基于BP神经网络的车牌识别技术研究(英文版).doc

基于BP神经网络的车牌识别技术研究(英文版).doc

《基于BP神经网络的车牌识别技术研究(英文版).doc》由会员分享，可在线阅读，更多相关《基于BP神经网络的车牌识别技术研究(英文版).doc（7页珍藏版）》请在冰点文库上搜索。

ResearchonlicenseplaterecognitiontechnologybasedonBPneuralnetwork

Withthecontinuousdevelopmentofscienceandtechnology,meansoftrafficmanagementisfrommanualmanagementgraduallytransformedintoautomaticallyorsemiautomatically,licenseplaterecognitionasoneofthekeyandhotissuesintheresearchfieldofmoderntrafficengineeringbymoreandmorepeople'sattention.Inrecentyears,neuralnetworkshavebeenappliedinmanyfields,andthecharacteristicsofneuralnetworksareusedtomakethecharacterrecognitionbasedonBPneuralnetwork.

Thisarticlethroughtoinlicenseplaterecognitionsystemimagepreprocessing,fourkeysteps:

licenseplatelocation,charactersegmentationandcharacterrecognitionofproposedakindoflicenseplatecharactersbasedonneuralnetworkrecognitionalgorithm.Usedthismethodoflicenseplateimageexperimentswereconductedtoextractthefeatureofthelicenseplatecharactersample,andundertheenvironmentofMATLABonthelicenseplatecharacterrecognitionwassimulated.Theresultsshowedthatthisalgorithmthecharactersonthelicenseplatelocationandsegmentationhasgoodeffect,thelicenseplatecharacterrecognitionwithcertainaccuracy.

Keywords:

BPneuralnetwork;licenseplatelocation;licenseplaterecognition;charactersegmentation;characterrecognition

1Introduction

Withtheincreaseofthenumberofcars,therearetrafficcongestionintheworld.Inordertosolvethisproblem,manycitieswillbewidenedlane,butstillfarfromsolvingtheproblem.Nottoincreasetheexistingroadfacilities,howtoimprovetheefficiencyoftransportationhasbecomethefocusofresearchintheworld.Intelligenttransportationsystem（Intelligent-TransportationSystemITS）isthemaindevelopmenttrendofthefuturetrafficregulationsystem.Vehiclelicenseplaterecognitiontechnology（License-PlateRecognitionLPR）isoneofthecoretechnologiesinITS.Therefore,theresearchanddevelopmentoflicenseplaterecognitionsystemisofgreatpracticalvalueforthedevelopmentofChina'strafficmanagementfield.

Atpresent,therearestillmanyproblemsinthelicenseplaterecognitionsystem.Recognitionrateisnotpossibletodoonehundredpercent,butwiththedeepeningofresearch,licenseplaterecognitiontechnologywillgraduallymature.Thedevelopmentofmodernintelligenttransportation,makeithasgreatpotentialforapplication,abroadermarket.Atthesametime,neuralnetworkinclassificationproblemsgetwidelyused,forlicenseplaterecognitionproblem,wemustfirstfindthelicenseplatefeatures,andcorrespondingevaluationdata,usingthesedatatotrainneuralnetwork.

Becausetheartificialneuralnetworkhasthecharacteristicsofparallelprocessing,distributedstorageandfaulttolerance,itiswidelyusedintheLPRsystem.Theparallelismofthestructuremakestheinformationstorageoftheneuralnetworkadoptthedistributedmode,thatis,thelicenseplatecharacterinformationisnotstoredinapartofthenetwork,butisdistributedinthenetworkofalltheconnections.Thesefeaturesareboundtomaketheneuralnetworkinthelicenseplaterecognitionofthetwoaspectsoftheperformanceofagoodfaulttolerance:

（1）becauseofthedistributedstorageofthecharactercharacteristicinformation,thewholeperformanceofthevehiclelicenseplaterecognitionsystemwillnotbeaffectedwhensomeoftheneuronsinthenetworkaredamaged.

（2）neuralnetworkthroughprestoredinformationandlearningmechanismsforadaptivetraining,cannevercompletelicenseplateinformationandnoiseofthelicenseplateimagebyLenovotorestorefullmemoriesoftheoriginal,inordertoachievethecorrectidentificationoftheincompleteinputinformation.

Basedontheabovecharacteristics,theapplicationofartificialneuralnetworkinthevehiclelicenseplaterecognitionsystemhasgreatresearchvalue.

2introductiontheprincipleofBPneuralnetwork

BP（backpropagation）networkisproposedthescientistsgroup1986byRumelhartandMcCellandheaded,isakindoferrorinversepropagationtrainingalgorithmforthemultilayerfeedforwardnetworkandiscurrentlythemostwidelyusedmodelsofneuralnetwork.BPnetworkcanlearnandstorealotofinput-outputmodelmapping,withoutpriormathematicsdescribingthismappingequation.Itslearningruleisthesteepestdescentmethodisusedtoadjusttheweightsandthresholdsofthenetworkthroughtheback-propagationnetwork,theminimumerrorsumofsquares.BPneuralnetworktopology,includinginputlayer,hiddenlayer（input）（hidelayer）andoutputlayer（outputlayer）.

