基于LVQ神经网络的手写英文字母识别
|
摘要
随着社会的发展,英语作为国际通用语言得到了日益广泛的应用,因此有大量的英文文档整理、查询、统计的工作需要完成,而英文文档识别系统可以轻而易举地完成很多以前难以想象的工作。一个完整的英文文档识别系统,包括文档切分、单词切分、单字母识别、单词识别、后处理等部分。其中单个字母的字符识别,是整个识别系统的核心。通常对英文文档的识别,都是基于单词。单词识别少数方法是以单词整体为识别对象,大部分是基于单字符识别,即先对单词进行切分,得到字符,然后再进行字符识别。
手写英文字母识别系统主要分为三个主要部分,分别是图像预处理,特征提取和判断识别。本文主要考虑特征提取和判断识别部分,特征提取主要提取统计特征和结构特征,并把它们相互融合得到训练与识别所需特征;判断识别采用常用的人工神经网络的识别方法,本文根据手写体字母的特点选取了LVQ神经网络对其进行识别,并对其作进一步的改进。
本文以一个26万左右的手写体字母图库作为研究对象,根据需要提取训练样本和测试样本。因为字母库中的图像还没有达到提取特征和输入识别的要求,所以要作适当的预处理。首先是字母图像大小不一,要进行归一化处理;其次是提取特征时需要图像的骨架和轮廓,要进行细化处理和轮廓跟踪。在细化和轮廓跟踪中本人对算法进行改进,使得字母图像在预处理后的效果更好,有助于提取更好的特征,提高识别率。
在预处理完成后要提取特征,本文对原图像提取投影特征和粗网格特征,对骨架图像提取方向特征,最后对轮廓提取外围特征。其中方向特征和粗网格特征是局部描述的统计特征,投影特征是全局描述的统计特征,而外轮廓特征是一种结构分析特征。本文将这三种特征相融合输入神经网络进行训练和识别,在实验中达到了一个很好的效果。
最后对神经网络进行训练识别。本文从字母库中随机抽取2600个字母图片作为训练样本,并抽取不同于训练样本的1040个作为测试样本。把训练样本和测试样本输入LVQ进行训练和识别,并与BP神经网络进行比较。
实验证明,LVQ神经网络与BP网络相比,结构简单,只需要三层网络就可以实现模式识别;不需要将输入向量进行归一化、正交化,只需要直接计算输入向量与竞争层之间的距离,从而实现模式识别;且收敛速度比BP网络更快,不存在BP网络的有可能陷入局部最小问题;也就是说,LVQ神经网络不但简单易行,而且识别效率更高。因此,采用LVQ网络进行手写字母识别是合适的。
With the development of society, English as an international common language has been increasingly widely used, so a large number of documents in English need to be tidied up , queried and statistic, and the English Document Recognition system can easily complete many Unimaginable work before. A complete English Document Recognition system is composed of documents segmentation, word segmentation, single-letter recognition, word recognition and lastly processing. And the recognition system for single letter is the core. The recognition of English document is usually based on words. And a small number of methods are depended on the object of the whole words, but most the single-character recognition, in other words you must gain the character after the segment of the word, and then recognize it.
Handwriting Recognition system for English letters can be divided into three main modules, namely the image preprocessing, feature extraction and identify. In this paper, the main consideration is feature extraction and recognition. The mainly work of feature extraction is to extract structure and statistics characteristics, then integrate them, and get that is needed for the training and recognition work in the end. The artificial neural network identification methods are commonly used for the recognition work. In this paper, according to the characteristics of the handwriting letters I have chose the LVQ neural network, and made to their further improvement.
In this paper, I use a research library of about 260,000 of a handwritten letters, and as required from training samples and test samples. Because the images of the letters library have not yet reached the feature extraction request, so I have to make appropriate pretreatment. First, the images are not the same size, and need to conduct a naturalization processing. Secondly it is necessary to do a Thinning Treatment and Contour Tracking work by the need to extract characteristics from their final images. In the procedure, I improve the algorithms, making the letter image better after pretreatment, help to extract better features, enhance recognition rate.
After the pretreatment, extracting the features is need. The Projection features and Coarse Grid features are extracted form the original images while the Direction from the skeleton ones, and External from the Contour. And Direction features and Coarse Grid is a partial description of the features of the demographic characteristics, the Projection is characterized by overall description of the demographic characteristics, External Contour characteristic features of structural analysis. In this paper, these three characteristics are integrated in order to import to the neural network for training and recognition, and eventually in the experiment has reached a very good results.
