DOC

Event

By Norma Sanders,2014-07-23 21:04
16 views 0
Eventevent,Event,eVENT,EVENT

    Event

Jun2008,Volume5,No.6(SerialNo.43)JournalofCommunicationandComputer,ISSN154

    8?7709,USA

    ;EventsemanticrecognizingbasedonMarkovchain

    ;ZENGCheng,OUWei-jie,CUIXiao-jun

    ;(StateKey6ofSoftwareEngineering,WuhanUniversity,Wuhan430072,ChinaJ ;Abstract:Owingtotheenormouslnformationandcomplex ;structure,videosemanticprocessingisatrickyissueallalong. ;Currentresearchesarerestrictedwithinrecognizingrelative ;simplesemanticinsomecertaindomains.Thispaperbrings ;forwardanovelmethodtransitinglow-levelfeatureto ;highlevelsemanticwithMarkovchainbystages,whichtakes ;objectsemanticasthecore.Thismethodisvalidforrecognizing ;complexeventsemantic.Semanticconceptmappingmechanism ;basedonsemantictemplateispresentedtorealizetheautomatic ;recognitionofvideosemantic.Intheexperimentcontrasting ;with1BM‟s1MAT.ourmethodshowsmoreextensive

    ;recognitionrangeandhigheraccuracy.Experimentalresultsare ;encouraging,andindicatethattheperformanceoftheproposed ;approachiseffective.

    ;Keywords:objectsemantic;semantictemplate;maximum ;1ikelihoodestimate;Markovchain

    ;1.Introduction

    ;Withtheexplosionofincreasinglyenormous

    ;amountandcapacityofvideodocumentsinInteractand ;PC,howtorapidlyandaccuratelyretrievethevideo ;documentsaccordingtouserrequirementbecomesa

    ;moreandmoreimportantproblem.Atpresent,the

    ;retrievaltechnologybasedonkeywordshasbeenadopted ;widelybymanycommercialsearchengines.Many

    ;retrievalsystemsofimagesorvideobasedoncontent ;havealsocomeintobeinginsoftieresearchinstitutions. ;However,thesetechnologiesdependtoomuchonuser‟s

    ;Acknowledgements:Thisworkwaspartiallysupportedby ;NationalBasicResearchProgramofChina(973ProgramNo. ;2007CB3l08O6),HubeiProvinceNaturalScienceFoundation ;ofChina(No.2007ABAO38).DoctorSubjectFundof

    ;EducationMinistryfNo.20070486064),thelllProjectof ;HighSchoolfNo.B07037).

    ;ZENGCheng(1978.),male,Ph.D.,lecturer;researchfields:

;cross-media,multimediadatamining.

    ;oUWei-jie(198l),male,Ph.D.candidate;researchfield: ;Webservices.

    ;CUIXiao-jun(1972.),male,Ph.D.candidate,associate ;professor;researchfields:deepWeb,datamining. ;participating,orretrievalresultsbetraytheuser ;requirement.Itisurgenttostrengthentheresearchabout ;mediasemanticinformationminingtechnology. ;Videosemanticminingisoneofthecrucial

    ;problemsofintendingcrossmediasearchengine.The

    ;currentresearchesaboutitfocusonminingobject ;semanticinformation,staticscene.Buttheyarerestricted ;withinsomecertainapplicationdomains,anddon‟tadapt

    ;todealwithcomplexsemanticminingsuchasdynamic ;structure,longterm,multiobjects,multirelationsandso

    ;on.Therefore,thispaperpresentsanovelmethodwhich ;integratesobjectsemanticwithrelationshipinformation ;amongobjectstorealizethetransitionfromlowlevel

    ;featuretohighlevelsemantic.Semantictemplate,which ;storesthemappingrelationshipbetweenphysicalfeatures ;andsemantic,isconstructedtoautomaticallyrecognize ;videosemanticinformation.Theseestablishthe ;theoreticalfoundationofvideoorcross.mediaretrieval ;basedonsemantic

    ;2.Relativeresearch

    ;Virage【】isavideoautomaticannotatingsystem

    ;whichprovidesanopenframeworktoexpandother ;videooraudioanalysistools.Butitdoesn‟trefertothe

    ;gapproblembetweenlow??levelfeaturesandhigh--level ;semantic,andpossesslimitedanalysistools. ;V.Mezaris[2correspondslowlevelfeatures

