号处理,(1998)14,117~122
[28]居琰,汪同庆,刘建胜等,基于集成RBF神经网络的小类别手写体汉字识别系统[J],(2002)23,100~102
[29]居琰,汪同庆,刘建胜等,有限集手写体汉字特征提取及分类器设计[J],重庆大学学报(自然科学版),(2002)25,96-129
[30]金连文,覃剑钊,手写汉字识别弹性网格Gabor特征提取方法研究[J],计算机应用研究,(2003)12,163~165
[31]李玉静,扬扬,基于矩和Gabor变换的手写体汉字识别方法[J],信息技术,(2003)12,44-46
[32]王学文,丁晓青,基]~Gabor变换的高鲁棒汉字识别新方法[J],电子学报,(2002)9,1317-1322
[33]田学东,郭宝兰,基于Gabor变换的汉字字体识别研究[J],计算机工程,(2002)20,89~91
[34]李晓辉,翟娟娟,历史文献识别中图像预处理方法的研究[J],微机发展,(2004)14,30~31
[35]N.Otsu.Traps System,A threshold selection method from gray-level histograms[J],IEEE Trans System Man arid Cybernetics,(1979)9,62-66
[36]杨淑莹,VC++图像处理程序设计[M],北京,清华大学出版社,(2005)1,56-107
[37]张世辉,汉字图像预处理算法的研究及实现[J],微机发展,(2003)13,53~58
[38]Lianwen Jin,Gang Wei,Handwritten Chinese Character Recognition with Directional Decomposition Cellular Features[J],Journal of Circuit System and Computer,(1998)8,517~524
[39]徐丽娜,神经网络控制[M],哈尔滨,哈尔滨工业大学出版社,(1998)12,126~127
[40]Kimura F,Shridhar M,Handwritten numerical recognition based on multiple algorithms[J],Pattern Recognition,(1991)24,969-983
[41]韩力群,人工神经网络教程[M],北京,北京邮电大学出版社,(2006)12,22-80