基于BP神经网络的车型识别外文翻译.docx

基于BP神经网络的车型识别外文翻译.docx

《基于BP神经网络的车型识别外文翻译.docx》由会员分享，可在线阅读，更多相关《基于BP神经网络的车型识别外文翻译.docx（22页珍藏版）》请在冰点文库上搜索。

基于BP神经网络的车型识别外文翻译

一、外文资料

LicensePlateRecognitionBasedOnPriorKnowledge

Abstract-Inthispaper,anewalgorithmbasedonimprovedBP（backpropagation）neuralnetworkforChinesevehiclelicenseplaterecognition（LPR）isdescribed.Theproposedapproachprovidesasolutionforthevehiclelicenseplates（VLP）whichweredegradedseverely.Whatitremarkablydiffersfromthetraditionalmethodsistheapplicationofpriorknowledgeoflicenseplatetotheprocedureoflocation,segmentationandrecognition.Colorcollocationisusedtolocatethelicenseplateintheimage.Dimensionsofeachcharacterareconstant,whichisusedtosegmentthecharacterofVLPs.TheLayoutoftheChineseVLPisanimportantfeature,whichisusedtoconstructaclassifierforrecognizing.Theexperimentalresultsshowthattheimprovedalgorithmiseffectiveundertheconditionthatthelicenseplatesweredegradedseverely.

IndexTerms-Licenseplaterecognition,priorknowledge,vehiclelicenseplates,neuralnetwork.

I.INTRODUCTION

VehicleLicense-Plate（VLP）recognitionisaveryinterestingbutdifficultproblem.Itisimportantinanumberofapplicationssuchasweight-and-speed-limit,redtrafficinfringement,roadsurveysandparksecurity[1].VLPrecognitionsystemconsistsoftheplatelocation,thecharacterssegmentation,andthecharactersrecognition.Thesetasksbecomemoresophisticatedwhendealingwithplateimagestakeninvariousinclinedanglesorundervariouslighting,weatherconditionandcleanlinessoftheplate.Becausethisproblemisusuallyusedinreal-timesystems,itrequiresnotonlyaccuracybutalsofastprocessing.MostexistingVLPrecognitionmethods[2],[3],[4],[5]reducethecomplexityandincreasetherecognitionratebyusingsomespecificfeaturesoflocalVLPsandestablishingsomeconstrainsontheposition,distancefromthecameratovehicles,andtheinclinedangles.Inaddition,neuralnetworkwasusedtoincreasetherecognitionrate[6],[7]butthetraditionalrecognitionmethodsseldomconsiderthepriorknowledgeofthelocalVLPs.Inthispaper,weproposedanewimprovedlearningmethodofBPalgorithmbasedonspecificfeaturesofChineseVLPs.TheproposedalgorithmovercomesthelowspeedconvergenceofBPneuralnetwork[8]andremarkableincreasestherecognitionrateespeciallyundertheconditionthatthelicenseplateimagesweredegradeseverely.

II.SPECIFICFEATURESOFCHINESEVLPS

A.Dimensions

Accordingtotheguidelineforvehicleinspection[9],alllicenseplatesmustberectangularandhavethedimensionsandhaveall7characterswritteninasingleline.Underpracticalenvironments,thedistancefromthecameratovehiclesandtheinclinedanglesareconstant,soallcharactersofthelicenseplatehaveafixedwidth,andthedistancebetweenthemediumaxesoftwoadjoiningcharactersisfixedandtheratiobetweenwidthandheightisnearlyconstant.Thosefeaturescanbeusedtolocatetheplateandsegmenttheindividualcharacter.

B.Colorcollocationoftheplate

TherearefourkindsofcolorcollocationfortheChinesevehiclelicenseplate.ThesecolorcollocationsareshownintableI.

TABLEI

Categoryoflicenseplate

Colorcollocation

smallhorsepowerplate

bluebackgroundandwhitecharacters

motortruckplate

yellowbackgroundandblackcharacters

militaryvehicleandpolicewagonplate

blackbackgroundandthewhitecharacters

embassyvehicleplate

whitebackgroundandblackcharacters

Moreover,militaryvehicleandpolicewagonplatescontainaredcharacterwhichbelongstoaspecificcharacterset.Thisfeaturecanbeusedtoimprovetherecognitionrate.

C.LayoutoftheChineseVLPS

ThecriterionofthevehiclelicenseplatedefinesthecharacterslayoutofChineselicenseplate.AllstandardlicenseplatescontainChinesecharacters,numbersandletterswhichareshowninFig.1.ThefirstoneisaChinesecharacterwhichisanabbreviationofChineseprovinces.ThesecondoneisaletterrangingfromAtoZexcepttheletterI.Thethirdandfourthonesarelettersornumbers.Thefifthtoseventhonesarenumbersrangingfrom0to9only.Howeverthefirstortheseventhonesmayberedcharactersinspecialplates（asshowninFig.1）.Aftersegmentationprocesstheindividualcharacterisextracted.Takingadvantageofthelayoutandcolorcollocationpriorknowledge,theindividualcharacterwillenteroneoftheclasses:

abbreviationsofChineseprovincesset,lettersset,lettersornumbersset,numberset,specialcharactersset.

