资源描述：
《TEXT LOCALIZATION USING IMAGE CUES AND TEXT LINE INFORMATION》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、Proceedingsof2010IEEE17thInternationalConferenceonImageProcessingSeptember26-29,2010,HongKongTEXTLOCALIZATIONUSINGIMAGECUESANDTEXTLINEINFORMATIONToanNguyenDinh,JonghyunPark,GueesangLeeDepartmentofElectronicsandComputerEngineering,ChonnamNationalUniversity,Kor
2、eaABSTRACTTextlocalizationisanimportanttaskintextunderstandingsystems.Inthispaper,anefficientconnectedcomponentbasedtextlocalizationmethodisproposed.Twoimagecues,edgeandcolor,arecombinedtoextractcandidatetextregions.Thenon-textregionsinthesecandidatetextregio
3、nsarethenremovedbythetextlineinformationgeneratedfrom2Dtensorvoting.TheimagesinICDAR2003competitiondatasetareusedtoevaluatetheproposedmethod.Theexperimentalresultsshowthattheproposedmethodextractsmoretextregionswithalowfalsepositiverate.IndexTerms—textlocaliz
4、ation,imagecues,tensorvoting,connectedcomponentFig.1.Proposedmethodflowchart.1.INTRODUCTIONcomponentsarehardtobesegmentedaccuratelybyusingasingleimagecuesuchasedge.TextlocalizationisanimportantpartintextimageInthispaper,anovelCC-basedtextlocalizationmethodinf
5、ormationextractionsystems[1].Textlocalizationinisproposed.Toovercometheseproblems,wecombinebothnaturalsceneimagesisachallengingtaskduetotheedgeandcolortogeneratemoretextregions.Insteadofclutteredbackground,noises,andvariationsoftexts’size,usingwell-definedheu
6、risticsrulesthathardtobedesignedfont,andorientation.Therearemanytextlocalizationforallimages,thetextlineinformationgeneratedby2Dmethodshavebeenreportedintheliterature.Theycanbetensorvotingisusedtoremovethenon-textregions.Tocategorizedintotwomaincategories:reg
7、ion-basedandevaluatetheperformance,selectedimagesintheICDARconnectedcomponent(CC)-based.Region-basedmethods2003competitiondatasetareused.arebasedontheobservationsthatthecharacteristicsofthetextregions,suchastexture,structure,andcoefficientvaluesintransformedd
8、omains,aredifferentfromthoseofthenon-textregions[2,3].CC-basedmethods[4-6],ontheotherhand,generateasetofseparateconnectedcomponentsbasedontheintensity,edge,andcolordistribution.Thesemetho