    ;descriptioninMPEG??7torelevantmiddle?-level ;description,andconstructsasimpleworddatabasecalled ;objectontology.M.etaltransformsthevideosemantic ;gapproblemintoprobabilitypatternrecognition. ;Benitez4lpresentsanintelligentinformation

    ;systemframeworkMediaNet,whichintegrates ;

    ;EventsemanticrecognizingbasedonMarkovchain ;low—levcII”eaturesandmediaknowledgeconceptslnto

    ;asinglesystem.Itutilizesentityconceptsinrealworld ;andrelationshipamongthemtosimulaterealworld. ;ASSAVIDproject‟supposedbyEuropeIST

    ;funddevelopsavidcorctrievalsystem.Itrealizesthe

    ;automaticclassificationofsportscenesbyresearching ;specialcharactersofdiffcrentsports.Eachclass ;correspondstoasportnamethatisconvenientforuser ;toretrievebasedonkeywords.Thelastresearchofthis

    61. ;projectisintroducedinreference

    ;Navid,etalfpresentedamethodwhichisableto

    ;recognizesomespecialregions,suchassky,grassand ;soon,anddistinguishbetweenindoorandoutdoor

    ;scenesbylowdimensionscolorfeaturebasedon ;suppoavectormachineandwavelettextureanalysis. ;Multilabelisutilizedtorecognizemoreobjectsand ;classifiesvisualmediacontainingmultiobjectswhich

    ;arepossiblyintersectingeachother.

    ;QianR.describesathreelevelstrategyfor

    ;detectingeventsemantic,correspondingtofeature ;detecting,markdetectingandclassifyingmanipulation, ;respectively.Inthefirstlevel,textureandcolor ;featuresareanalyzedandintegratedwithobjectmotion ;features;inthesecondlevel,neuralnetworkisutilized ;tocombinealllowlevelfeaturesandproducesimple ;sceneannotation.However,thismethodisonly ;adaptedtothespoavideo.

    ;IBMdevelopsaMPEG7videoannotation

    ;prototypesystemlo1.whichrealizesvideoretrieval ;basedonobjectoreventconceptsbymanual

    ;annotatingbeforehand.Inaddition,thesystem ;providesatrainingandlearningmechanismbasedon ;HMMforautomaticannotation.Wecontrastthe ;systemwiththispaper‟smethodinrecognitionrange

    ;andaccuracybyexperimentwhichwillbeintroduced ;nthelastsection.

    ;Jurgen,etalll1recognizedgoal

    ;,football,position

    ;andcolorofathleteinfootballgame.Theinformation ;willbecombinedwithspacerelationshipamongthem ;torealizeautomaticmarkingofwonderfulscene. ;2

    ;Visualandaudiofcaturesareutilizedatthesame ;timetorccognizevidcoeventsemanticinreference ;121.Thefirstoneisinclinedtodetectshoot,fouland ;otherevcnts.Anotheroneexpandstherecognizing ;scopetonewsandadvertisementdomain.

    ;3.Objectsemantictemplate

    ;3.1Semanticconceptsontologydatabase

    ;Thispaperconstructsthreesemanticconcept ;ontology:object,scene,andevent,todescribe ;semanticinformationimpliedinvideo.Object ;semanticistakenasthefoundationofothersemantic ;types.Theyareinclinedtocollectcommonconcepts ;emergingindifferentdomainsandaredefinedas ;hierarchicstructureintermsofabstractdegree. ;3.2Semantictemplatetrainingsystem

    ;Semantictemplateislookedasthebridgebetween ;featurelayerandsemanticlayer.Eachtemplatemapsto ;severalsemanticconcepts.Thispaperconstructs ;differentsemantictemplaterespectivelycorresponding ;toobjectandeventsemantic.0bjectSemanticTemplate ;(OST)containsnotonlystaticfeaturerelationship ;obtainedbystatisticlearning,butalsobestsuitable ;segmentgranularityforeachobjectregion,shapesets ;description,objectsub-regionfeatures,space ;relationshipdescriptionandsoon.Thesearecontributed ;torelatetime-basefeaturestoobjectsemantic.In ;addition,bestsuitablefeaturewillalsobestoredto ;realizeretrievingwithmultigranularities.Event

    ;semantictemplateisbasedonOST.Itstorestimespace

    ;topologyrelationamongregions,statetransferring ;throughtime,objectsappearinglistandsoon. ;ThekeyproblemofconstructingOSTishowto ;correspondobjectregioninvideowithobjectconcept ;inreal-world.WetrainvisualmediasetbySemantic ;TemplateTraining(STT)system,showninFig.1, ;whichisalsoadaptedtovideoretrievalbasedon ;example.

