




已阅读5页,还剩41页未读, 继续免费阅读
版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
外文文献资料TheInternationalJournalofSoftComputingandSoftwareEngineeringJSCSE,Vol.3,No.3,SpecialIssue:TheProceedingofInternationalConferenceonSoftComputingandSoftwareEngineering2013SCSE13,SanFranciscoStateUniversity,CA,U.S.A.,March2013Doi:10.7321/jscse.v3.n3.15e-ISSN:2251-754583LicensePlateRecognition(LPR):AReviewwithExperimentsforMalaysiaCaseStudyNuzulhaKhilwaniIbrahim,EmalianaKasmuri,NoraziraAJalil,MohdAdiliNorasikin,SazilahSalamFacultyofInformationandCommunicationTechnology,UniversitiTeknikalMalaysiaMelaka(UTeM),HangTuahJaya,76100Melaka,Malaysianuzulha|emaliana|adili|.myMohamadRiduwanMdNawawiFacultyofElectricalEngineering,UniversitiTeknikalMalaysiaMelaka(UTeM),HangTuahJaya,76100Melaka,M.myAbstractMostvehiclelicenseplaterecognitionuseneuralnetworktechniquestoenhanceitscomputingcapability.Theimageofthevehiclelicenseplateiscapturedandprocessedtoproduceatextualoutputforfurtherprocessing.Thispaperreviewsimageprocessingandneuralnetworktechniquesappliedatdifferentstageswhicharepreprocessing,filtering,featureextraction,segmentationandrecognitioninsuchwaytoremovethenoiseoftheimage,toenhancetheimagequalityandtoexpeditethecomputingprocessbyconvertingthecharactersintheimageintorespectivetext.AnexemplarexperimenthasbeendoneinMATLABtoshowthebasicprocessoftheimageprocessingespeciallyforlicenseplateinMalaysiacasestudy.Analgorithmisadaptedintothesolutionforparkingmanagementsystem.Thesolutionthenisimplementedasproofofconcepttothealgorithm.Keyword-imageprocessing,preprocessing,filtering,featureextraction,segmentation,recognition,experimentI.INTRODUCTIONTheadvancedofcomputerapplicationprocessedmorethantextualdatasolvingeverydayproblems.Inputsfromopticaldeviceareusedindomainapplicationsuchasmedical,security,monitoringandcontrolandengineering.Abilityforcomputertoprocessimageandtranslateitintosomethingmeaningfulhasbecomemorepopular.Therefore,thetechnologyofimageprocessinghasadoptedinmanagingvehicleparkingsystem,vehicleaccesstorestrictedarea,trafficmonitoringsystemandhighwayelectronictollcollection.Forthispurpose,thecomputerneedstocapturethevehiclelicenceplatenumberandprocessitinthecomputer.Acameracapturestheimageofvehiclelicenseplate.Theimagethenfeedintothecomputerforfurtherprocessing.Theoutputoffromtheprocessisthevehiclelicenseplatenumberintextualform.Foraparkingsystem,theoutputisusedforcaridentification,parkingpaymentandauthorizationtoaccessintotheparkingspace.Thispaperreviewstheprocessingofvehiclelicenseplatethatusesimageprocessingandneuralnetworktechnique.Theframeworkforthisresearchisadaptedfrompreviousstudies1-4asshowninFigure1whichincludes5stages:pre-processing,(b)filtering,(c)featureextraction,(d)segmentationand(e)characterrecognition.Thefinaloutputofthesampleexperimentistorecognizethealphanumericcharactersonthelicenseplate.Thestructureofthispaperisorganizedbythestagesoftheprocess.II.PREPROCESSINGDigitalimagepreprocessingisaninitialsteptoimageprocessingimprovingthedataimagequalityformoresuitableforvisualperceptionorcomputationalprocessing.Preprocessingremoveunwanteddataandenhancetheimagebyremovingbackgroundnoise,normalizingtheintensityofindividualimageparticles,imagedeblurandremoveimagereflections.Preprocessingforcarlicenseplatenumberusesthreecommonsubprocesses,whicharegeometricoperation,grayscalingprocessandbinarizationprocess.Manyneuralnetworktechniqueshavebeenappliedtothesepreprocessingtechniquesmainlytoproducebetterimageandtoincreasethespeedofconvergenceofanimage.A.GeometricOperationGeometricoperationisaprocesstolocatethecarlicenseplate.Thepurposeofthisoperationistolocalizethecarplateforfastercharacteridentificationoverasmallregion.AnimprovedBackPropagationnetworkisusedtoovercometheweaknessofconvergencespeedin1.Geneticalgorithmandmomentumtermisintroducedtothecurrentnetworktoincreasethespeedofconvergencerate.ThecurrentBPnetworklearningprocessissaidtobeeasilyproduceerrorifinitialweightsisnotsetproperly1anditisdifficulttodeterminethenumberofhiddenlayerandhiddennodes.TheimprovednetworkusingBPmomentumincreasethespeedandtheaccuracytolocalizethecarlicenseplacelocation.Agrayscaleimageextractstheedgeofthelicenseplateusingsobeloperator1.MalviyaandBhirudin2usesiterativethresholdingoperationtoidentifylicenseplateofavehicle.Objectswithgeometriccharacteristicsarelabelledandselected.Theprocesstakesintoaccountaspectratio,totalpixelperobject,height,widthandthepresenceofcharactersintheregion.Forthis,weproposethefollowingalgorithm,wherethepseudo-codecanbesimplifiedasthefollowing:togetthescaleoftheimageforx-axisandy-axistoassignthenewvalueofhorizontalandverticalaxisbasedonthescaleofthex-axisandy-axistogetthegrayscalethresholdingvalueoftheimageTheinputoftheexperimentisshownasFigure2whiletheexampleoutputcanbeviewedasFigure3.Figure2.InputdataforLicensePlateImageProcessingFigure3.ImageoutputaftergeometricoperationprocessTheoutputfromtheextractionprocesswillbeusedinthenextstagewhichisgrayscalingprocess.B.GrayscalingProcessGrayscalingisaprocesstoproduceagrayscaleimagefromamulticolorimage.Inthisprocess,thethresholdofanimageiscalculated.Ifitislessthanthethreshold,theimagedataisrecalculatedtogetthecorrectgrayscalevalue.Thepurposeofthresholdingistoseparatetheobjectofinterestfromthebackground.Thresholdingisimportanttoprovidesufficientcontrastfortheimagesothatdifferentlevelofintensitybetweenobjectandthebackgroundcanbedifferentiatedforlatercomputationalprocessing.Differentintensitydeterminesthevalueofthethreshold.Grayscalingprocessimprovesthequalityoftheimageforlatercomputationalprocessing.Otherpreprocessingtechniquestoimprovethequalityoftheimageincludingimagedeblurring,imageenhancement,imagefusionandimagereconstruction.Imagefusionisaprocesstoenhancetheimagewithmultiplecombinationsofimages2-3.Thisprocessissuitabletoidentifythecarlicenseregistrationnumberfromamovingcar.Thetechniqueintegratesmultiresolutionimageandproduceacompositeimageusinginversemultiresolutiontransform3.Atemplateofimagefromagrayscaleisshiftedtoverticalandhorizontaldirection.Thecontrastfrequencyiscalculatedforeachpositioninthetemplateandcreatesanewimageusingthresholdingprocedure.Anycolorbelowthethresholdissettoback(zero)andabovethresholdissettowhite(one).Thevaluedeterminesthegraylevelresultingblackandwhiteimage.Atrainedfeedforwardneuralnetwork(FFN)withBlockRecursiveLSalgorithmisusedtoprocesscarlicenseplate4.Theapproachistoimprovetheconvergencerateandstabilizetherobustnessofthesolution.ThelocationofthecarlicenseplateisextractedusingDiscreteFourierTransform(DFT).DFTidentifiesmaximumvalueofhorizontalandverticaledges.Priortothattoneequalizationandcontrastreductionisusedtoimprovetheimage.Thesetechniquesarepreferredbecauseitismorerobustandsuitablecomparedtoedgeenhancement.Forthis,weproposethefollowingalgorithm,wherethepseudo-codecanbesimplifiedasthefollowing:toconvertintograyscaleimageThepseudo-codecanbetranslatedinMATLABsuchasfollowing:TestImg1=rgb2gray(TestImg1);C.BinarizationProcessBinarizationisaprocessofconvertinggrayscaleimageintoblackandwhiteimageor“0”and“1”.Previously,thegrayscaleimageconsistsofdifferentlevelofgrayvalues;from0to255.Toimprovethequalityandextractsomeinformationfromtheimage,theimageneedstobeprocessafewtimesandthusmakethebinaryimagemoreuseful.Graythresholdvalueofanimageisrequiredinthebinarizationprocessasitisimportanttodeterminewhetherthepixelsthathavinggrayvalueswillbeconvertedtoblackorwhite.Forthis,weproposethefollowingalgorithm,wherethepseudo-codecanbesimplifiedasthefollowing:toconvertintoblackandwhiteimageThepseudo-codecanbetranslatedinMATLABsuchasfollowing:ImgBW=im2bw(TestImg1,thresholdTheexampleoutputcanbeviewedasFigure4.Figure4.ImageoutputafterbinarizationprocessTheoutputfromtheextractionprocesswillbeusedinthenextstageoftheprocessinginthisframeworkwhichisfiltering.III.FILTERINGToenhancethequalityofprocessingimage,filteringisrequiredtosolvecontrastenhancement,noisesuppression,blurryissueanddatareduction.ItisreportedthatmostofpreprocessingactivitiesconductedinimagerestorationapplyNeuralNetworkapproach5.Rectanglesfilteringimplementedontherealplatenumberinvolvesconvolutionmatrix,binarizationfilterwithverticalandhorizontalprojectionabletoenhancetheimagequalityandeliminatesunwantedpiecesontheplate.Itisalsorecognizethenumberofrowsandsymbolsintheplatenumber6.In7,asimplefilterisdesignedbyimplementingintensityvarianceandedgedensitytoovercomeilluminationissue,distancechangedandcomplexbackground.Itisproposedthatthisapproachconvenientforreal-timeapplication.Thequalityandselectionofparametersonthecameraextremelycontributesthedesiredpreprocessingimagequality8.TheexampleoutputcanbeviewedasFigure5.Theoutputfromtheextractionprocesswillbeusedinthenextstagewhichisimagesegmentation.Figure5.ImageoutputafterfilteringprocessIV.FEATUREEXTRACTIONFeaturesextractionisthepartofmeasuringthoserelevantfeaturestobeusedinrecognitionprocess.Selectionoftherightfeaturesisimportantinordertoobtainbestresultsinlicenseplaterecognitionstudy.Colourfeaturesareverygoodpotentialforobjectdetection.Howevertheparametersuchascolourofcar,illuminationconditionandthequalityofimagingsystemhasbeenlimiteditspractice9.Accordingto10,colourfeatureshavebeenstudiedby11and12butfromthestudy,thisfeaturenotrobustenoughtovariousenvironments.However,therearemanytypesoffeaturesthatcanaidlicenseplaterecognitionsuchasaspectratio,texture,edgedensity,andsizeofregion10.Inordertoachievebetterdetectionrateinlicenseplaterecognition,researchersin10and13hadsuggestedacombinationoffeatures.Forinstance,apromisingresultforcombinationofcolourandedgehasbeenreportedin14.Moreover,9hasreportedthattheuseofsimplegeometricalfeaturessuchasshape,aspectratio,andsizeareenoughtofindgenuinelicenseplate.Howevertheresearchersfaceproblemsuchasclutterpartsintheimageandovercomeitwithedgedensity.Edgefeaturesofthecarimageareveryimportant,andedgedensitycanbeusedtosuccessfullydetectanumberplatelocationduetothecharacteristicsofthenumberplate9.Theedgedensityfeatureshadbeenusedin9,10,13becausethedensityofverticaledgesatthelicenseplateareaisconsiderablyhigherthanitsneighbourhood.Inaddition,thisfeatureismorereliableandabletoreduceprocessingtime.Littlecomputationaltimeisoneofimportantelementinrecognitionespeciallyinreal-timedetection.However,thereisalwaystrade-offbetweenthenumberoffeaturesusedinthesystemandthecomputationaltime9,13.Forthis,weproposethefollowingalgorithm,wherethepseudo-codecanbesimplifiedasthefollowing:Tocomparetheverticalandhorizontalhistogramingettingtherequiredfeatures.toextractthemeaningfulimagebasedonthefeaturesselectedThen,thehorizontalandverticalhistogramsarecombinedtogetthematchingregionofalicenseplateiskeptascandidateregionoralsoknownasmeaningfulimage.TheexampleoutputcanbeviewedasFigure6andFigureFigure6.ImageoutputafterfeatureselectionprocessFigure7.ImageoutputafterfeatureextractionprocessTheoutputfromtheextractionprocesswillbeusedinthenextstagewhichisimagesegmentation.V.IMAGESEGMENTATIONOneofthemostpopulartopicsinimageprocessingstudyisimagesegmentation.Thesegmentationprocessbecomesimportanttotheprocessingoftheimagetofindthemeaningfulinformationwhereitcomesfromthemeaningfulregionswhichrepresenthigherlevelofdata.Theanalysisofimagerequireslargeamountoflowlevelofdatawhichisinpixeltobeextractedintomeaningfulinformation.Higher-levelobjectpropertiescanbeincorporatedintosegmentationprocess,aftercompletingcertainpreliminarysegmentationprocess.Examplesofhigher-levelpropertiesareasfollow:i.shape,orii.colourfeaturesThen,itcomestothegoalofsegmentationwhichistofindregionsthatrepresentmeaningfulpartsofobjects.Insegmentation,theimagewillbedividedintoregionsbasedontheinterestofthestudy.Imagesegmentationmethodswilllookforobjectsthateitherhavesomemeasureofhomogeneity(withinthemselves),orcontrast(withtheobjectsontheirborder).Mostimagesegmentationalgorithmscanbedividedasthefollowing:i.modifications,ii.extensions,binationofthese2basicconceptsClassically,Umbaughin15divideimagesegmentationtechniquesintothree(3)whichare:i.Regiongrowingandshrinking:subsetofclusteringii.Clusteringmethodsiii.Boundarydetection:extensionsoftheedgedetectiontechniquesAtthesamepoint,HaralickandShapiro16categorizedimagesegmentationtechniquesintosix(6)whichare:i.Measurementspaceguidedspatialclusteringii.Singlelinkageregiongrowingschemesiii.Hybridlinkageregiongrowingschemesiv.Centroidlinkageregiongrowingschemesv.Spatialclusteringschemesvi.SplitandmergeschemesClusteringisoneofthesegmentationtechniqueasHaralickandShapiro16differentiatedclusteringandsegmentationsuchasfollow:i.Inclustering:thegroupingisdoneinmeasurementspaceii.Insegmentation:thegroupingisdoneinthespatialdomainoftheimageClusteringtechniquescanbeusedtoanydomain,eg:anyN-dimensionalcolororfeaturespace,includingspatialdomainscoordinates.Thistechniquesegmentstheimagebyplacingsimilarelementsintogroups,orclusters,basedonsomesimilaritymeasure.Clusteringisdifferfromregiongrowingandshrinkingmethods,wherethemathematicalspaceusedforclustering.Thedetailsofeachmethodsinsegmentationareexplainedinthenextsections.A.ThresholdingThresholdingisoneofthesimplestandmostpopularmethodinimagesegmentation.Twocommontypesofthresholdingareoutlinedasfollow:i.Localthresholdingisreferredwhenanimageispartitionedintosubregions,andeachsubregioncarrydifferentvalueofthreshold.Localthresholdmethodalsocalledasadaptivethresholdingschemes17-19.ii.Globalthresholdingisreferringtoassigningonlyonethresholdvaluetotheentireimage.Thresholdingtechniquesalsocanbecategorizedintotwolevels:i.Bilevelthresholding:theimageistwo(2)regionswhichareobject(black)andbackground(white).ii.Multithresholding:theimageiscomposedoffewobjectswithdifferentsurfacecharacteristicsthusneedmultiplevalueofthreshold.Thresholdingalsocanbeanalyzedasclassificationproblem,suchthatclassifiyingbilevelsegmentationofanimageintoobjectandbackground.Amongthemostcommonmethodsfoundforthresholdinginimagesegmentationarelistedasthefollowing:i.maximumentropymethod20-22ii.Otsusmethod(maximumvariance)23-26iii.k-meansclustering27-35B.EdgeDetectionTherewillbeedgeandlinedetectioninsegmentationtodivideregionsintomeaningfulinformation.Edgedetectiontechniques:Linedetection/linefinding=Houghtransform37.Houghtransformisdesignedspecificallytofindlines.Alineisacollectionofedgepoints(thatareadjacentandhavethesamedirection).TheHoughalgorithmwilltakeacollectionoffewedgepoints.Edgedetectiontechniques38-53havebeenusedasthebaseofanothersegmentationtechnique.Basically,edgedetectionisalsoanindependentprocessinimageprocessing.Edgedetection,orsometimesitiscalledasedgefindingisalsocloselyrelatedtoregiondetection.Weneedtofindtheregionboundariesfirstbeforewecanproceedtosegmentanobjectfromanimage.Thisisbecausetheedgesidentifiedbyedgedetectionarefrequentlydisconnected.Itmeansthatwehavetofindtheboundariesinordertogettheedges.Insegmentation,linedetectionisdonetodivideregionsintomeaningfulinformation.OneoflinedetectiontechniqueisHoughtransform.Houghtransformisdesignedspecificallytodetectlines.Alineisacollectionofedgepoints(thatareadjacentandhavethesamedirection).TheHoughalgorithmwilltakeacollectionoffewedgepoints.C.Region-basedimagesegmentationThistechniqueattempttoclassifyaparticularimageintoseveralregionsorclassesaccordingtothecommonpropertiesoftheimage.Therearefewpropertiesconsideredforthisprocesswhicharepatternandtexture,intensityvaluesandspectralprofilesoftheimage.Inthismethod,wewanttogrouptheregionssothateachofthepixelsintheregionwillhavesimilarvalueoftheproperties.Therearemanyrealapplicationsusedthismethodsuchasremotesensing,2Dand3Dimages54-55whiletherearevariousmodelsandalgorithmsusedforthistechniquesuchasMarkovRandomFieldModel