Finally, the neural network is trained to identify. In this article, 2,600 letters picture are randomly selected from the letters library as training samples and 1,040 other samples for testing. The samples are then imported to LVQ neural network for training and recognition, and compare to BP neural network.
It is proved that, compared with BP network, simple structure, the LVQ neural network, just three-tier network, can achieved pattern recognition; and it does not need the input vector for normalization of orthogonal, only need to calculate the distance between the entering vector and competition layer so as to achieve pattern recognition; besides, the convergence rate is faster than the BP network, and the network does not exist the possibility of problem that BP may have a local minimum; in other words, LVQ neural network is not only simple, but also more efficient to identify. Therefore, the use of LVQ network for handwritten letter recognition is appropriate.
手写英文字母识别系统主要分为三个主要部分,分别是图像预处理,特征提取和判断识别。本文主要考虑特征提取和判断识别部分,特征提取主要提取统计特征和结构特征,并把它们相互融合得到训练与识别所需特征;判断识别采用常用的人工神经网络的识别方法,本文根据手写体字母的特点选取了LVQ神经网络对其进行识别,并对其作进一步的改进。
本文以一个26万左右的手写体字母图库作为研究对象,根据需要提取训练样本和测试样本。因为字母库中的图像还没有达到提取特征和输入识别的要求,所以要作适当的预处理。首先是字母图像大小不一,要进行归一化处理;其次是提取特征时需要图像的骨架和轮廓,要进行细化处理和轮廓跟踪。在细化和轮廓跟踪中本人对算法进行改进,使得字母图像在预处理后的效果更好,有助于提取更好的特征,提高识别率。
在预处理完成后要提取特征,本文对原图像提取投影特征和粗网格特征,对骨架图像提取方向特征,最后对轮廓提取外围特征。其中方向特征和粗网格特征是局部描述的统计特征,投影特征是全局描述的统计特征,而外轮廓特征是一种结构分析特征。本文将这三种特征相融合输入神经网络进行训练和识别,在实验中达到了一个很好的效果。
最后对神经网络进行训练识别。本文从字母库中随机抽取2600个字母图片作为训练样本,并抽取不同于训练样本的1040个作为测试样本。把训练样本和测试样本输入LVQ进行训练和识别,并与BP神经网络进行比较。
实验证明,LVQ神经网络与BP网络相比,结构简单,只需要三层网络就可以实现模式识别;不需要将输入向量进行归一化、正交化,只需要直接计算输入向量与竞争层之间的距离,从而实现模式识别;且收敛速度比BP网络更快,不存在BP网络的有可能陷入局部最小问题;也就是说,LVQ神经网络不但简单易行,而且识别效率更高。因此,采用LVQ网络进行手写字母识别是合适的。
With the development of society, English as an international common language has been increasingly widely used, so a large number of documents in English need to be tidied up , queried and statistic, and the English Document Recognition system can easily complete many Unimaginable work before. A complete English Document Recognition system is composed of documents segmentation, word segmentation, single-letter recognition, word recognition and lastly processing. And the recognition system for single letter is the core. The recognition of English document is usually based on words. And a small number of methods are depended on the object of the whole words, but most the single-character recognition, in other words you must gain the character after the segment of the word, and then recognize it.
Handwriting Recognition system for English letters can be divided into three main modules, namely the image preprocessing, feature extraction and identify. In this paper, the main consideration is feature extraction and recognition. The mainly work of feature extraction is to extract structure and statistics characteristics, then integrate them, and get that is needed for the training and recognition work in the end. The artificial neural network identification methods are commonly used for the recognition work. In this paper, according to the characteristics of the handwriting letters I have chose the LVQ neural network, and made to their further improvement.
In this paper, I use a research library of about 260,000 of a handwritten letters, and as required from training samples and test samples. Because the images of the letters library have not yet reached the feature extraction request, so I have to make appropriate pretreatment. First, the images are not the same size, and need to conduct a naturalization processing. Secondly it is necessary to do a Thinning Treatment and Contour Tracking work by the need to extract characteristics from their final images. In the procedure, I improve the algorithms, making the letter image better after pretreatment, help to extract better features, enhance recognition rate.
After the pretreatment, extracting the features is need. The Projection features and Coarse Grid features are extracted form the original images while the Direction from the skeleton ones, and External from the Contour. And Direction features and Coarse Grid is a partial description of the features of the demographic characteristics, the Projection is characterized by overall description of the demographic characteristics, External Contour characteristic features of structural analysis. In this paper, these three characteristics are integrated in order to import to the neural network for training and recognition, and eventually in the experiment has reached a very good results.
