自动车牌识别系统
自动车牌识别系统[20200907152442]
摘要:本文提出板式鉴定和认可的视频图像的软件系统.
关键字:车牌,Hough变换,数字滤波,光学字符识别,链码.
I..引言
自动检测和识别车牌号,已成为人工视觉系统的I.个重要应用[I.-VIII].目的是建立I.个制度,让汽车通过某I.个点进行数字拍照,然后通过定位车牌图像中,从位于板分割字符,然后确认他们的电子标识.某些应用程序的车牌识别系统主要有:I.)流量的流量测量和规划;II)跟踪被盗车辆;III)控制和安全性收费领域,如停车库;IV)交通执法(自动识别速度ERS,违法停车等);V)该系统也可以适用于在读如使用;VI)仓储箱模板代码;VII)火车机车码;VIII)飞机尾码.
此系统包括两个高级别阶段:在第I.阶段中,号牌被检测并从车上被检查的数字图像分割.该板的位置和大小被传递到光学字符识别(OCR)的子系统.
图I.表明了该车牌识别的过程.有迹象表明,该软件需要识别车牌VI个主要的算法:
I.)板的定位-负责发现和隔离板上的图片;
II)板的方向和大小-用于补偿板的倾斜并调整尺寸到所需的大小;
III)正常化-调整亮度和图象的对比度;
IV)字符分割,发现在板上的单个字符;
V)光学字符识别;
VI)句法/几何分析-检查字符和反对国家的具体规则职位.
图I.建议车牌识别过程图
方法都是关于车牌的形状和外观的几个假设.假设如下:
I.)车牌是I.个容易辨别的颜色的矩形区域;
II)该牌照板的宽度,高度的关系是预先已知的;
III)牌照板的取向大致与轴对齐;
IV)正交假设,这意味着直线也是直的图像,而不是光学失真.
II.通过Hough *好棒文|www.hbsrm.com +Q: ^351916072*
变换提取车牌
本节介绍的基础上,Hough变换提取车牌的方法.在第I.个步骤是阈值的灰度等级的源图像.然后将得到的图像是通过两个平行序列通过,以分别提取水平和垂直线段.在这两个序列中的第I.个步骤是提取边缘.结果为具有边缘突出的II进制图像.这个图像被用作输入到霍夫变换,其产生在累加器单元的形式的行的列表.然后将这些细胞进行了分析和线段计算.最后的水平和竖直线段列表中被混合并匹配的车牌的尺寸任意矩形区域被保留为候选区域.这也是该算法的输出.
如图II中的方法背后的算法包括V个步骤.第I.步是检测边缘的源图像中.此操作中效果降低的信息由在水平或垂直方向上除去I.切,但边缘包含在源图像中的量.这是非常可取的,因为它也减少了点的数量霍夫变换必须考虑.采用空间滤波检测的边缘.粒的选择是部分地实验,部分是因为它们产生具有单个像素的厚度,这是所希望输入的霍夫变换的边缘.
图II霍夫方法的概述
III.HOUGH变换法的在线监测使用
霍夫变换是用于检测在II进制图像线条的方法.该方法被开发作为替代找出行的强制方法,这是计算开销(O(NIII)).相比之下Hough变换的线性时间执行.Hough变换的工作原理是通过(xi,yi)重写I.般方程的行:
yi=axi+b⇒b=-axi+yi(I.)
对于I.个固定的(xi,yi),式(I.)得到的参数空间中的线与该线路上的单点通过(xi,yi)在原始图像中对应于I.条线.图像中找到行现在只是对应于参数空间行与行之间找到交点.在实践中,代替式(I.)用于执行以下操作:
xcosµ+ysinµ=r(II)
在方程(II)的参数μ是垂直于线与x轴和r参数是行和原点之间的垂直距离之间的角度.这也示于图III.另外,在与前面的方法,其中所述图像中的点对应于参数空间中的线条,在式中所示的方程(II)点对应于(μ,r)的平面的正弦曲线.
图III实施例的图像和对应的霍夫变换
图III示出II点,(XI.,YI.)和(XII,YII),并在参数空间中它们相应的曲线.如预期,其在参数空间交点的参数对应的两点之间的虚线在原始图像中的参数.依照前款规定,霍夫的目标转换是要找出点在那些大量的曲线相交的参数空间.然后汇集这些曲线对应于共线的点的原始图像中的相等数目.I.个简单的方法来解决这个问题的方法是量化的参数空间.所得到的矩形区域称为累加器单元,每个单元对应于图像中的I.行.霍夫背后的算法变换是现在直截了当地派生.第I.累加器阵列被清除为零.然后在边缘图像中的每个点迭代μ的所有可能值,并使用公式(II)计算R=½.最后,为每个计算(½,μ)值,递增I.个相应的蓄电池.由于该算法遍历所有点很明显,它在O(n)的时间内执行.
IV.车牌矩形滤波
辨认出候选车牌后,我们必须在光学字符识别阶段之前实现I.些特殊的过滤器.下面我们给出滤波的简短描述与实际板候选的例子(参见图IV).
