美元纸币号码识别方法的研究
|
摘要
纸币上的编码是纸币的重要特征之一,其不仅是印刷数量的标识,而且在统计货币流通量,货币管理等方面也具有重要作用。每张纸币编码没有重复,因此可以用来标识纸币的身份;纸币上的币值号码表示纸币所代的货币价值。本文将纸币上的编码和币值号码统称为号码,利用计算机视觉技术开发一种智能纸币号码识别系统,自动识别纸币上的号码,就可以有效地实现对纸币的管理,具有广泛的应用价值。本文完成了一套纸币号码采集识别系统的设计与研制,包括图像分割、定位、识别算法,用C++Builder6和Matlab7.0.1语言编写了号码识别程序,建立了可运行的纸币号码识别系统。
本文通过CIS传感器进行图像采集,在图像预处理阶段,考虑到在实际应用中对识别的实时性要求很高,要求每一张纸币的所有识别过程必须在很短的时间内完成(实际系统要求30ms以内),因此放弃了诸如二值化、边缘检测等常规图像处理方法,而直接采用分段线性拟合的方法确定图像中纸币的四个边界,以完成纸币的定位和倾斜矫正。由于采集到的号码图像存在多种噪声,通过均值滤波和中值滤波等去噪声算法的对比和实验,选取了以中值滤波为主,高斯滤波辅助的方法去除噪声。
由于神经网络其固有的自组织、自适应、和容错性,只须重新训练就可以完成对新类别的正确识别,因此本文确定神经网络分类器作为分类器的首选加以研究和应用。本文在实际应用中引入了改进的BP网分类器对待识别的纸币样本做自组织特征映射,实现特征训练和匹配。实验表明,本文提出的算法能够满足识别的实时性要求,且较好的保证了识别率和稳定性。
在号码识别阶段,本文提出了在水平方向采用投影的方法进行定位,垂直方向采用改进的基于穿越号码体距离的方法进行垂直定位的方法,避免了双峰或多峰的干扰,定位号码准确。此方法处理速度快,识别纸币号码准确率高。
识别算法中除了采用神经网络进行识别外,还尝试了模板匹配的方法。传统的二维模板匹配虽然实现简单,但计算量十分庞大,花费时间太长。针对此情况提出了采用一维灰度投影的模板匹配方法。实验结果表明,后者在保证了匹配准确率略高的前提下,识别速度也明显高于前者。用BP神经网络的识别方法来提取纸币字符特征,与一维投影模板匹配识别一种面额需要考虑正面正向、正面反向、反面正向、反面反向相比,在识别率基本相同的情况下,当用于美元纸币识别时,推广能力明显高于后者。本文在基于弧体、圆体的结构识别基础上运用了基于穿越号码次数的结构识别方法。通过对300幅美元号码图像实验表明,识别率达98%。
The numbers on the paper currency is one of the important characters for it. It is not only the amount identifier of the paper currency issuance, but also for paper currency's turnover and manage has important effect, too. Each sheet has its own numbers and the same numbers haven't been used twice, so the numbers on it can identify paper currency's identity. In this paper the numbers on the paper currency and value number called "number". Developed a aptitude system by computer vision for paper currency numbers recognition, the number is recognized and recorded with binding by the system, then the organization of the paper currency could be easy. So the automatic register system has utilize tremendous and amplitude application foreground. In this paper, the automatic recognition system for paper currency numbers was built by the technique of image processing and pattern recognition.
Firstly, the system samples the images of paper currency passing quickly on the CIS. Then in pre-processing stage, for the allowance of the real time demand, we directly apply linear fitting to determine the border of paper currency to achieve localization and adjustment. The image gotten by CIS has many noises. Through the experiment of average filter and median filter, the algorithm of Gauss was adopted to dispel the noises.
Since the Artificial Neural Network has the property of self-organization, self-adaptation, and fault tolerance, it classifies the samples by training, the paper applies the ANN as the final classifier of samples. The system uses the modified BP network to classify the feature. The experiment shows that the algorithm in the paper is quite effective in recognition rate as well as real time property.
In the number identify phase, the paper advance in level location, the gray projection method was adopted, in perpendicular orientation, an improved method of distance of traversing number body was adopted. Then every number has a notable peak, and the two or more peaks made by projection directly is avoided. The algorithm is simply and fast enough to content the processing system.