Finally, the neural network is trained to identify. In this article, 2,600 letters picture are randomly selected from the letters library as training samples and 1,040 other samples for testing. The samples are then imported to LVQ neural network for training and recognition, and compare to BP neural network.
It is proved that, compared with BP network, simple structure, the LVQ neural network, just three-tier network, can achieved pattern recognition; and it does not need the input vector for normalization of orthogonal, only need to calculate the distance between the entering vector and competition layer so as to achieve pattern recognition; besides, the convergence rate is faster than the BP network, and the network does not exist the possibility of problem that BP may have a local minimum; in other words, LVQ neural network is not only simple, but also more efficient to identify. Therefore, the use of LVQ network for handwritten letter recognition is appropriate.
引文
[1]荆涛,王仲,光学字符识别技术与展望[J].计算机工程,2003,2:1-2.
[2]胡小锋,赵辉.图像处理与识别实用案例精选[M].北京:人民邮电出版社,2004.264-384
[3]白利达.神经网络集成识别手写体数字研究[D]:[硕士学位论文]。保存地点:北京邮电大学,2007
[4]李冰.基于多神经网络集成的手写体字符识别[D]:[硕士学位论文]。保存地点:华中科技大学,2005
[5]飞思科技产品研发中心.神经网络理论与MATLAB7实现[M].北京:电子工业出版社,2005.183-189
[6]何东健,耿楠,张义宽.数字图像处理[M].西安:西安电子科技大学出版社,2003.193-202
[7]欧阳应华.一种基于特征提取的脱机手写汉字识别技术[D]:[硕士学位论文].保存地点:兰州大学,2007
[8]吕俊白.一种有效的二值图像细化算法[J].计算机工程,2003.10,29(18):147-148.
[9]ZHANG Y Y,SUEN C Y.A fast parallel algorithm for thinning digital pattern[J].Communications of the ACM,1984,27(6):236-539.
[10]王家隆,郭成安.一种改进的图像模板细化算法[J].中国图形图像学报,2004,9(3):297-301
[11]LUHE,WANG P S P.An improved fast parallel algorithm for thinning digital patterns[C]//Proc IEEE Conference on Computer Vision and Pattern Recognition.1985:364-367.
[12]Shaikh,N.A,Shaikh,Z.A.A generalized thinning algorithm for cursive and non-cursive language scripts.[C]2005 Pakistan Section Multitopic Conference(IEEE Cat.No.O5EXII83).2005,4pp.
[13]包建军,樊菁.鲁棒的二值图像并行细化算法[J].计算机辅助工程,2006,15(4):43-46.
[14]苏彦华等.数字图像识别技术典型案例[M].北京:人民邮电出版社,2004.456-639
[15]吴佑寿,丁晓青.汉字识别原理方法与实现[M].北京:高等教育出版社.1992.1-10.
[16]张忻中.汉字识别技术[M].北京:清华大学出版社,1992.1-8.
[17]厄尔曼著,刘定一译.文字、图形识别技术[M].北京:人民邮电出版社,1983.1-7.
[18]Martin T.Hagan Howard B.Demuth Mark H.Beale.Neural Network Design[M].戴葵.北京:机械工业出版社,2005.285-310
[19]徐飞.施晓红等.MATLAB应用图像处理[M].西安:西安电子科技大学出版社,2003.216-217
[20]缪青.基于LVQ模型的模式分类方法及其在大规模学习问题中的应用[D]:[硕士学位论文].
[21]杨兴炜,刘文子,白翔.一种有效的快速细化算法[J].小型微型计算机系统,2006,27(7):1343-1346.
[22]张平,潘保昌,汪同庆等.一种自由手写体数字识别方法研究[J].光电工程,1995,3(22):43-46
[23]N.Ariea,F.T.Yarman-Vural.Optical character recognition for cursive handwriting[J].IEEE Trans.Pattern Analysis and Machine Intelligence,2002,24(6):801-813.