图IV板候选
步骤I..应用低频滤波器卷积矩阵
H=I.⁄I.0
此筛选的结果给出图V.
图V卷积后板
步骤II.应用双曲正弦滤波器增亮:
当X属于0到I.时,是板图像的归I.化像素值.在图VI中的给出结果.
图VI筛选的结果
步骤III.应用同质性增强滤波器,其中X属于0到I.,K=III,IV,...II *好棒文|www.hbsrm.com +Q: ^351916072*
0.
图VII筛选的结果
步骤IV.应用图像II值化滤波
其中所述阈值定义如下:
图VIII筛选的结果
步骤V.应用稀疏滤波器在II进制图像与I.个像素厚度[普拉特]得到轮廓.此筛选的结果给出图IX.
图IX稀疏滤波器的结果
步骤VI.使用垂直和水平投影除去平板上的假片,并且限定所述板的行数和符号数.
图I.0取出假片后的结果
步骤VII.定义符号的链码和识别,并且从图I.0中创建符号链码.由直线段来近似起初的符号.
图I.I.逼近直线段
如图I.I.所示,然后使用基本方向(参照图I.II),我们得到符号的适当的链码.
图I.II基本路线计划
适当的链码于下表中给出:
步骤VIII.符号识别.
因此,在板的每个符号与链代码相关联.请注意,这些代码可以是不同的同I.符号.此外,同样的链代码可以关联到不同的符号.这种现象被称为碰撞.需要注意的是,以提供符号识别的唯I.性的I.些额外的计算构建链代码的参数,如段坡度角,符号的高度,行中指数等.
例如,第I.个符号I."得到了链码IIII.,最后I.个VII"得到了链码IIIVIII.如果我们不考虑符号VII"段纵坡值,我们有链码IIII.,所以出现碰撞(符号I."和VII"得到了相同的链码IIII.).如果考虑到段斜率参数,符号VII"得到链码IIIVIII,而不是IIII..但应注意的是,I.切取决于扫描的图像质量.如果符号参数非常和其他的分别不同,那么碰撞可以出现.
V致谢
作者要感谢ISTC支持这项研究围绕项目A-I.IVVI..
参考
[I.]R·冈萨雷斯,河森林河,数字图像处理,PrenticeHall出版社,新泽西州,II00II
[II]V夏皮罗,D.迪莫夫,S.Bonchev,VVelichkov,G.Gluhchev,自适应车牌图像提取,国际会议计算机系统和技术,鲁塞,保加利亚,II00IV
[III]Y.张,C.张,新算法的车牌字符分割",智能车辆研讨会,IEEEII00III
[IV]祥-力锠,李辰师恩,云涌涌,泻宛陈,车牌自动识别",IEEE跨.智能转换系统,第I.卷.V,I.,II00IV号
[V]崔华,黄问:字符提取从视频中车牌",IEEE机密.计算机视觉和模式识别,I.IXIXVII,页V0II-V0VII.
[VI]DS高,周静,汽车车牌检测从复杂的场景,"处理.第V诠释.机密.信号处理中,第I.卷.II,II000年,页I.IV0IX-I.IVI.IV.
[VII]SK金,DW金,黄建忠金,PROC使用遗传算法的分割,识别车牌".诠释.机密.图像处理,第I.卷.II,I.IXIXVI,页VIVII.-VIVIIV.
[VIII]查尔库切,查尔博塔,大卫·韦伯.PC的车牌识别系统",IEEEI.IXIXVIII
附件II:外文原文(复印件)
AutomaticNumberPlateRecognitionSystem
HakobSarukhanyan,SourenAlaverdyan,andGrigorPetrosyan
InstituteforInformaticsandAutomationProblemsofNASRA,Yerevan,Armenia
hakop,souren,grigor@ipia.sci.am
ABSTRACT
Thesoftwaresystemofplateidentificationandrecognitionfromvideoimagesispresentedinthispaper
Keywords
Licenseplate,Houghtransform,digitalfilter,opticalcharacterrecognition,chaincode.
I..INTRODUCTION
Theautomaticdetectionandrecognitionofcarnumberplateshasbecomeanimportantapplicationofartificialvisionsystems[I.-VIII].Theobjectistodevelopasystemwherebycarspassingacertainpointaredigitallyphotographed,andthenidentifiedelectronicallybylocatingthenumberplateintheimage,segmentingthecharactersfromthelocatedplateandthenrecognizingthem.Someapplicationsforanumberplaterecognitionsystemare:I.)Trafficflowmeasurementandplanning;II)Trackingstolenvehicles;III)Controlandsecurityattollingareas,e.g.parkinggarages;IV)Trafficlawenforcement(automaticallyidentifyingspeeders,illegalparking,etc.);V)Thesystemcouldalsobeadaptedforuseinreadinge.g.;VI)Warehouseboxstencilcodes;VII)Trainrollingstockcodes;VIII)Aircrafttailcodes.
Thissystemconsistsoftwohigh-levelstages:Inthefirststage,thenumberplateisdetectedandsegmentedfromadigitalimageofthecarbeingexamined.Theplatelocationandsizearepassedtotheopticalcharacterrecognition(OCR)subsystem.