In recognition algorithm we also try the method of template matching besides neural network. Although the traditional method of two-dimensional template matching
is simply achieved, the calculating quantities are much larger; it may take very long time. Under this circumstance, the author puts forward the template-matching algorithm based on one-dimensional gray projection. The result of experiments indicates that the latter on the premise that the right ratio of matching is higher, the recognition speed is obviously higher than the former, too. The method that based on BP neural network only considers two conveyed directions that are head and reverse because it adopts the technique of identify the character. Under the circumstance that recognition rates doesn't change, it has better ability of popularization than one-dimensional gray projection which need consider four directions such as head upright, head reverse, tail upright, and tail reverse when it is used in recognizing new kinds of dollar and other countries' banknote. A new stricture-analyzing algorithm was be put forward by calculate the transfer of the number body times.
Take the dollar as an example, 300 sheets of 100 par values was random chose, the recognition ratio is 98 percent.
本文通过CIS传感器进行图像采集,在图像预处理阶段,考虑到在实际应用中对识别的实时性要求很高,要求每一张纸币的所有识别过程必须在很短的时间内完成(实际系统要求30ms以内),因此放弃了诸如二值化、边缘检测等常规图像处理方法,而直接采用分段线性拟合的方法确定图像中纸币的四个边界,以完成纸币的定位和倾斜矫正。由于采集到的号码图像存在多种噪声,通过均值滤波和中值滤波等去噪声算法的对比和实验,选取了以中值滤波为主,高斯滤波辅助的方法去除噪声。
由于神经网络其固有的自组织、自适应、和容错性,只须重新训练就可以完成对新类别的正确识别,因此本文确定神经网络分类器作为分类器的首选加以研究和应用。本文在实际应用中引入了改进的BP网分类器对待识别的纸币样本做自组织特征映射,实现特征训练和匹配。实验表明,本文提出的算法能够满足识别的实时性要求,且较好的保证了识别率和稳定性。
在号码识别阶段,本文提出了在水平方向采用投影的方法进行定位,垂直方向采用改进的基于穿越号码体距离的方法进行垂直定位的方法,避免了双峰或多峰的干扰,定位号码准确。此方法处理速度快,识别纸币号码准确率高。
识别算法中除了采用神经网络进行识别外,还尝试了模板匹配的方法。传统的二维模板匹配虽然实现简单,但计算量十分庞大,花费时间太长。针对此情况提出了采用一维灰度投影的模板匹配方法。实验结果表明,后者在保证了匹配准确率略高的前提下,识别速度也明显高于前者。用BP神经网络的识别方法来提取纸币字符特征,与一维投影模板匹配识别一种面额需要考虑正面正向、正面反向、反面正向、反面反向相比,在识别率基本相同的情况下,当用于美元纸币识别时,推广能力明显高于后者。本文在基于弧体、圆体的结构识别基础上运用了基于穿越号码次数的结构识别方法。通过对300幅美元号码图像实验表明,识别率达98%。
The numbers on the paper currency is one of the important characters for it. It is not only the amount identifier of the paper currency issuance, but also for paper currency's turnover and manage has important effect, too. Each sheet has its own numbers and the same numbers haven't been used twice, so the numbers on it can identify paper currency's identity. In this paper the numbers on the paper currency and value number called "number". Developed a aptitude system by computer vision for paper currency numbers recognition, the number is recognized and recorded with binding by the system, then the organization of the paper currency could be easy. So the automatic register system has utilize tremendous and amplitude application foreground. In this paper, the automatic recognition system for paper currency numbers was built by the technique of image processing and pattern recognition.
Firstly, the system samples the images of paper currency passing quickly on the CIS. Then in pre-processing stage, for the allowance of the real time demand, we directly apply linear fitting to determine the border of paper currency to achieve localization and adjustment. The image gotten by CIS has many noises. Through the experiment of average filter and median filter, the algorithm of Gauss was adopted to dispel the noises.
Since the Artificial Neural Network has the property of self-organization, self-adaptation, and fault tolerance, it classifies the samples by training, the paper applies the ANN as the final classifier of samples. The system uses the modified BP network to classify the feature. The experiment shows that the algorithm in the paper is quite effective in recognition rate as well as real time property.
