naturemediawiki

TouchingTextSegmentationandShapeAnalysisLeKangAdvisor:DavidDoermannUniversityofMaryland,CollegeParkSponsorship:DepartmentofDefense,DARPA,BBN,NationalScienceFoundationMotivationSegmenttouchingtextbetweentextlines.
Needageneralsolutionformultipleenvironments.
ShapematchingandtransformShapecontextcharacterizespointsdistribution,invarianttorotationandscaleThinPlateSpline(TPS)isacoordinatestransformthatwarpsoneshapeintoanotherChallengesNumerousstrokeshapesNumeroustouchingconfigurationIrregularnatureofhandwritingSystemflowApproachConstructadictionaryofLTPsForaninputLTP,findthebestmatchinthedictionaryTransformthetemplate'sknownsegmentationtoobtaintheinputLTP'ssegmentation.
ExamplesDictionaryandexemplarsClusteringusingAffinityPropagation.
deterministicnoneedtoprescribethenumberofclusters.
Exemplars(i.
e.
centroids)representdifferenttouchingconfigurationsPreviousworkEmpiricalmethodsefficientnotrobusttovariationsRecognitionbasedmethodsrequirecandidatesegmentationshavehighcomputationalcost.
ProblemAssumeatextlinesegmentationalgorithmisabletoroughlydetectthetouchinglocationFocusonsegmentingtheLocalTouchingPatterns(LTPs)thatonlycoverpartoftheconnectedcomponentssharedbytwodifferentcharactersDatacollectionReducethespacingbetweentextlinestocreatetouchingLTPsareobtainedwithknownsegmentations(groundtruths).
Email:lekang@umd.
eduhttp://lamp.
cfar.
umd.
edu3348A.
V.
WilliamsBuildingUniversityofMaryland,CollegePark,MD20742EvaluationSimilartoICDARtextsegmentationcontestcriterionfulldictionaryexemplar-onlydictionarySummaryTheproposedapproachprovidesaextensibleframeworkforsegmentingtouchingcomponents.
Varioustouchingpatchesarereasonablysegmentedwithoutad-hocheuristics.
Forfuturework,wewillincorporatepriorknowledgeofshapesforefficientsegmentationofsimpletouchingpatterns.
EvaluationontrainingdatainputLTP&warpedtemplatecomponentsmatchingtemplateresultsegmentationgroundtruth