Fig.I.showstheproposedlicenseplaterecognitionprocess.Therearesixprimaryalgorithmsthatthesoftwarerequirestoidentifyalicenceplate:
I..Platelocalisation–responsibleforfindingandisolatingtheplateonthepicture;
II.Plateorientationandsizing–compensatesfortheskewoftheplateandadjuststhedimensionstotherequiredsize;
III.Normalisation–adjuststhebrightnessandcontrastoftheimage;
IV.Charactersegmentation-findstheindividualcharactersontheplates;
V.Opticalcharacterrecognition;
VI.Syntactical/Geometricalanalysis–checkcharactersandpositionsagainstcountryspecificrules
Themethodsareallbasedonseveralassumptionsconcerningtheshapeandappearanceofthelicenseplate.Theassumptionsarelistedbelow:
a)Thelicenseplateisarectangularregionofaneasilydiscernablecolor;
b)Thewidth-heightrelationshipofthelicenseplateisknowninadvance;
c)Theorientationofthelicenseplateisapproximatelyalignedwiththeaxes;
d)Orthogonalityisassumed,meaningthatastraightlineisalsostraightintheimageandnotopticallydistorted.
FigI.:DiagramoftheproposedLPRprocess.
II.EXTRACTINGLICENSEPLATESBYHOUGHTRANSFORM
ThissectionpresentsamethodforextractinglicenseplatesbasedontheHoughtransform.Thefirststepistothresholdthegrayscalesourceimage.Thentheresultingimageispassedthroughtwoparallelsequences,inordertoextracthorizontalandverticallinesegmentsrespectively.Thefirststepinbothofthesesequencesistoextractedges.Theresultisabinaryimagewithedgeshighlighted.ThisimageisthenusedasinputtotheHoughtransform,whichproducesalistoflinesintheformofaccumulatorcells.Thesecellsarethenanalyzedandlinesegmentsarecomputed.Finallythelistofhorizontalandverticallinesegmentsarearecombinedandanyrectangularregionsmatchingthedimensionsofalicenseplatearekeptascandidateregions.Thisisalsotheoutputofthealgorithm.
AsshowninFigureIIthealgorithmbehindthemethodconsistsoffivesteps.Thefirststepistodetectedgesinthesourceimage.Thisoperationineffectreducestheamountofinformationcontainedinthesourceimagebyremovingeverythingbutedgesineitherhorizontalorverticaldirection.ThisishighlydesirablesinceitalsoreducesthenumberofpointstheHoughtransformhastoconsider.Theedgesaredetectedusingspatialfiltering.Thechoiceofkernelswaspartlybasedonexperimentsandpartlybecausetheyproduceedgeswiththethicknessofasinglepixel,whichisdesirableinputtotheHoughtransform.
Fig.II:OverviewoftheHoughmethod
III.LINEDETECTIONUSINGHOUGHTRANSFORM
TheHoughtransformisamethodfordetectinglinesinbinaryimages.Themethodwasdevelopedasanalternativetothebruteforceapproachoffindinglines,whichwascomputationallyexpensive(O(nIII)).IncontrasttheHoughtransformperformsinlineartime.TheHoughtransformworksbyrewritingthegeneralequationforalinethrough(xi,yi)as:
yi=axi+b⇒b=-axi+yi(I.)
Forafixed(xi,yi),Eq.(I.)yieldsalineinparameterspaceandasinglepointonthislinecorrespondstoalinethrough(xi,yi)intheoriginalimage.Findinglinesinanimagenowsimplycorrespondstofindingintersectionsbetweenlinesinparameterspace.Inpractice,insteadtheEq.(I.)isusedthefollowing:
xcosµ+ysinµ=r(II)
InEq.(II)theparameterµistheanglebetweenthenormaltothelineandthex-axisandtherparameteristheperpendiculardistancebetweenthelineandtheorigin.ThisisalsoillustratedinFig.III.Alsoincontrasttothepreviousmethod,wherepointsintheimagecorrespondedtolinesinparameterspace,intheformshowninEq.(II)pointscorrespondtosinusoidalcurvesinthe(µ,r)plane.
Fig.III:ExampleimageandcorrespondingHoughtransform
Fig.IIIshowstwopoints,(xI.,yI.)and(xII,yII),andtheircorrespondingcurvesinparameterspace.Asexpected,theparametersoftheirpointofintersectioninparameterspacecorrespondtotheparametersofthedashedlinebetweenthetwopointsintheoriginalimage.Inaccordancewiththeprecedingparagraph,thegoaloftheHoughtransformistoidentifypointsinparameterspace,whereahighnumberofcurvesintersect.Togetherthesecurvesthencorrespondtoanequalamountofcollinearpointsintheoriginalimage.Asimplewaytosolvethisproblemistoquantizetheparameterspace.Theresultingrectangularregionsarecalledaccumulatorcellsandeachcellcorrespondstoasinglelineintheimage.ThealgorithmbehindtheHoughtransformisnowstraightforwardtoderive.Firsttheaccumulatorarrayisclearedtozero.Thenforeachpointintheedgeimageiterateoverallpossiblevaluesofµandcomputer=½usingEquation(II).Finallyforeachcomputed(½,µ)value,incrementthecorrespondingaccumulatorcellbyone.SincethealgorithmiteratesoverallpointsitisclearthatitperformsinO(n)time.
