Image segmentation is a mixture of complex sub-problems. Following the concept of perceptual organization[1], we first discuss categorization of these problems. Then we propose a method to bind most of all pixels in each image regions. It is based on image
ANIMAGESEGMENTATIONANDTRACKINGMETHODBYSEGMENTSCANNING
WAVEPROPAGATIONNETWORK
ToshinaoIshii(a)andKazuoKyuma(b)
Multi-ModulFunctionMitsubishiLaboratory,RealWorldComputingPartnership
c/oAdvancedTechnologyR&DCenter,MitsubishiElectricCorporation
ABSTRACT
Imagesegmentationisamixtureofcomplexsub-problems.Followingtheconceptofperceptualorganization[1],we rstdiscusscategorizationoftheseproblems.Thenweproposeamethodtobindmostofallpixelsineachimageregions.Itisbasedonimageregionscanningdynamicwaveprop-agationandregionlabelinformationrepresentedinproba-bilitydistribution.Thiswavepropagationmethodrealizesboundarynoiserobustnessaswellasitsinterpolation.Thesegmentrepresentationinprobabilitydistributionenhancesstabilityofsegmentationresult.
1.INTRODUCTION
Segmentingimageobjectisanimportantpartofearlyvi-sion.Psychologicalstudiesdiscoveredevidencesthatcer-tainimagesegmentationisdoneintheearlystageandtheyareeffectivetohighlevelvisualrecognition.Oneoftheinterestingdiscoveriesisdescribedin[2].Theycomparedquicknessandaccuracyofrecognizingapathformedbyfragmentsofedgepatches.Accordingtotheirexperimen-talresults,humanrecognizespathbetterwhenthepathisaclosedone.Thisimpliesthathumanearlyvisionisusingsegmentinformation,whichisaglobalproperty,for ndingboundaryofthesegmentandthisprocessisfasterthanjust ndingaopenpathinthenoisyconditionoftheexperiment.
Inarti cialvision,manysegmentationrelatedstudieshavebeendonecorrespondingly.Itisfoundthatproblemofvisualsegmentationisnotamonolithicone.“Perceptualorganization”[1]istheconcepttodescribethatitconsistsofmanystagesandthatdifferenttypesandlevelsofimagefeatureareinvolved.In[1],perceptualorganizationisclas-si edintofourlevelsof(1)signallevel,(2)primitivelevel,(3)structuralleveland(4)assemblylevel.Thesignallevelbelongstoorganizationofrawsignal.Theprimitivelevelconsistsofprocessestoproducerepresentationwhosedi-mensionalityislowerthanrealobjects.Thestructurallevelorganizesrepresentationinthesamedimensionalityasof
(a)ici@qua.crl.melco.co.jp(b)kyuma@ar.sow.melco.co.jp
objects.Assemblylevelorganizesglobalinformationforestimatingspatialrelationsofobjects.
Inthisarticle,weproposeamethodofimagesegmenta-tionintheprimitivelevel.Themethodisbasedonthelocalimagefeaturessuchasedgeintensityandtheoutputofthemethodisregionlabelofpixels.Thisisessentiallyequiv-alenttobindingallpixelsofeachimagesegments.Themethodproposedinthisarticlehaspropertiesofnoisero-bustnessandboundaryinterpolation.Theyarecrucialbe-causenoiseandincompletenessofboundaryofobjectsinitsprojectiontoimagedataisunavoidable.
Thecomputationalmodelofthemethodconsistsoftwolayersofnetwork.Thelowerleveltwodimensionalnet-workcorrespondstolocalimagefeaturemapandthehigherlevelonecorrespondstoregionlabels.Inthelowerlevelnetwork,cellshavenon-uniformlocalconnectionsthroughwhichcertainpatternpropagatesonthistwodimensional eld.Thelocalconnectionissoorganizedthatpatterndoesnotpropagateeasilyoverimagesegmentboundaries.Inthehigherlevelnetwork,informationofthescanningsofpat-ternpropagationsisintegratedtobindallthecellsineachsegments.
2.THEPRIMITIVELEVELIMAGE
SEGMENTATION
Inthisarticleweprescribeimagesegmentintheprimitivelevelasspatiallyconnectedareainwhichcertainimagefea-tureiscontinuous.Sinceourtargetofprocessisrealworldimage,wecannotde nestrictlythede nitionofconnectedareaorcontinuityofimagefeaturebecauserealworldim-ageistoovaluable.Inotherwords,evenincasethere’srealsegmentsin3Dphysicalspace,itsopticalmappingto2Dspaceincludesverycomplexnoisesandtheexistingbound-aryismappedincompletelytothe2Ddata eld.
Wecategorizethesourceofincompletenessintothreetypes.Theyareoptical,structuralandthehigherlevelin-completeness.Structuralsegmentboundaryincompletenessoriginatesfromthedimensionaldegeneration.Forexample,asurfacepartiallyoccludedbyotherobjectscanbeseenastwoormoredifferentsegments.Thehigherlevelisknowl-
什么时候用an_英语学习_外语学习_教育专区。-什么时候用 an,什么时候用 a 1.( )The old lady in brown is ( ) university professor. A.a B.an C.the...
不定冠词a/an的区别 特别提醒:归纳: a用在以辅音音素开头的单词前,an用在以 元音音素开头的单词前,判断一个单词是元音开 头还是辅音开头,要根据其读音,而不...
一个”特别指定的目标物前面 通常第一次提到一样东西或一个人的时候,我们是在介绍全新或者是听者不知 道的人事物,因此不需要特别指定只要用不定冠词 a/an 就...
拼写以辅音字母开头读音却以元音开头的单词 hour, honest, honor 等单词的拼写虽然以辅音字母 h 开头,但其读音却以元 音开头,因此,前面要用 an。可熟记下面这...
不定冠词a an的用法_英语学习_外语学习_教育专区。不定冠词 a\an 的用法冠词分不定冠词(The Indefinite Article)和定冠词(The Definite Article)a, an 是不...
a与an... the的用法_中考_初中教育_教育专区。a 与 an... the 的用法 a 用于辅音前;an 用于元音前。a 是用在辅音开始的词前,an 是用在元音开始的词...
an ang教案_一年级语文_语文_小学教育_教育专区。an ang 一 设计理念 an 是前鼻韵母,ang 是后鼻韵母。教材把 an 和 ang 放到了一课来学习,目 的是通过...
n ? 2? , 求通项公式 an 3、若数列的递推公式为 a1 ? 3, an?1 ? an ? 2 ? 3n?1 (n ? N * ) ,则求这个数列的通项公式 a n ?1 ? a ...
在英语中, a, an和 the称为冠词,其中 a和 an是不定冠词,the是定冠词。今天 我们先了解不定冠词。不定冠词a和an 都可以表示“一(个、支、张、块……)”...
Lesson 5-Make an appointment 商务英语-约会_英语学习_外语学习_教育专区。REVIEW Conversation 1 ? ? ? ? 你好,我想找苏珊。 不好意思,她现在不在,可以留言...
热门文档