Automatic_License_Plate_Recognition_System1

AutomaticLicensePlateRecognitionSystemBoLi,Zhi-yuanZeng,Hua-liDong,Xiao-mingZengDigitalEngineeringandEmulationResearchCenter，HuazhongUniversityofScienceandTechnology,[e-mail]:licenseplaterecognition,licenseplatelocation,charactersegmentation,characterrecognition,imageprocessingAbstract.Thispaperproposedanalgorithmforlicenseplaterecognitionsystem(LPRS).Theverticaledgewasfirstdetectedbysobelcoloredgedetector.Then,theinvalidedgewasremovedregardingedgedensity.Next,thelicenseplate(LP)imagewasconvertedintoHSVcolormodel,andbyedgedensitytemplateandfuzzycolorinformationjudgement,theLPregionwaslocated.Then,color-reversingjudgementandtiltcorrectionwasconducted.Afterward,charactersweresegmentedbymeansofverticalprojectionandconvolution,bywhichcharacterwidthandpositioncanbeexactlyconfirmed,andcharacterrecognitionwasconductedbasedonradialbasisfunction(RBF)neuralnetwork.Withalotofsamplesverifiedinnighthoursanddaytimeunderrealconditions,theexperimentalresultsshowthattheproposedmethodcanachieveaccuracyandeffectivenessinLPRS.IntroductionLPRSisveryimportantinintelligenttransportationsystem.Itconsistsofthreepartsnamely,LPlocation,charactersegmentationandcharacterrecognition.TherearemanyalgorithmsforLPlocation,amongthem,widelyusedareasfollows,firstly,algorithmbasedonLPbackgroundcolororcoloredge[1].Secondly,algorithmbasedonthecharacteristicofedgeandtexture,levelgrayscalejumping[2].Thirdly,algorithmonthebasisofmathematicalmorphology[2].Also,othermethods,suchasalgorithmbasedonthecornerfeatureinLP[3],andonthewaveletanalysis[4],etc.Forcharactersegmentation,therearemanyalgorithmslikeprojection[5],methodscarriedonthepriorknowledgesuchaswidthofcharacter[5].Andforcharacterrecognition,theback-propagationneuralnetworkisoftenused[6],furthermore,supportvectormachineisalsooftenemployed[7].TheaimofthisstudyistopresentanalgorithmforLPRS.EdgedetectionandinvalidedgeremovalInLPregion,animportantcharacteristicistheregulararrangementofverticaltexture,therefore,verticaledgewasdetectedbysobelcoloredgedetector,andtheinvalidedgewasremovedonthebasisofedgedensitytemplate.Then,thecandidateLPregionswerejudgedbyedgedensityandcolorinformation.ThealgorithmisshownasFig.1.Fig.1TheflowchartofalgorithmAppliedMechanicsandMaterialsVols.20-23(2010)pp438-444Onlineavailablesince2010/Jan/12at©(2010)TransTechPublications,Switzerlanddoi:10.4028/(ID:116.251.214.47,NationalChengKungUniversity,Tainan,Taiwan-20/02/14,06:25:15)Verticalsobelcoloredgedetection.BecauseitismainlyverticaldirectionedgeinLPregion,thispaperchosethesobelverticalkernelG,whichisdesignedtorespondmaximallytoedgesrunningverticallyrelativetothepixelgrid,theGconvolutionmaskisappliedtoeverycolorchannel.1100011kGk=−−−(1)Where,kcanbeassigned2.Apixelisregardedasedgepointifthemeanofthegradientmagnitudevaluesinthethreecolorchannelsexceedsagiventhreshold.Invalidedgeelimination.Generallyspeaking,thecharacteristicsoftextureandgrayscaleinLPareapparentlydifferentfromotherregions,forinstance,itisobviousthatgrayscaledifferenceexistsbetweenthecharacterandbackground.AndedgedensityinLPisstableandmuchmorethananyotherregions.Also,thecolorinLPisobviouslydifferentfromotherregions.ThereareplentyofinvalidedgesoutsideLPregionincarimage,whichoftendisturbthejudgementforLP.Therefore,somestepsweretakentoeliminatetheinvalidanddisturbingedge.Scanningfrombottomtotopinanimage,foreveryrow:Adoptahorizontaledgedensitytemplate,thenmovefromlefttoright,iftheedgedensityinthetemplateislessthanThreshold1,then,thecenterofthetemplateisassigned0;Fortheirow,theThreshold1isself-adaptivelyadjustedby(2):1((,)),(0,)TemplateWidthThresholdSumedgeijiWidthWidth=∈(2)Where,Sumstandsforthesumofthenumberofedgepoints,Widthisthewidthofwholeimage,andWidthTemplateisthewidthofedgedensitytemplate.Theinvalidanddisturbingedgewasremovedgreatlyaftertheprocessing.LicenseplatelocatingTheLPlocatingmethodexploitsthefactthat,theedgesaredenserintheregioncontainingtheLPthananyotherregion,andthecolorinLPisspecifically,inChina,therearefourkindsofcolorofLP,thatis,blue,yellow,white,black.Thecandidateregionjudgementismadebymeansof1)edgedensitytemplatematching,2)colorjudgement.Thealgorithmisasfollows:Aftertheaboveprocessing,muchnoiseandinvalidedgewereremoved.Then,Scanfrombottomtotopoftheimage,andtakealinearityedgedensitytemplatewhosewidthisabout1/20ofwidthofthewholeimage,andsearchtheregionswhichedgedensityisstableandobviouslyhigherthanotherregions.Then,keepthebordervalueofthecandidateregion.SincethecolorjudgeforLPisoftennotunambiguous,uncertaintiescouldbetakenintoaccountduringthecomputationofthecharacteristicmap.Fuzzysetshavebeenknowntoprovideagoodtoolformodelinguncertainties.Inthisstudy,weintroducefuzzyintotheentriesofthecolorcharacteristicmapandrefertotheresultasthefuzzymap.ThecolorfuzzymapwasconductedinHSVmodelaccordingtothecolorinLP.ThetransformfromRGBtoHSVisasfollo