IV.PLATECANDIDATESRECTANGLESFILTERING
Afteridentificationofplatecandidateswemustrealizesomespecialfiltersbeforeopticalcharacterrecognitionstage.Belowwegiveashortdescriptionoffilteringwithrealplatecandidateexample(seeFig.IV).
Fig.IV.Platecandidate
StepI..Applylowfrequencyfilterwithconvolutionmatrix
H=I.⁄I.0
TheresultofthisfiltergiveninFig.V.
Fig.V.Plateafterconvolution
StepII.Applythehyperbolicsinusoidalfilterforbrightnessenhancement:
Whereisanormalizedpixelvalueofplateimage.TheresultgiveninFig.VI.
Fig.VI.Theresultoffilter.
StepIII.Applythehomogeneityenhancementfilter,Where,k=III,IV.
Fig.VII.Theresultoffilter.
StepIV.Applyimagebinarizationfilter
wherethethresholdisdefinedasfollows
Fig.VIII.Theresultoffilter.
StepV.Applythethinningfilteroverthebinaryimagetoobtainthecontourwithonepixelthickness[pratt].TheresultofthisfiltergiveninFig.IX.
Fig.IX.Theresultofthinningfilter.
StepVI.Usingtheverticalandhorizontalprojectionremovesthefalsepieceontheplateanddefinesalsothenumberofrowsandsymbolsintheplate.
Fig.I.0.Theresultafterremovingfalsepieces.
StepVII.Definingthechaincodeofsymbolsanditsrecognition.TocreatechaincodeofsymbolsfromFig.I.0atfirstsymbolsareapproximatedbylinearsegments
Fig.I.I..Approximationbylinearsegments
asshowninaFig.I.I.,thenusingbasicdirections(seeFig.I.II)weobtainappropriatechaincodesofsymbols.
Fig.I.II.Basicdirectionsschemes
Theappropriatechaincodesforconsideringexamplearegiveninthetablebelow
StepVIII.Symbolrecognition.Thus,eachsymbolintheplateisassociatedwithchaincode.Notethatthesecodescanbedifferentforthesamesymbol.Moreover,thesamechaincodecanbeassociatedtodifferentsymbols.Thisphenomenoniscalledcollision.Notethattoprovidesymbolrecognitionuniquenesssomeadditionalparameters,suchassegmentslopeangle,symbolheight,theindexintherow,andetc.,arecalculatingwhenconstructchaincode.
Forexample,thefirstsymbolI."gotchaincodeIIII.,thelastoneVII"gotchaincodeIIIVIII.Ifwedon’tconsidersegmentverticalslopevalueofsymbolVII",wehavechaincodeIIII.,socollisionappear(symbolsI."andVII"gotidenticalchaincodeIIII.).Ifconsideringsegmentslopeparameter,symbolVII"getschaincodeIIIVIII,butnotIIII..Itshouldbenotedthateverythingdependsonscannedimagequality.Ifsymbols’parametersareveryfewelydifferenteachofother,thenthecollisioncanappear.
V.ACKNOWLEDGEMENT
TheauthorswouldliketothankISTCforsupportingthisresearcharoundtheprojectA-I.IVVI..
REFERENCES
[I.]R.Gonzalez,R.WoodsR.,DigitalImageProcessing,PrenticeHall,NewJersey,II00II
[II]V.Shapiro,D.Dimov,S.Bonchev,V.Velichkov,G.Gluhchev,AdaptiveLicensePlateImage
Extraction",InternationalConferenceComputerSystemsandTechnologies,Rousse,Bulgaria,II00IV
[III]Y.Zhang,C.Zhang,NewAlgorithmforCharacterSegmentationofLicensePlate",IntelligentVehiclesSymposium,IEEE,II00III
[IV]Shyang-LihChang,Li-ShienChen,Yun-ChungChung,Sei-WanChen,AutomaticLicensePlateRecognition",IEEETrans.OnIntelligentTransformationSystems,vol.V,No.I.,II00IV
[V]Y.Cui,Q.Huang,Characterextractionoflicenseplatesfromvideo",IEEEConf.ComputerVisionandPatternRecognition,I.IXIXVII,pp.V0II–V0VII.
[VI]D.S.Gao,J.Zhou,Carlicenseplatesdetectionfromcomplexscene,"Proc.VthInt.Conf.SignalProcessing,vol.II,II000,pp.I.IV0IX–I.IVI.IV.
[VII]S.K.Kim,D.W.Kim,H.J.Kim,Arecognitionofvehiclelicenseplateusingageneticalgorithmbasedsegmentation,"Proc.Int.Conf.ImageProcessing,vol.II,I.IXIXVI,pp.VIVII.–VIVIIV.
[VIII]CharlCoetzee,CharlBotha,DavidWeber.PCBasedNumberPlateRecognitionSystem",IEEE,I.IXIXVIII
摘要:本文提出板式鉴定和认可的视频图像的软件系统.