In the number identify phase, the paper advance in level location, the gray projection method was adopted, in perpendicular orientation, an improved method of distance of traversing number body was adopted. Then every number has a notable peak, and the two or more peaks made by projection directly is avoided. The algorithm is simply and fast enough to content the processing system.
In recognition algorithm we also try the method of template matching besides neural network. Although the traditional method of two-dimensional template matching
is simply achieved, the calculating quantities are much larger; it may take very long time. Under this circumstance, the author puts forward the template-matching algorithm based on one-dimensional gray projection. The result of experiments indicates that the latter on the premise that the right ratio of matching is higher, the recognition speed is obviously higher than the former, too. The method that based on BP neural network only considers two conveyed directions that are head and reverse because it adopts the technique of identify the character. Under the circumstance that recognition rates doesn't change, it has better ability of popularization than one-dimensional gray projection which need consider four directions such as head upright, head reverse, tail upright, and tail reverse when it is used in recognizing new kinds of dollar and other countries' banknote. A new stricture-analyzing algorithm was be put forward by calculate the transfer of the number body times.
Take the dollar as an example, 300 sheets of 100 par values was random chose, the recognition ratio is 98 percent.
引文
[1]周翟和,刘建业,肖纪立.纸币号码读入识别系统的单片机实现.工业控制计算机,2002,15(1):69-71
[2]殷瑞祥,李国华.身份证号码的自动识别系统.华南理工大学学报,2002,30(2):94-96
[3]吕松茂,韩震宇,吕津等.身份证号码的快速识别系统.实用测试技术,2001,(1):31-33
[4]李雪巍,孙玉兰,吴胜权等.机车车厢号码的自动识别系统.测试技术学报,1996,(3):480-484
[5]周玉玲.货车车厢号码CCD自动识别系统.西安工业学院学报,1988,(2):99-104
[6]杨波,彭健,叶俊勇等.枪号识别系统的设计.重庆大学学报,2001,24(5):137-141
[7]刘振杰,乐宁,王海丽等.汽车变速箱壳体号码识别中输入输出控制系统的研制.机械设计与制造工程,2000,29(1):40-42
[8]王砚坤.机号识别的若干问题的研究,沈阳工业大学学报,2001,(1):55-58
[9]张宏林.Visual C++数字图像模式识别技术及工程实践.北京:人民邮电出版社,2002
[10]王晓华,高庆吉,赵卫平.利用模糊统计法识别货车列车车号的研究.中国图像图形学报,2000,5(11):963-967
[11]Fumiaki Takeda,Sigeru Omatu.High speed paper currency recognition by neural networks.IEEE Trans on Neural Network,1995,6(1):73-77
[12]谭锐,宣国荣.切线距离在印刷体数字识别中的应用.微型电脑应用,2002,18(12):7-8
[13]段会川,张海波,张曙光.基于模糊理论的仪表数字识别研究.仪表技术与传感器,2004,(4):37-39
[14]F.Takeda,S.Omatu,T.Inoue etc.High-speed Conveyed Bill Money Recongniotin with Neural Networks,in Proc.IMACS/SCINE Int.Symp.Robot,Mechatron.Manufact.Syst.,Kobe,Japan,1992:16-20
[15]F.Takeda,S.Omatu,T.Inoue etc.Bill Money Recognition using Neural Network with FFT as Pre-processor.Trans.Inst.Syst,Control and info.Eng,Japan,1992,5(7):265-273
[16]徐勇等译.神经网络棋式识别及其实现.电子工业出版社,99-6-1,第一版
[17]Sonehara.An application of neural network to image data.Trans.Inst.Syst,Control and Info,Eng,Vol.35,N0.1,pp.11-18,1991,Japan
[18]S.Amari.Yhe Present and Future of Neural Network.Kyoritsu Press,1990,in Japan
[19]戴葵.神经网络实现技术.国防科技大学出版社,1998年,第一版
[20]Fumiaki Takeda,Sigeru Omatu.High Speed Paper Currency Recognition by Neural Networks.IEEE Trans Neural Network,1995,Vo 16,No 1,Jan
[21]朱方.基于彩色图像处理的车牌识别系统的研究与开发:[硕士学位论文].西安:西安交通大学,2001
[22]容观澳.计算机图像处理.北京:清华大学出版社,2000
[23]夏良正.数字图像处理.南京:东南大学出版社,1999
[24]张兆礼,赵春晖,梅晓丹.现代图像处理技术及Matlab实现.北京:人民邮电出版社,2001
[26]何斌,马天予,王运坚等.VC++数字图像处理,第2版.北京:人民邮电出版社,2003
[27]宋铭利.关于去除图像噪音的中值滤波算法.洛阳师范学院学报,2002(5):67-69
[28]高湘玲.PCB数字图像检测与识别研究:[硕士学位论文].成都:西南交通大学,2002
[29]权炜,郑南宁,贾新春.复杂背景下车辆牌照字符提取方法研究.信息与控制,2002,31(1):25-29
[30]程杰.一种基于直方图的分割方法.华中理工大学学报,1999,27(1):20-23
[31]F.Takeda,S.Omatu,Y.Matsumoto.Development of High Speed Neuro-Recogniton Board and Application for Paper Currency.The International Workshop on Signal Processing Application and Technology,1998,49-56
[32]Mauro Barini,Rossana Gualtieri.A new possibalistic clustering algorithm for line detection in real world image.Pattern Recognition,Vol.32,No.11,Nov.1999
[33]Anca L.Ralescu,James G.Shanahan.Perceptual organization for inferring object boundaries in an image.Pattern Recognition,Vol.32,No.11,Nov.1999
[34]J.Illing Worth,J.Kittler.A Survey of the Hough Transform.Compute Vision and Image Process,44(1988):87-116
[35]V.F.Leavers.Which Hough Transform.Compute Vision and Image Process,58(1993)
[36]沈清,汤霖.模式识别导论.北京:国防科技大学出版社,1991
[37]郑霞.车牌照识别关键技术研究:[硕士学位论文],西安:西北大学计算机软件与理论,2004
[38]谢鸣.基于形态学的车牌自动识别系统的研究:[硕士学位论文].武汉:武汉理工 大学通信与信息系统,2002
[39]陶军.车牌识别技术研究与实现:[硕士学位论文].南京:南京理工大学模式识别与智能系统,2004
[40]周翟和.纸币读入识别系统的研究和实现:[硕士学位论文].南京:南京航空航天大学,2002,3
[41]哈勤和,张畅,王泽宾等.图像的快速一维模板匹配.计算机工程与应用,1998(1):36-37
[42]桂志国,薄瑞峰,韩焱.基于投影特征的图像匹配的快速算法.华北工学院测试技术学报,2000,14(1),18-20
[43]孙远,周刚慧,赵立初等.灰度图像匹配的快速算法.上海交通大学学报,2000,34(5),702-704
[44]王强,宋京民,胡建平.一种快速模板匹配目标识别算法.计算机工程与应用,2000(5):42-43
[45]赵楠楠.人民币图像识别系统:[硕士学位论文].鞍山:鞍山科技大学,2000,3
[2]殷瑞祥,李国华.身份证号码的自动识别系统.华南理工大学学报,2002,30(2):94-96
[3]吕松茂,韩震宇,吕津等.身份证号码的快速识别系统.实用测试技术,2001,(1):31-33
[4]李雪巍,孙玉兰,吴胜权等.机车车厢号码的自动识别系统.测试技术学报,1996,(3):480-484
[5]周玉玲.货车车厢号码CCD自动识别系统.西安工业学院学报,1988,(2):99-104
[6]杨波,彭健,叶俊勇等.枪号识别系统的设计.重庆大学学报,2001,24(5):137-141
[7]刘振杰,乐宁,王海丽等.汽车变速箱壳体号码识别中输入输出控制系统的研制.机械设计与制造工程,2000,29(1):40-42
[8]王砚坤.机号识别的若干问题的研究,沈阳工业大学学报,2001,(1):55-58
[9]张宏林.Visual C++数字图像模式识别技术及工程实践.北京:人民邮电出版社,2002
[10]王晓华,高庆吉,赵卫平.利用模糊统计法识别货车列车车号的研究.中国图像图形学报,2000,5(11):963-967