    ;

    ;EventsemanticrecognizingbasedonMarkovcha

    ;in

    ;

    ;……:t0ee

    ;r(0,f1

    ;0IS

    ;_?m,:

    ;m

    ;E…RTOF.__

    ;F

    ;Sl_F121

    ;h

    ;Plt*?

;I‟

    ;$.

    ;

    ;

    ;.

     ;j舞簧‟.

    ;圈霸

    ;

    ;Fig.1Semantictemplatetrainingsystem

    ;3.3Videopre-processingandinteractive

    ;conceptmatching

    ;Videopreprocessingincludesvideofeatures

    ;extracting,videosegmenting,videoframesegmenting ;basedonobjectandotherprocessing.Wetakethe ;begin..frameofvideosegmentasoriginalinformation. ;STTsystemwillbeutilizedtoobtainthemapping ;relationship,betweenobjectregioninformationand ;objectsemanticconcept,whichisthefoundationof ;constructingOST.

    ;Afterascertainingtheobjectregions,wewill ;dynamicallyadjustthesegmentinggranularityofthis ;regionandverifyitstrackingeffectivenessin ;followupframesforeachsemanticobject.

    ;3.4ConstructingOST

    ;Thispaperimprovesparametersaccuracyin ;semantictemplatebyceaselesslyiterativeprocessing ;basedonstaffsticalprobabilityuntilallparameters ;leveloff.Theaimisthateachobjectregionsegmented ;couldautomaticallybemappedtothebestsuitable ;semanticconcept.Moreover,wetakeintoaccountthe ;relationshipamongdifferentobjectsemanticconcepts ;whichwillimprovetheeffectofmappingandenhance ;thevalidityofvideosemanticexpression. ;Wedefine=,,?,(i?t1,N)to

    ;denotevideofeaturevector,whereNisthenumberof ;videosegmentsintrainingsetandXisfeature ;dimension.cI(j?【1,Q])denotesaoNectsemantic

    ;concept,Qistheamountofobjectsemanticconcepts. ;OSTisrepresentedaso?o={Ol,02,…oK),whereKis

    ;theamountofobjectsemantictemplatewaitingfor ;beingconstructed.Eachfeaturefisabletobelongto ;severalOST,buteachOSTcorrespondstoasingle ;concept.Theessencethatlowlevelfeaturetransitsto

    ;objectsemanticistocomputetheprobabilityvalue

;P(clf),whereOSTplaysthebridgerole.

;(,cj)constructedbyfeaturevectorand

    ;semanticconceptissupposedtobeindependenteach ;otherthatcouldberepresentedasfollows: ;P(L,clo)=(Io)(cIo)(1)

     ;Featurevectorandconceptarelookedasrandom;distributinginvideoframe.Theprobability,thataOST ;ischosen.isrepresentedasP(ok).Informula(1), ;(Iot)and(cIot)respectivelydenote

    ;conditionalprobabilityoffeatureandconceptc1.If ;theeffectofOSTokisabletobeignored,wewill ;computetheprobabilityof(fi,c):

    ;,)=:()Id)

    ;=

    ;P(cj)?,(Io)Jp(Ic)(2)

    ;P(oIc)couldbetransformedbyBayesian ;formu1a:

    ;P(,c)P()(Io,)(cIok)

    ;Supposethatfeaturevectorsinwholemedia ;documentaccordwithKGaussianmixturedistribution. ;SoeachOSTOkwillcorrespondtoaGaussian ;distribution.Theconditionalprobabilityofregion ;featurefiinOkis:

    ;(lOk)=

    ;(2X)x/”P()?((4)

    ;“Where?trespectivelydenotethemeanand

    ;covariancematrixoffjintheobjectregion ;correspondingtook.

    ;=

    ;4Lp(l)/?N4p(IL)(5)

    ;3

    ;

    ;EventsemanticrecognizingbasedonMarkovchain ;?=

    ;?.p(oI)()()

    ;?5(.I(6)

    ;Whereistheamountofconceptsinvideo

    ;flame.Intermofmaximumlikelihoodprinciple, ;functionL(O)andIn(L(0))willtendtomaximumwith ;thesameindependentvariable.Therefore,thispaper ;takesPr(Io?)asindependentvariabletoform

    ;theprobabilityformulaforannotatingthewhole ;trainingset.