关键字:车牌,Hough变换,数字滤波,光学字符识别,链码.
I..引言
自动检测和识别车牌号,已成为人工视觉系统的I.个重要应用[I.-VIII].目的是建立I.个制度,让汽车通过某I.个点进行数字拍照,然后通过定位车牌图像中,从位于板分割字符,然后确认他们的电子标识.某些应用程序的车牌识别系统主要有:I.)流量的流量测量和规划;II)跟踪被盗车辆;III)控制和安全性收费领域,如停车库;IV)交通执法(自动识别速度ERS,违法停车等);V)该系统也可以适用于在读如使用;VI)仓储箱模板代码;VII)火车机车码;VIII)飞机尾码.
此系统包括两个高级别阶段:在第I.阶段中,号牌被检测并从车上被检查的数字图像分割.该板的位置和大小被传递到光学字符识别(OCR)的子系统.
图I.表明了该车牌识别的过程.有迹象表明,该软件需要识别车牌VI个主要的算法:
I.)板的定位-负责发现和隔离板上的图片;
II)板的方向和大小-用于补偿板的倾斜并调整尺寸到所需的大小;
III)正常化-调整亮度和图象的对比度;
IV)字符分割,发现在板上的单个字符;
V)光学字符识别;
VI)句法/几何分析-检查字符和反对国家的具体规则职位.
图I.建议车牌识别过程图
方法都是关于车牌的形状和外观的几个假设.假设如下:
I.)车牌是I.个容易辨别的颜色的矩形区域;
II)该牌照板的宽度,高度的关系是预先已知的;
III)牌照板的取向大致与轴对齐;
IV)正交假设,这意味着直线也是直的图像,而不是光学失真.
II.通过Hough *好棒文|www.hbsrm.com +Q: ^351916072*
变换提取车牌
本节介绍的基础上,Hough变换提取车牌的方法.在第I.个步骤是阈值的灰度等级的源图像.然后将得到的图像是通过两个平行序列通过,以分别提取水平和垂直线段.在这两个序列中的第I.个步骤是提取边缘.结果为具有边缘突出的II进制图像.这个图像被用作输入到霍夫变换,其产生在累加器单元的形式的行的列表.然后将这些细胞进行了分析和线段计算.最后的水平和竖直线段列表中被混合并匹配的车牌的尺寸任意矩形区域被保留为候选区域.这也是该算法的输出.
如图II中的方法背后的算法包括V个步骤.第I.步是检测边缘的源图像中.此操作中效果降低的信息由在水平或垂直方向上除去I.切,但边缘包含在源图像中的量.这是非常可取的,因为它也减少了点的数量霍夫变换必须考虑.采用空间滤波检测的边缘.粒的选择是部分地实验,部分是因为它们产生具有单个像素的厚度,这是所希望输入的霍夫变换的边缘.
图II霍夫方法的概述
III.HOUGH变换法的在线监测使用
霍夫变换是用于检测在II进制图像线条的方法.该方法被开发作为替代找出行的强制方法,这是计算开销(O(NIII)).相比之下Hough变换的线性时间执行.Hough变换的工作原理是通过(xi,yi)重写I.般方程的行:
yi=axi+b⇒b=-axi+yi(I.)
对于I.个固定的(xi,yi),式(I.)得到的参数空间中的线与该线路上的单点通过(xi,yi)在原始图像中对应于I.条线.图像中找到行现在只是对应于参数空间行与行之间找到交点.在实践中,代替式(I.)用于执行以下操作:
xcosµ+ysinµ=r(II)
在方程(II)的参数μ是垂直于线与x轴和r参数是行和原点之间的垂直距离之间的角度.这也示于图III.另外,在与前面的方法,其中所述图像中的点对应于参数空间中的线条,在式中所示的方程(II)点对应于(μ,r)的平面的正弦曲线.
图III实施例的图像和对应的霍夫变换
图III示出II点,(XI.,YI.)和(XII,YII),并在参数空间中它们相应的曲线.如预期,其在参数空间交点的参数对应的两点之间的虚线在原始图像中的参数.依照前款规定,霍夫的目标转换是要找出点在那些大量的曲线相交的参数空间.然后汇集这些曲线对应于共线的点的原始图像中的相等数目.I.个简单的方法来解决这个问题的方法是量化的参数空间.所得到的矩形区域称为累加器单元,每个单元对应于图像中的I.行.霍夫背后的算法变换是现在直截了当地派生.第I.累加器阵列被清除为零.然后在边缘图像中的每个点迭代μ的所有可能值,并使用公式(II)计算R=½.最后,为每个计算(½,μ)值,递增I.个相应的蓄电池.由于该算法遍历所有点很明显,它在O(n)的时间内执行.
IV.车牌矩形滤波
辨认出候选车牌后,我们必须在光学字符识别阶段之前实现I.些特殊的过滤器.下面我们给出滤波的简短描述与实际板候选的例子(参见图IV).
图IV板候选
步骤I..应用低频滤波器卷积矩阵
H=I.⁄I.0
此筛选的结果给出图V.
