英文原文1.IntroductionNowadayslicenseplaterecognitionbecomesakeytechniquetomanyautomatedtransportsystemssuchasroadtrafficmonitoring,automaticpaymentoftollsonhighwaysorbridgesandparkinglotsaccesscontrol.Licenseplatelocationisanessentialandimportantstageinthistechnique,andithasreceivedconsiderableattention.Researchershavefoundmanydiversemethodsoflicenseplatelocation.RodolfoandStefano(2000)devisedamethodbasedonvectorquantization(VQ).VQimagerepresentationisaquadtreerepresentationbythespecificcodingmechanism,anditcangiveasystemsomehintsaboutthecontentsofimageregions,andsuchinformationboostslocationperformance.Parketal.(1999)usedneuralnetworkstolocatelicenseplate.Neuralnetworkscanbeusedasfiltersforanalyzingsmallwindowsofanimageanddecidingwhethereachwindowcontainsalicenseplate,andtheirinputsareHSIvalues；apost-processorcombinesthesefilteredimagesandlocatestheboundingboxesoflicenseplatesintheimage.Besidesneuralnetworks,otherfiltershavebeenconsideredtoo.Forexample,someauthorsusedlinesensitivefilterstoextracttheplateareas.Licenseplatesareidentifiedasimageareaswithhighdensityofratherthindarklinesorcurves.Therefore,localizationishandledlookingforrectangularregionsintheimagecontainingmaximaofresponsetotheselinefilters,whichiscomputedbyacumulativefunction(Luisetal.,1999).Platecharacterscanbedirectidentifiedbyscanningthroughtheinputimageandlookingforportionsoftheimagethatwerenotlinkedtootherpartsoftheimage.Ifanumberofcharactersarefoundtobeinastraightline,theymaymakeupalicenseplate(Limetal.,1998).FuzzylogichasbeenappliedtotheproblemoflocatinglicenseplatebyZimicetal.(1997).Theauthorsmadesomeintuitiverulestodescribethelicenseplate,andgavesomemembershipfunctionsforthefuzzysetsbrightanddark,brightanddarksequencetogetthehorizontalandverticalplatepositions.Butthismethodissensitivetothelicenseplatecolorandbrightnessandneedsmuchprocessingtime.UsingcolorfeaturestolocatelicenseplatehasbeenstudiedbyZhuetal.(2002)andWeietal.(2001),butthesemethodsarenotrobustenoughtothedifferentenvironments.Edgefeaturesofthecarimageareveryimportant,andedgedensitycanbeusedtosuccessfullydetectanumberplatelocationduetothecharacteristicsofthenumberplate.Mingetal.(1996)developedamethodtoimprovetheedgeimagebyeliminatingthehighestandlowestportionsoftheedgedensitytosimplifythewholeimage.Butsomeoftheplateregionidentitywillbelostinthismethod.Thispaperfurtherresearchesthesubjectoflicenseplatelocation.Therectanglelicenseplatecontainsrichedgeandtextureinformation,soweconsideritinitsedgeimagebutverydifferenttoMingetal.(1996).Wefirstenhancetheoriginalcarimagetoboostuptheplatearea,thenextracttheverticaledgeimageusingSobeloperator,andthenremovethebackgroundcurvesandnoiseintheedgeimage,andfinallyslidearectanglewindowtosearchtheplateintheresidualimageandsegmentitoutfromtheoriginalcarimage.Section2describesourmethodoflicenseplatelocation,anditcontainsfourparts:imageenhancement,verticaledgeextraction,backgroundcurveandnoiseremoving,platesearchandsegmentation.ExperimentswiththreesetsofcarimagesareperformedinSection3.Section4givesthediscussionandconclusions.1.TheproposedmethodforlicenseplatelocationAlltheinputcarimageshave384288pixelsand256graylevels,andanexampleimageisgiveninFig.1.Thelicenseplateofthecarconsistsofseveralcharacters(suchasLatinletters,Arabicnumerals,etc.),sotheplateareacontainsrichedgeinformation.Butsometimesthebackgroundofthecarimageholdsmuchedgeinformationtoo.Therearetwofactsthatattractourattention:oneisthatthebackgroundareasaroundthelicenseplatemainlyincludesomehorizontaledges；theotheristhattheedgesinthebackgroundaremainlylongcurvesandrandomnoises,whereastheedgesintheplateareaclustertogetherandproduceintensetexturefeature.Ifonlytheverticaledgesareextractedfromthecarimage(althoughtheplatewilllosealittlehorizontaledgeinformation,thislittlelossistobevaluable)andmostofthebackgroundedgesareremoved,theplateareawillbeisolatedoutdistinctlyinthewholeedgeimage.Thusweproposetolocatethelicenseplateinitsverticaledgeimageasthefollowingfourstages.2.1.ImageenhancementInFig.1,thegradientsinthelicenseplateareaaremuchlowerthanthoseinthecontourareasofthecar,whichiscausedbythecarshadowinthedazzlingsunshine.Thecarimagescapturedinthegloomydaysordimnightsoftenbringoutweakgradientsinplateareastoo.Afewverticaledgeswillappearintheplateareas,ifweextractedgeimagesdirectlyfromthesecarimages.Thereforeitisimportanttoenhancethecarimagesfirstly.Thelocalareasthatneedtobeenhancedinacarimagehavelowvariances.HereweuseIi,jtodenotetheluminanceofthepixelPi,j(row:06i288,column:06j384)inthecarimage,anduseI1i；jtodenotetheluminanceintheenhancedimage.WeletIi,jandI1i；jsatisfyEq.(1),whereWi,jisawindowcenteredonpixelPi,j,IWi；jandrWi；jarethemeanluminanceandstandarddeviationofthepixelsinthewindowWi,j,I0andr0aretheexpectedmeanandstandarddeviation,respectively.0,0,1)(,IIIIjijiwjiwji（1）Inordertorepresentthelocalinformationbetter,thesizeofthewindowshouldbesmallerthantheestimatedsizeoftheplate.Inthispaper,weselecta4836rectangleasthewindowWi,jandthus88windowscancoveroverthewhole384288carimage.LetI0beequaltoIWi；jandr0beaconstantindependentofpixelPi,j.NowweneedtoknowthevaluesIWi；jandrWi；jateachpixel.Computingoutallthevaluesisnotadvisable,andwecanusethebilinearinterpolationalgorithmtogetthem.Firstwecutthecarimageinto88blocksequably；andthencomputeouttheIWi；jandrWi；jvaluesatthevertexesofblocks,wherei=36m,j=48n,m,n=0,1,2,.,8；finallycomputeouteveryIWi；jandrWi；jbythebilinearinterpolationEqs.(2)and(3)(Fig.2),where36m6i36(m+1),48n6j0)Mi,j=maxMi1,j1,Mi1,j,Mi1,j+1,Mi,j1+1；elseMi,j=maxMi2,j1,Mi2,j,Mi2,j+1,Mi1,j2,Mi1,j+2,Mi,j2+1；endendendend3.foreachrowifrombottom-to-topdoforeachcolumnjfr