[24]Andrew W Senior et al.An Of-Line Cursive Handwriting Recognition System[J].IEEE Tran on PAMI,1998;20(3):309-321
[25]P V S Rao.A Knowledge-Based Approach for Script Recognition without Training[J].IEEE Tran on PAMI,1995;17(12):1233-1239
[26]Christophe Parisse.Global Word Shape Processing in Of-Line Recognition of Handwriting[J].IEEE Tran on PAMI,1996;18(4): 460-464
[27]Alessandro Vineiarelli.A Survey on Of-line Cursive Word Recognition[J].Pattern Recognition,2002;35:1433-1446
[28]Radmilo M Bozinovie.SargurNSrihari.Of-Line Cursive Script Word Recognition[J].IEEE Transaction on Pattern Analysis and Machine Intelligence,1989;11(1):68-83
[29]伍振军,丁晓青.鲁棒的多体印刷英文识别系统的实现[J].计算机工程与应用,2001;37(20):120-122
[30]吴锐,赵巍,尹芳等.特征融合及相似度判据在英文识别中的应用[J].计算机工程与应用,2005:55-57
[31]郑南宁,王龙,胡超等.BP神经网络的改进及其用于手写数字识别的研究[J].西安交通大学学报,1992,2:1-12
[32]严国莉,黄山等,印刷体数字快速识别算法在身份证编号数字识别中的应用[J].计算机工程,2003,1:178-179
[33]陈国平,张明新,付跃文等,高性能的多体印刷英文识别系统的实现[J].计算机工程与应用,2006,12:183-186
[34]柳回春,马树元,吴平东等,手写体数字识别技术的研究[J].计算机工程,2003,3:24-61.
[35]周治紧,李玉槛,基于投影归一化的字符特征提取方法[J].计算机工程,2006,1:197-199.
[36]杜彦蕊,李珍,宁伟宏,基于特征编码的手写字符识别技术[J].计算机工程,2004,3:156-158.
[37]武强,童学锋,季隽,基于人工神经网络的数字字符识别[J].计算机工程,2003,8:112-132.
[38]Martin T.Hagan,Howard B.Demuth,Mark H.Beale.Neural Network Design[J].PWS Publishing Company.1996
[39]GAADER P.Recognition of handwritten digits using template and model matching[J].Pattern recognition,1991,24(5)421-431
[40]SUEN C Y,Legault R,Nadal Cetal.Building a mew generation of handwriting recognition system[J].Pattern Rec2ognition Letters,1993,14(4):303-315
[41]ALMALLIM H,YAMAGUCHI S.A method of recognition of arabic cursive handwriting[J],IEEE Trans Patternand Machine Intelligence,1987,PAMI-9(5):715-722
[42]求是科技张宏林Visual C++数字图像模式识别技术及工程实践[M].北京:人民邮电出版社,2003
[43]焦李成著.神经网络系统理论[M].西安:西安电子科技大学出版社1992.34-41
[44]刘云飞.脱机手写体汉字识别中细化、特征提取和相似字识别算法研究[D]:[硕士学位论文]。保存地点:湖南大学,2006
[45]袁氢.基于特征融合与神经网络的手写体数字识别技术研究[D]:[硕士学位论文].保存地点:武汉科技大学,2007
[46]白利达.神经网络集成识别手写体数字研究[D]:[硕士学位论文]。保存地点:北京邮电大学,2007
[2]胡小锋,赵辉.图像处理与识别实用案例精选[M].北京:人民邮电出版社,2004.264-384
[3]白利达.神经网络集成识别手写体数字研究[D]:[硕士学位论文]。保存地点:北京邮电大学,2007
[4]李冰.基于多神经网络集成的手写体字符识别[D]:[硕士学位论文]。保存地点:华中科技大学,2005
[5]飞思科技产品研发中心.神经网络理论与MATLAB7实现[M].北京:电子工业出版社,2005.183-189
[6]何东健,耿楠,张义宽.数字图像处理[M].西安:西安电子科技大学出版社,2003.193-202
[7]欧阳应华.一种基于特征提取的脱机手写汉字识别技术[D]:[硕士学位论文].保存地点:兰州大学,2007
[8]吕俊白.一种有效的二值图像细化算法[J].计算机工程,2003.10,29(18):147-148.
[9]ZHANG Y Y,SUEN C Y.A fast parallel algorithm for thinning digital pattern[J].Communications of the ACM,1984,27(6):236-539.
[10]王家隆,郭成安.一种改进的图像模板细化算法[J].中国图形图像学报,2004,9(3):297-301
[11]LUHE,WANG P S P.An improved fast parallel algorithm for thinning digital patterns[C]//Proc IEEE Conference on Computer Vision and Pattern Recognition.1985:364-367.
[12]Shaikh,N.A,Shaikh,Z.A.A generalized thinning algorithm for cursive and non-cursive language scripts.[C]2005 Pakistan Section Multitopic Conference(IEEE Cat.No.O5EXII83).2005,4pp.
[13]包建军,樊菁.鲁棒的二值图像并行细化算法[J].计算机辅助工程,2006,15(4):43-46.
[14]苏彦华等.数字图像识别技术典型案例[M].北京:人民邮电出版社,2004.456-639
[15]吴佑寿,丁晓青.汉字识别原理方法与实现[M].北京:高等教育出版社.1992.1-10.