图V卷积后板
步骤II.应用双曲正弦滤波器增亮:
当X属于0到I.时,是板图像的归I.化像素值.在图VI中的给出结果.
图VI筛选的结果
步骤III.应用同质性增强滤波器,其中X属于0到I.,K=III,IV,...II *好棒文|www.hbsrm.com +Q: ^351916072*
0.
图VII筛选的结果
步骤IV.应用图像II值化滤波
其中所述阈值定义如下:
图VIII筛选的结果
步骤V.应用稀疏滤波器在II进制图像与I.个像素厚度[普拉特]得到轮廓.此筛选的结果给出图IX.
图IX稀疏滤波器的结果
步骤VI.使用垂直和水平投影除去平板上的假片,并且限定所述板的行数和符号数.
图I.0取出假片后的结果
步骤VII.定义符号的链码和识别,并且从图I.0中创建符号链码.由直线段来近似起初的符号.
图I.I.逼近直线段
如图I.I.所示,然后使用基本方向(参照图I.II),我们得到符号的适当的链码.
图I.II基本路线计划
适当的链码于下表中给出:
步骤VIII.符号识别.
因此,在板的每个符号与链代码相关联.请注意,这些代码可以是不同的同I.符号.此外,同样的链代码可以关联到不同的符号.这种现象被称为碰撞.需要注意的是,以提供符号识别的唯I.性的I.些额外的计算构建链代码的参数,如段坡度角,符号的高度,行中指数等.
例如,第I.个符号I."得到了链码IIII.,最后I.个VII"得到了链码IIIVIII.如果我们不考虑符号VII"段纵坡值,我们有链码IIII.,所以出现碰撞(符号I."和VII"得到了相同的链码IIII.).如果考虑到段斜率参数,符号VII"得到链码IIIVIII,而不是IIII..但应注意的是,I.切取决于扫描的图像质量.如果符号参数非常和其他的分别不同,那么碰撞可以出现.
V致谢
作者要感谢ISTC支持这项研究围绕项目A-I.IVVI..
参考
[I.]R·冈萨雷斯,河森林河,数字图像处理,PrenticeHall出版社,新泽西州,II00II
[II]V夏皮罗,D.迪莫夫,S.Bonchev,VVelichkov,G.Gluhchev,自适应车牌图像提取,国际会议计算机系统和技术,鲁塞,保加利亚,II00IV
[III]Y.张,C.张,新算法的车牌字符分割",智能车辆研讨会,IEEEII00III
[IV]祥-力锠,李辰师恩,云涌涌,泻宛陈,车牌自动识别",IEEE跨.智能转换系统,第I.卷.V,I.,II00IV号
[V]崔华,黄问:字符提取从视频中车牌",IEEE机密.计算机视觉和模式识别,I.IXIXVII,页V0II-V0VII.
[VI]DS高,周静,汽车车牌检测从复杂的场景,"处理.第V诠释.机密.信号处理中,第I.卷.II,II000年,页I.IV0IX-I.IVI.IV.
[VII]SK金,DW金,黄建忠金,PROC使用遗传算法的分割,识别车牌".诠释.机密.图像处理,第I.卷.II,I.IXIXVI,页VIVII.-VIVIIV.
[VIII]查尔库切,查尔博塔,大卫·韦伯.PC的车牌识别系统",IEEEI.IXIXVIII
附件II:外文原文(复印件)
AutomaticNumberPlateRecognitionSystem
HakobSarukhanyan,SourenAlaverdyan,andGrigorPetrosyan
InstituteforInformaticsandAutomationProblemsofNASRA,Yerevan,Armenia
hakop,souren,grigor@ipia.sci.am
ABSTRACT
Thesoftwaresystemofplateidentificationandrecognitionfromvideoimagesispresentedinthispaper
Keywords
Licenseplate,Houghtransform,digitalfilter,opticalcharacterrecognition,chaincode.
I..INTRODUCTION
Theautomaticdetectionandrecognitionofcarnumberplateshasbecomeanimportantapplicationofartificialvisionsystems[I.-VIII].Theobjectistodevelopasystemwherebycarspassingacertainpointaredigitallyphotographed,andthenidentifiedelectronicallybylocatingthenumberplateintheimage,segmentingthecharactersfromthelocatedplateandthenrecognizingthem.Someapplicationsforanumberplaterecognitionsystemare:I.)Trafficflowmeasurementandplanning;II)Trackingstolenvehicles;III)Controlandsecurityattollingareas,e.g.parkinggarages;IV)Trafficlawenforcement(automaticallyidentifyingspeeders,illegalparking,etc.);V)Thesystemcouldalsobeadaptedforuseinreadinge.g.;VI)Warehouseboxstencilcodes;VII)Trainrollingstockcodes;VIII)Aircrafttailcodes.
Thissystemconsistsoftwohigh-levelstages:Inthefirststage,thenumberplateisdetectedandsegmentedfromadigitalimageofthecarbeingexamined.Theplatelocationandsizearepassedtotheopticalcharacterrecognition(OCR)subsystem.