（a）Typicallayout

（b）Specialcharacter

Fig.1ThelayoutoftheChineselicenseplate

III.THEPROPOSEDALGORITHM

Thisalgorithmconsistsoffourmodules:

VLPlocation,charactersegmentation,characterclassificationandcharacterrecognition.ThemainstepsoftheflowchartofLPRsystemareshowninFig.2.

Firstlythelicenseplateislocatedinaninputimageandcharactersaresegmented.Theneveryindividualcharacterimageenterstheclassifiertodecidewhichclassitbelongsto,andfinallytheBPnetworkdecideswhichcharacterthecharacterimagerepresents.

Imageacquisition

Platelocation

Characterssegmentationsegmentation

classifier

Chinese

character

Letter

Letteror

number

Number

Specialcharacter

Charactersrecognition

Fig.2TheflowchartofLPRsystem

A.Preprocessingthelicenseplate

1）VLPLocation

Thisprocesssufficientlyutilizesthecolorfeaturesuchascolorcollocation,colorcentersanddistributionintheplateregion,whicharedescribedinsectionII.Thesecolorfeaturescanbeusedtoeliminatethedisturbanceofthefakeplate’sregions.TheflowchartoftheplatelocationisshowninFig.3.

Charactersedgedetection

Binaryimagesegmenting

Candidateimagedetection

Vehicleplateextraction

Fig.3Theflowchartoftheplatelocationalgorithm

Theregionswhichstructureandtexturesimilartothevehicleplateareextracted.Theprocessisdescribedasfollowed:

（1）

（2）

Here,theGaussianvariance

issettobelessthanW/3（Wisthecharacterstrokewidth）,so

getsitsmaximumvalueMatthecenterofthestroke.Afterconvolution,binarizationisperformedaccordingtoathresholdwhichequalsT*M（T<0.5）.Medianfilterisusedtopreservetheedgegradientandeliminateisolatednoiseofthebinaryimage.AnN*Nrectanglemedianfilterisset,andNrepresentstheoddintegermostlyclosetoW.

Morphologyclosingoperationcanbeusedtoextractthecandidateregion.Theconfidencedegreeofcandidateregionforbeingalicenseplateisverifiedaccordingtotheaspectratioandareas.Here,theaspectratioissetbetween1.5and4forthereasonofinclination.Thepriorknowledgeofcolorcollocationisusedtolocateplateregionexactly.ThelocatingprocessofthelicenseplateisshowninFig.4.

Fig.4Thewholeprocessoflocatinglicenseplate

2）Charactersegmentation

Thispartpresentsanalgorithmforcharactersegmentationbasedonpriorknowledge,usingcharacterwidth,fixednumberofcharacters,theratioofheighttowidthofacharacter,andsoon.TheflowchartofthecharactersegmentationisshowninFig.5.

Fig.5Theflowchartofthecharactersegmentation

Firstly,preprocessthelicensetheplateimage,suchasunevenilluminationcorrection,contrastenhancement,inclinecorrectionandedgeenhancementoperations;secondly,eliminatingspacemarkwhichappearsbetweenthesecondcharacterandthethirdcharacter;thirdly,mergingthesegmentedfragmentsofthecharacters.InChina,allstandardlicenseplatescontainonly7characters（seeFig.1）.Ifthenumberofsegmentedcharactersislargerthanseven,themergingprocessmustbeperformed.TableIIshowsthemergingprocess.Finally,extractingtheindividualcharacter’imagebasedonthenumberandthewidthofthecharacter.Fig.6showsthesegmentationresults.（a）Theinclineandbrokenplateimage,（b）theinclineanddistortplateimage,（c）theseriousfadeplateimage,（d）thesmutlicenseplateimage.

TABLEII

GetNf

IfNF>MaxF

Foreachcharactersegments

Calculatethemediumpoint

Foreachtwoconsecutivemediumpoints

Calculatethedistance

Calculatetheminimumdistance

Mergethecharactersegmentkandthecharactersegmentk+1

NF=NF-1

Endofalgorithm

whereNfisthenumberofcharactersegments,MaxFisthenumberofthelicenseplate,andiistheindexofeachcharactersegment.

Themediumpointofeachsegmentedcharacterisdeterminedby:

（3）

where

istheinitialcoordinatesforthecharactersegment,and

isthefinalcoordinateforthecharactersegment.Thedistancebetweentwoconsecutivemediumpointsiscalculatedby:

（4）

Fig.6Thesegmentationresults

B.Usingspecificpriorknowledgeforrecognition

ThelayoutoftheChineseVLPisanimportantfeature（asdescribedinthesectionII）,whichcanbeusedtoconstructaclassifierforrecognizing.Therecognizingprocedureadoptedconjugategradientdescentfastlearningmethod,whichisanimprovedlearningmethodofBPneuralnetwork[10].Conjugategradientdescent,whichemploysaseriesoflinesearchesinweightorparameterspace.Onepicksthefirstdescentdirectionandmovesalongthatdirectionuntiltheminimuminerrorisreached.Theseconddescentdirectionisthencomputed:

thisdirectionthe“conjugatedirection”istheonealongwhichthegradientdoesnotchangeitsdirectionwillnot“spoil”thecontributionfromthepreviousdescentiterations.Thisalgorithmadoptedtopology625-35-NasshowninFig.7.Thesizeofinputvalueis625（25*25）andinitialweightsarewithrandomvalues,desiredoutputvalueshavethesamefeaturewiththeinputvalues.

Fig.7Thenetworktopology

AsFig.7shows,thereisathree-l