资源描述:
《Text Region Extraction and Text Segmentation on Cameracaptured》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、TextRegionExtractionandTextSegmentationonCamera-capturedDocumentStyleImages12123,444Y.J.Song,K.C.Kim,Y.W.Choi,H.R.Byun,S.H.KimS.Y.Chi,D.K.Jang,Y.K.Chung,1Dept.ofComputerScience,SookmyungWomen’sUniversity,Korea2Dept.ofComputerScience,YonseiUniversity,Korea3Dept.ofComputerScience,UniversityofDenver,US
2、A4VisualRecognitionResearchTeam,ElectronicsandTelecommunicationsResearchInstitute,Koreakimkch@cs.yonsei.ac.kr,ywchoi@sookmyung.ac.kr,seonkim@cs.du.edu,ykchung@etri.re.krABSTRACTbasedmethods,A.K.Jain[4]presentedatextextractionmethodthattreatstextasadistinctivetextureandusedanunsupervisedInthispaper,w
3、eproposeatextextractionmethodfromclusteringtoclassifyeachpixelastextornon-text.However,incamera-captureddocumentstyleimagesandproposeatextvideoframes,naturalscenesliketheleavesofatreeorgrassinasegmentationmethodbasedonacolorclusteringmethod.Thefieldhavetexturessimilartotext,andtextandnon-textoftenpr
4、oposedextractionmethoddetectstextregionsfromtheimagesoverlapinthefeaturespace.usingtwolow-levelimagefeaturesandverifiestheregionsthroughahigh-leveltextstrokefeature.ThetwolevelfeaturesForrecognizingtexts,weneedtoseparatetextsfromthearecombinedhierarchically.Thelow-levelfeaturesareintensitybackground
5、s.Mostofthepreviousapproachesusedbinarizationvariationandcolorvariance.And,weusetextstrokesasahigh-forthispurpose.However,sincethecameraimagescanhavelevelfeatureusingmulti-resolutionwavelettransformsonlocalmanydifferentcolors,mostofthepreviousbinarizationmethodsimageareas.Thestrokefeaturevectorisani
6、nputtoaSVMmaynotprovidedesirableresults.M.Seeger[5]proposeda(SupportVectorMachine)forverification,whenneeded.Theglobalthresholdingmethodfortheentireimage,butsincetheproposedtextsegmentationmethodusescolorclusteringtothepixelsofthetextsareaffectedbyilluminationchanges,non-textsextractedtextregions.We
7、improvedK-meansclusteringmethodcanbeeasilydeterminedastexts.C.Yan[6]proposedmulti-anditselectsKandinitialseedvaluesautomatically.Wetestedstageapproachthatrecursivelybreaksdownanimageintosub-thepropose