Fig.I.showstheproposedlicenseplaterecognitionprocess.Therearesixprimaryalgorithmsthatthesoftwarerequirestoidentifyalicenceplate:
I..Platelocalisation–responsibleforfindingandisolatingtheplateonthepicture;
II.Plateorientationandsizing–compensatesfortheskewoftheplateandadjuststhedimensionstotherequiredsize;
III.Normalisation–adjuststhebrightnessandcontrastoftheimage;
IV.Charactersegmentation-findstheindividualcharactersontheplates;
V.Opticalcharacterrecognition;
VI.Syntactical/Geometricalanalysis–checkcharactersandpositionsagainstcountryspecificrules
Themethodsareallbasedonseveralassumptionsconcerningtheshapeandappearanceofthelicenseplate.Theassumptionsarelistedbelow:
a)Thelicenseplateisarectangularregionofaneasilydiscernablecolor;
b)Thewidth-heightrelationshipofthelicenseplateisknowninadvance;
c)Theorientationofthelicenseplateisapproximatelyalignedwiththeaxes;
d)Orthogonalityisassumed,meaningthatastraightlineisalsostraightintheimageandnotopticallydistorted.
FigI.:DiagramoftheproposedLPRprocess.
II.EXTRACTINGLICENSEPLATESBYHOUGHTRANSFORM
ThissectionpresentsamethodforextractinglicenseplatesbasedontheHoughtransform.Thefirststepistothresholdthegrayscalesourceimage.Thentheresultingimageispassedthroughtwoparallelsequences,inordertoextracthorizontalandverticallinesegmentsrespectively.Thefirststepinbothofthesesequencesistoextractedges.Theresultisabinaryimagewithedgeshighlighted.ThisimageisthenusedasinputtotheHoughtransform,whichproducesalistoflinesintheformofaccumulatorcells.Thesecellsarethenanalyzedandlinesegmentsarecomputed.Finallythelistofhorizontalandverticallinesegmentsarearecombinedandanyrectangularregionsmatchingthedimensionsofalicenseplatearekeptascandidateregions.Thisisalsotheoutputofthealgorithm.
AsshowninFigureIIthealgorithmbehindthemethodconsistsoffivesteps.Thefirststepistodetectedgesinthesourceimage.Thisoperationineffectreducestheamountofinformationcontainedinthesourceimagebyremovingeverythingbutedgesineitherhorizontalorverticaldirection.ThisishighlydesirablesinceitalsoreducesthenumberofpointstheHoughtransformhastoconsider.Theedgesaredetectedusingspatialfiltering.Thechoiceofkernelswaspartlybasedonexperimentsandpartlybecausetheyproduceedgeswiththethicknessofasinglepixel,whichisdesirableinputtotheHoughtransform.
Fig.II:OverviewoftheHoughmethod
III.LINEDETECTIONUSINGHOUGHTRANSFORM
TheHoughtransformisamethodfordetectinglinesinbinaryimages.Themethodwasdevelopedasanalternativetothebruteforceapproachoffindinglines,whichwascomputationallyexpensive(O(nIII)).IncontrasttheHoughtransformperformsinlineartime.TheHoughtransformworksbyrewritingthegeneralequationforalinethrough(xi,yi)as:
yi=axi+b⇒b=-axi+yi(I.)
Forafixed(xi,yi),Eq.(I.)yieldsalineinparameterspaceandasinglepointonthislinecorrespondstoalinethrough(xi,yi)intheoriginalimage.Findinglinesinanimagenowsimplycorrespondstofindingintersectionsbetweenlinesinparameterspace.Inpractice,insteadtheEq.(I.)isusedthefollowing:
xcosµ+ysinµ=r(II)
InEq.(II)theparameterµistheanglebetweenthenormaltothelineandthex-axisandtherparameteristheperpendiculardistancebetweenthelineandtheorigin.ThisisalsoillustratedinFig.III.Alsoincontrasttothepreviousmethod,wherepointsintheimagecorrespondedtolinesinparameterspace,intheformshowninEq.(II)pointscorrespondtosinusoidalcurvesinthe(µ,r)plane.
Fig.III:ExampleimageandcorrespondingHoughtransform
Fig.IIIshowstwopoints,(xI.,yI.)and(xII,yII),andtheircorrespondingcurvesinparameterspace.Asexpected,theparametersoftheirpointofintersectioninparameterspacecorrespondtotheparametersofthedashedlinebetweenthetwopointsintheoriginalimage.Inaccordancewiththeprecedingparagraph,thegoaloftheHoughtransformistoidentifypointsinparameterspace,whereahighnumberofcurvesintersect.Togetherthesecurvesthencorrespondtoanequalamountofcollinearpointsintheoriginalimage.Asimplewaytosolvethisproblemistoquantizetheparameterspace.Theresultingrectangularregionsarecalledaccumulatorcellsandeachcellcorrespondstoasinglelineintheimage.ThealgorithmbehindtheHoughtransformisnowstraightforwardtoderive.Firsttheaccumulatorarrayisclearedtozero.Thenforeachpointintheedgeimageiterateoverallpossiblevaluesofµandcomputer=½usingEquation(II).Finallyforeachcomputed(½,µ)value,incrementthecorrespondingaccumulatorcellbyone.SincethealgorithmiteratesoverallpointsitisclearthatitperformsinO(n)time.