[16]张忻中.汉字识别技术[M].北京:清华大学出版社,1992.1-8.
[17]厄尔曼著,刘定一译.文字、图形识别技术[M].北京:人民邮电出版社,1983.1-7.
[18]Martin T.Hagan Howard B.Demuth Mark H.Beale.Neural Network Design[M].戴葵.北京:机械工业出版社,2005.285-310
[19]徐飞.施晓红等.MATLAB应用图像处理[M].西安:西安电子科技大学出版社,2003.216-217
[20]缪青.基于LVQ模型的模式分类方法及其在大规模学习问题中的应用[D]:[硕士学位论文].
[21]杨兴炜,刘文子,白翔.一种有效的快速细化算法[J].小型微型计算机系统,2006,27(7):1343-1346.
[22]张平,潘保昌,汪同庆等.一种自由手写体数字识别方法研究[J].光电工程,1995,3(22):43-46
[23]N.Ariea,F.T.Yarman-Vural.Optical character recognition for cursive handwriting[J].IEEE Trans.Pattern Analysis and Machine Intelligence,2002,24(6):801-813.
[24]Andrew W Senior et al.An Of-Line Cursive Handwriting Recognition System[J].IEEE Tran on PAMI,1998;20(3):309-321
[25]P V S Rao.A Knowledge-Based Approach for Script Recognition without Training[J].IEEE Tran on PAMI,1995;17(12):1233-1239
[26]Christophe Parisse.Global Word Shape Processing in Of-Line Recognition of Handwriting[J].IEEE Tran on PAMI,1996;18(4): 460-464
[27]Alessandro Vineiarelli.A Survey on Of-line Cursive Word Recognition[J].Pattern Recognition,2002;35:1433-1446
[28]Radmilo M Bozinovie.SargurNSrihari.Of-Line Cursive Script Word Recognition[J].IEEE Transaction on Pattern Analysis and Machine Intelligence,1989;11(1):68-83
[29]伍振军,丁晓青.鲁棒的多体印刷英文识别系统的实现[J].计算机工程与应用,2001;37(20):120-122
[30]吴锐,赵巍,尹芳等.特征融合及相似度判据在英文识别中的应用[J].计算机工程与应用,2005:55-57
[31]郑南宁,王龙,胡超等.BP神经网络的改进及其用于手写数字识别的研究[J].西安交通大学学报,1992,2:1-12
[32]严国莉,黄山等,印刷体数字快速识别算法在身份证编号数字识别中的应用[J].计算机工程,2003,1:178-179
[33]陈国平,张明新,付跃文等,高性能的多体印刷英文识别系统的实现[J].计算机工程与应用,2006,12:183-186
[34]柳回春,马树元,吴平东等,手写体数字识别技术的研究[J].计算机工程,2003,3:24-61.
[35]周治紧,李玉槛,基于投影归一化的字符特征提取方法[J].计算机工程,2006,1:197-199.
[36]杜彦蕊,李珍,宁伟宏,基于特征编码的手写字符识别技术[J].计算机工程,2004,3:156-158.
[37]武强,童学锋,季隽,基于人工神经网络的数字字符识别[J].计算机工程,2003,8:112-132.
[38]Martin T.Hagan,Howard B.Demuth,Mark H.Beale.Neural Network Design[J].PWS Publishing Company.1996
[39]GAADER P.Recognition of handwritten digits using template and model matching[J].Pattern recognition,1991,24(5)421-431
[40]SUEN C Y,Legault R,Nadal Cetal.Building a mew generation of handwriting recognition system[J].Pattern Rec2ognition Letters,1993,14(4):303-315
[41]ALMALLIM H,YAMAGUCHI S.A method of recognition of arabic cursive handwriting[J],IEEE Trans Patternand Machine Intelligence,1987,PAMI-9(5):715-722
[42]求是科技张宏林Visual C++数字图像模式识别技术及工程实践[M].北京:人民邮电出版社,2003
[43]焦李成著.神经网络系统理论[M].西安:西安电子科技大学出版社1992.34-41
[44]刘云飞.脱机手写体汉字识别中细化、特征提取和相似字识别算法研究[D]:[硕士学位论文]。保存地点:湖南大学,2006
[45]袁氢.基于特征融合与神经网络的手写体数字识别技术研究[D]:[硕士学位论文].保存地点:武汉科技大学,2007
[46]白利达.神经网络集成识别手写体数字研究[D]:[硕士学位论文]。保存地点:北京邮电大学,2007