[11]Fumiaki Takeda,Sigeru Omatu.High speed paper currency recognition by neural networks.IEEE Trans on Neural Network,1995,6(1):73-77
[12]谭锐,宣国荣.切线距离在印刷体数字识别中的应用.微型电脑应用,2002,18(12):7-8
[13]段会川,张海波,张曙光.基于模糊理论的仪表数字识别研究.仪表技术与传感器,2004,(4):37-39
[14]F.Takeda,S.Omatu,T.Inoue etc.High-speed Conveyed Bill Money Recongniotin with Neural Networks,in Proc.IMACS/SCINE Int.Symp.Robot,Mechatron.Manufact.Syst.,Kobe,Japan,1992:16-20
[15]F.Takeda,S.Omatu,T.Inoue etc.Bill Money Recognition using Neural Network with FFT as Pre-processor.Trans.Inst.Syst,Control and info.Eng,Japan,1992,5(7):265-273
[16]徐勇等译.神经网络棋式识别及其实现.电子工业出版社,99-6-1,第一版
[17]Sonehara.An application of neural network to image data.Trans.Inst.Syst,Control and Info,Eng,Vol.35,N0.1,pp.11-18,1991,Japan
[18]S.Amari.Yhe Present and Future of Neural Network.Kyoritsu Press,1990,in Japan
[19]戴葵.神经网络实现技术.国防科技大学出版社,1998年,第一版
[20]Fumiaki Takeda,Sigeru Omatu.High Speed Paper Currency Recognition by Neural Networks.IEEE Trans Neural Network,1995,Vo 16,No 1,Jan
[21]朱方.基于彩色图像处理的车牌识别系统的研究与开发:[硕士学位论文].西安:西安交通大学,2001
[22]容观澳.计算机图像处理.北京:清华大学出版社,2000
[23]夏良正.数字图像处理.南京:东南大学出版社,1999
[24]张兆礼,赵春晖,梅晓丹.现代图像处理技术及Matlab实现.北京:人民邮电出版社,2001
[26]何斌,马天予,王运坚等.VC++数字图像处理,第2版.北京:人民邮电出版社,2003
[27]宋铭利.关于去除图像噪音的中值滤波算法.洛阳师范学院学报,2002(5):67-69
[28]高湘玲.PCB数字图像检测与识别研究:[硕士学位论文].成都:西南交通大学,2002
[29]权炜,郑南宁,贾新春.复杂背景下车辆牌照字符提取方法研究.信息与控制,2002,31(1):25-29
[30]程杰.一种基于直方图的分割方法.华中理工大学学报,1999,27(1):20-23
[31]F.Takeda,S.Omatu,Y.Matsumoto.Development of High Speed Neuro-Recogniton Board and Application for Paper Currency.The International Workshop on Signal Processing Application and Technology,1998,49-56
[32]Mauro Barini,Rossana Gualtieri.A new possibalistic clustering algorithm for line detection in real world image.Pattern Recognition,Vol.32,No.11,Nov.1999
[33]Anca L.Ralescu,James G.Shanahan.Perceptual organization for inferring object boundaries in an image.Pattern Recognition,Vol.32,No.11,Nov.1999
[34]J.Illing Worth,J.Kittler.A Survey of the Hough Transform.Compute Vision and Image Process,44(1988):87-116
[35]V.F.Leavers.Which Hough Transform.Compute Vision and Image Process,58(1993)
[36]沈清,汤霖.模式识别导论.北京:国防科技大学出版社,1991
[37]郑霞.车牌照识别关键技术研究:[硕士学位论文],西安:西北大学计算机软件与理论,2004
[38]谢鸣.基于形态学的车牌自动识别系统的研究:[硕士学位论文].武汉:武汉理工 大学通信与信息系统,2002
[39]陶军.车牌识别技术研究与实现:[硕士学位论文].南京:南京理工大学模式识别与智能系统,2004
[40]周翟和.纸币读入识别系统的研究和实现:[硕士学位论文].南京:南京航空航天大学,2002,3
[41]哈勤和,张畅,王泽宾等.图像的快速一维模板匹配.计算机工程与应用,1998(1):36-37
[42]桂志国,薄瑞峰,韩焱.基于投影特征的图像匹配的快速算法.华北工学院测试技术学报,2000,14(1),18-20
[43]孙远,周刚慧,赵立初等.灰度图像匹配的快速算法.上海交通大学学报,2000,34(5),702-704
[44]王强,宋京民,胡建平.一种快速模板匹配目标识别算法.计算机工程与应用,2000(5):42-43
[45]赵楠楠.人民币图像识别系统:[硕士学位论文].鞍山:鞍山科技大学,2000,3