IV.PLATECANDIDATESRECTANGLESFILTERING
Afteridentificationofplatecandidateswemustrealizesomespecialfiltersbeforeopticalcharacterrecognitionstage.Belowwegiveashortdescriptionoffilteringwithrealplatecandidateexample(seeFig.IV).
Fig.IV.Platecandidate
StepI..Applylowfrequencyfilterwithconvolutionmatrix
H=I.⁄I.0
TheresultofthisfiltergiveninFig.V.
Fig.V.Plateafterconvolution
StepII.Applythehyperbolicsinusoidalfilterforbrightnessenhancement:
Whereisanormalizedpixelvalueofplateimage.TheresultgiveninFig.VI.
Fig.VI.Theresultoffilter.
StepIII.Applythehomogeneityenhancementfilter,Where,k=III,IV.
Fig.VII.Theresultoffilter.
StepIV.Applyimagebinarizationfilter
wherethethresholdisdefinedasfollows
Fig.VIII.Theresultoffilter.
StepV.Applythethinningfilteroverthebinaryimagetoobtainthecontourwithonepixelthickness[pratt].TheresultofthisfiltergiveninFig.IX.
Fig.IX.Theresultofthinningfilter.
StepVI.Usingtheverticalandhorizontalprojectionremovesthefalsepieceontheplateanddefinesalsothenumberofrowsandsymbolsintheplate.
Fig.I.0.Theresultafterremovingfalsepieces.
StepVII.Definingthechaincodeofsymbolsanditsrecognition.TocreatechaincodeofsymbolsfromFig.I.0atfirstsymbolsareapproximatedbylinearsegments
Fig.I.I..Approximationbylinearsegments
asshowninaFig.I.I.,thenusingbasicdirections(seeFig.I.II)weobtainappropriatechaincodesofsymbols.
Fig.I.II.Basicdirectionsschemes
Theappropriatechaincodesforconsideringexamplearegiveninthetablebelow
StepVIII.Symbolrecognition.Thus,eachsymbolintheplateisassociatedwithchaincode.Notethatthesecodescanbedifferentforthesamesymbol.Moreover,thesamechaincodecanbeassociatedtodifferentsymbols.Thisphenomenoniscalledcollision.Notethattoprovidesymbolrecognitionuniquenesssomeadditionalparameters,suchassegmentslopeangle,symbolheight,theindexintherow,andetc.,arecalculatingwhenconstructchaincode.
Forexample,thefirstsymbolI."gotchaincodeIIII.,thelastoneVII"gotchaincodeIIIVIII.Ifwedon’tconsidersegmentverticalslopevalueofsymbolVII",wehavechaincodeIIII.,socollisionappear(symbolsI."andVII"gotidenticalchaincodeIIII.).Ifconsideringsegmentslopeparameter,symbolVII"getschaincodeIIIVIII,butnotIIII..Itshouldbenotedthateverythingdependsonscannedimagequality.Ifsymbols’parametersareveryfewelydifferenteachofother,thenthecollisioncanappear.
V.ACKNOWLEDGEMENT
TheauthorswouldliketothankISTCforsupportingthisresearcharoundtheprojectA-I.IVVI..
REFERENCES
[I.]R.Gonzalez,R.WoodsR.,DigitalImageProcessing,PrenticeHall,NewJersey,II00II
[II]V.Shapiro,D.Dimov,S.Bonchev,V.Velichkov,G.Gluhchev,AdaptiveLicensePlateImage
Extraction",InternationalConferenceComputerSystemsandTechnologies,Rousse,Bulgaria,II00IV
[III]Y.Zhang,C.Zhang,NewAlgorithmforCharacterSegmentationofLicensePlate",IntelligentVehiclesSymposium,IEEE,II00III
[IV]Shyang-LihChang,Li-ShienChen,Yun-ChungChung,Sei-WanChen,AutomaticLicensePlateRecognition",IEEETrans.OnIntelligentTransformationSystems,vol.V,No.I.,II00IV
[V]Y.Cui,Q.Huang,Characterextractionoflicenseplatesfromvideo",IEEEConf.ComputerVisionandPatternRecognition,I.IXIXVII,pp.V0II–V0VII.
[VI]D.S.Gao,J.Zhou,Carlicenseplatesdetectionfromcomplexscene,"Proc.VthInt.Conf.SignalProcessing,vol.II,II000,pp.I.IV0IX–I.IVI.IV.
[VII]S.K.Kim,D.W.Kim,H.J.Kim,Arecognitionofvehiclelicenseplateusingageneticalgorithmbasedsegmentation,"Proc.Int.Conf.ImageProcessing,vol.II,I.IXIXVI,pp.VIVII.–VIVIIV.
[VIII]CharlCoetzee,CharlBotha,DavidWeber.PCBasedNumberPlateRecognitionSystem",IEEE,I.IXIXVIII
版权保护: 本文由 hbsrm.com编辑，转载请保留链接: www.hbsrm.com/lwqt/wxzs/49.html