2.1BPalgorithm

Theerrorback-propagationalgorithm（BPalgorithm）ofthelearningprocess,bythereverseforwardpropagationanderrorinformationtransmissionconsistsoftwoprocesses.Inputlayerneuronsreceivestheinputinformationfromtheoutsideworld,andpassedtothemiddlelayerneurons;intermediatelayerisinternalinformationprocessinglayerandisresponsiblefortheinformationtransform,accordingtothedemandoftheinformationchanges,themiddlelayercanbedesignedforsinglehiddenlayerormultihiddenlayerstructure;thelasthiddenlayertransfertooutputlayerneurons,afterfurtherprocessing,tocompletealearningforwardpropagationprocess,fromtheoutputlayeroutputtotheoutsideinformationprocessingresults.Whentheactualoutputisnotinconformitywiththeexpectedoutput,thereversepropagationphaseoftheerrorisentered.Theerroriscorrectedbytheoutputlayer,andtheweightofeachlayeriscorrectedbytheerrorgradientdescentmethod.Thecycleofinformationforwardpropagationanderrorbackpropagationprocess,theconstantadjustmentoftheweightsofeachlayer,isthelearningandtrainingofneuralnetworkprocess,thisprocesshasbeencarriedouttonetworkoutputerrorreducedtoanacceptablelevel,orpre-setlearningtimessofar.

3licenseplaterecognitionprinciple

Acompletevehiclelicenseplaterecognitionsystemisdividedintothefollowingfoursteps:

[4].Asshownbelow:

车牌定位

字符分割

图像处理

车牌识别

识别训练

1）imageprocessing:

Nomatterontheimprovementofthelicenseplateimageidentifiabledegree,orsimplifiedlocationandsegmentationofthecharacters,imageconversionanddatacompression,imagecorrectionandimageenhancementprocessingisverynecessary.

（2）licenseplatelocation:

Mainlyincludingtheedgeofthelicenseplateimageextractionandtwovalues,thelicenseplateleveldirectionofthepositioningalgorithm,theverticaldirectionofthelicenseplatelocationalgorithm.Finallydeterminetherelativepositionofthelicenseplateintheentireimage,theoutputoftherectangularlicenseplateimage.

（3）charactersegmentation:

Asinglecharacterisobtainedbyusingthecharacterlocationandsegmentationmethod,whichisusedtodetectthenumberofpixels.

（4）characterrecognition:

Thetemplatematchingmethodisusedtomatchthecharactersintheneuralnetworkdatabasetoconfirmthecharacter,getthefinallicense,includingtheEnglishlettersandnumbers.

4systemdesignandImplementation

Theestablishmentof4.1BPneuralnetwork

BPnetworkisisappliedverywidelyusedasafeedforwardneuralnetwork,issimilartothehumanbrainandhighdegreeofparallelism.Goodfaulttoleranceandassociativememoryfunction,adaptivelearningandfaulttoleranceabilityarestrong,fromthetheoreticalresearchshows,withasinglehiddenlayerneuralnetworkenoughtoperformarbitrarilycomplexfunctionmappingsystem.Therefore,wechoosethehasahiddenlayerofthreelayerBPneuralnetworktorealizethecharacterrecognition.Withartificialneuralnetworkcharacterrecognitionmainlyhastwokindsofmethods:

onemethodistotreatthecharacterrecognitionfeatureextraction,andthentotraintheneuralnetworkclassifierwiththefeature.Theextractionandrecognitioneffectofcharacterfeatures,andcharacterfeatureextractionisoftentime-consuming.Therefore,thecharacterfeatureextractionbecomesthekeyresearch.Theotherwayistomakefulluseofthecharacteristicsoftheneuralnetwork,directlytotheprocessingofimageinputnetwork,automaticallybythenetworktorealizethefeatureextractionandrecognition.Here,Iusedsecondmethodstoidentifythecharacter.

与模板样版进行计算

寻找相关度最大的模块

读入字符

根据模块输出值

Theneuralnetworkiscomposedoftwostages:

（1）learningperiod:

Theconnectionweightsbetweenneuronscanbemodifiedbylearningrulesinordertominimizetheobjectivefunction.

Vehiclelicenseplatecharactersseven,mostlicenseplatefirstChinesecharacters,usuallyrepresentthevehiclebelongstotheprovinces,orisservices,policedon'thavereferredtoasthespecificmeaningofthecharacters,followedbythelettersandnumbers.Licenseplatecharacterrecognitionandgeneralcharacterrecognitionisthatithasalimitednumberofcharacters,atotalofaboutmorethan50Chinesecharacters,26Englishletters,numbers10.Soitisveryconvenienttosetupthecharactertemplatelibrary.

ThelicenseplaterecognitionofChinesecharacters,lettersandnumbers,butthenumberisnotverylarge,butforChinesecharacters,thereisonlya"Su".Lettersandnumbersarethe"numbers"and"letters"setupbytheCSPHOTOSHOPprocess,whicharecollectedfromtheInternet,andusedtobuildthetemplatelibrary.

Chinesecharactersincludedinthelibraryare:

Beijing,Zhejiang,Jiangsu,Henan,Henan,Shaanxi,Shaanxi,Lu,lettersare:

A-Z,thenumberoflibrariesare:

0-9.

（2）workingperiod:

Inthispaper,thenumberofhiddenlayerneuronsis13,andthenumberofoutputneuronsis6