56-60andMumford-ShahAlgorithm61-64.D.Compression-basedmethodsInthismethod,segmentationwillbedoneinawaytheimagewillbecompressedbasedonthesimilarityofthepatternsoftexturesorboundaryshapeoftheimage.Thismethodaimstominimizethelengthofthedatawheretheoptimalsegmentationcanbeachieved.TherearefewwaysonhowtocalculatethecodinglengthofthedatasuchasHuffmancodingorMDL(MinimumDescriptionLength)principle65-66,wheretheycanbefoundinpreviousstudies67-70.E.Histogram-basedmethodsHistogram-basedmethod71-76isoneofthefrequentlyusedforimagesegmentationtechniques.Inthismethod,wewillproduceaverticalandahorizontalhistogramaccordingly.Thisprocessistogetagroupofpixelsinverticalandhorizontalregionswheretheywillleadtodistinguishingthegraylevelsoftheimage.Incommon,animagewillhavetworegions:backgroundandobject.Normally,thebackgroundisassignedasonegraylevelwhilestheobject(oralsocalledassubject)isanothergraylevel.Usually,backgroundwillsecurethelargestpartoftheimagesothegraylevelofitwillhavelargerpeakinthehistogramcomparedtotheobjectoftheimage.F.Region-growingmethodsRegionGrowingandShrinking77-99techniqueuserowandcolumn(r,c)basedimagedomain.Itcanbeconsideredassubsetofclusteringmethods,butlimitedtospatialdomain.Themethodscanbe:Local:operatingonsmallneighbourhoods,orGlobal:operatingontheentireimage,orCombinationofbothG.Split-and-mergemethodsThereisanalternativeforsegmentationmethodcalledsplitandmerge100-108.Splitandmergeisalsocalledasquadtreesegmentationwhereitbasedonquadtreepartition.Thedatastructureusedinsplitandmergeiscalledquadtreewhereatreewhichhasnodesandeachnodecanhavefourchildren.Itdividesregionsthatdonotpassahomogeneitytest,andcombinesregionsthatpassthehomogeneitytest.Forthis,weproposethefollowingalgorithm,wherethepseudo-codecanbesimplifiedasthefollowing:togetwidthofy-axisoftheimagetodivideintosubregiontogetwidthofy-axisoftheimagetodivideintosubregiontodivideintosubregiontoremoveblankspacetogetsamesizeafterregionhasbeendividedTheexampleoutputcanbeviewedasFigure8Figure8.ImageoutputaftersegmentationprocessTowrapup,agoodsegmentationprocessshouldturnoutuniformandhomogeneousregionswithrespecttosomecharacteristicssuchasgraytoneortextureaswellassimpleregionswithoutmanysmallholes.Theoutputfromthesegmentationwillbeusedinthenextstagewhichischaracterrecognition.CHARACTERRECOGNITIONCharacterrecognitionisthemostimportanttaskinrecognizingtheplatenumber109.Therecognitionofcharactershasbeenaproblemthathasrecei
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2025绿色农产品买卖合同
- 部编统编一下语文口语交际请你帮个忙名师教学设计公开课教案(2025-2026学年)
- 二手房屋买卖标准合同
- Unit 5 Experiencing Literature说课稿-2025-2026学年高中英语冀教版必修二-冀教版2004
- 主题班会教案主题班会我想飞得更高教案(2025-2026学年)
- 选听 心依恋说课稿-2025-2026学年初中音乐人教版七年级下册-人教版
- 小海豚顶球规则游戏公开课教案教学设计(2025-2026学年)
- Unit 5 What were you doingSection A (3a-3c) 教学设计2023-2024学年人教版英语八年级下册
- 跨学科实践活动:栽种草莓探究影响草莓生长的环境条件 说课稿-2024-2025学年苏科版生物七年级下册
- 浙教版科学八上1.3 水的浮力 配套说课稿
- 农民田间学校
- 各类应急演练方案脚本大全
- 雅思词汇一本全(打印珍藏版)
- 高速磁浮大跨度桥梁设计关键技术介绍
- 红蓝简明万人计划青年拔尖人才答辩PPT模板
- DB23T 2550-2020 政务服务大厅建设和管理规范
- 生态系统服务功能与生态保护
- 基于PLC的物料分拣系统设计论文
- Kobe·Bryant科比英文介绍教学课件
- 《教育统计与测量》笔记(一).
- 统编版六年级语文上册第10课《竹节人》优质课件
评论
0/150
提交评论