    ;Inn.n(cIO)P(cIc”)

;=

    ;?.?.In(?.P()(clot))

    ;+In

    ;-t.

    ;P(clcs)1

    ;P(cIc)denotesrelationshipbetweenobject ;semanticconcepts.Forreducingthecomplexity,we ;dontconsidertherelationshipamongcomposite ;concepts,theadjacencyrelationamongtheseregions ;correspondingtodifl~rentconceptsandsoon.By ;similarprinciple,(Io)couldbecalculated: ;ln.P(Io)

    ;P(Ol,)=n,n,P(.j,l)

    ;Wheres(ci)denotestheprobabilitythatthe ;conceptcorrespondingtofeaturevectorfiemergesin ;media,oi.

    ;idenotestheOSTcorrespondingto(,c

    ;Withthesimilarmethod,wegaintheexpectationof ;likelihoodestimateInbasedonP(OI):

    ;??,1nP(.)(Io,)]P(.I)(12)

    ;Formula(13)providestwoaccessionaIconditions. ;WeutilizeLagrangemultipliermethodwithpartial ;derivativeP(Ok),(Io),(cIo?)to

    ;calculatetherelevantvariableswhenformulas(11)and ;(12)emergemaximum.

    ;P(o)=1

    ;lP(oI,c)=1

    ;ThecalculatedP(oI)andformula(4)will ;providetheinitialdistributionoffeaturesinOST. ;AndP(oI,c)providestheprobabilitythatOST ;ischosen.

    ;:

    ;?,In?P(ok)PAf,Iok)](8)P(.)=

    ;Theaboveformulasgaintheprobabilitythat ;featureCandconceptarechosen,whileOkiStakenas ;priorprobability.However,itisnecessarytogainthe ;posteriorprobabilityofOkfortransitinglowlevel

    ;featuretoobjectsemantic.Wededucethefollowing ;formulasbasedonBayesianrule:

    ;P(ol)=P(o)(lo)

    ;?KP(.)(

    ;P(cj)P(ok)Pa(fiIo,)~(cjIo,)

    ;(1.)

    ;Theexpectationoffulllikelihoodestimateof

;ln,iscalculatedbasedonP(Ol,):

    ;...

    ;()1nIp(,

    ;,)(l,,)(lq,)|p(Dl)

    ;4

    ;?Z,Ls(c)P(.I

    ;??(c)

    ;3.5FeedbackstudyofOST

    ;Itisunavoidablethattrainingdatacontainnoise. ;AtieranOSThasbeenconstructed,itmustbevalidated ;byretrievinginlargermediasettoupdateitself.Since ;OSTrealizesthemappingbetweenobjectregionand ;semanticconcept,wecoulddirectlyretrievalwith ;conceptword.Theformulaisasfollows:

    ;fiIcj=flIo)P(oIc~)do

    ;=

    ;f))

    ;I.)(15)

    ;Theresultofformula(15)istheexpectationwith ;independentvariableP(0).Thispaperusesthe ;

    ;EventsemanticrecognizingbasedonMarkovchain ;cxpectationestimatingmethodinIMCC96ltocompute ;thcapproximationoftbrmula(15):

    ;Intermsofformula(I6),systemwillreturnM ;documentsofmaximalP(lc)invideodatabase. ;Thenusermarksthosesatisfyingresults.Wedesigna ;decisionfunction((),()=whichdenotesthat

    ;belongingtothetemplateOkisaccepted. ;SupposetheprobabilitythatgainstemplateOkis ;P(okIq)basedontheinputci.Ifthedecision(jleadsto

    ;anexpense(lo),thetotalriskof(iwillbe:

    ;R(P,():?P(oIc)(()(17)=I

    ;Where

    ;R(giok)=2(eok)P(41.)

    ;Thekeythatjudgesthesuitabilityofthe

    ;constructedOSTforobjectsemanticrecognitionisto ;maketheriskminimalorlessthanacertainthreshold. ;Otherwise,itisnecessarytoupdateOSTbyincreasing ;trainingexamples.

    ;Theweightisassignedtoeachtypeoffeaturein ;termsoftheirdifferentsuitabilityfordifferentobject ;semantic.Souserscouldcustomthevectordimension ;toadjusttheproportionbetweenretrievalspeedand

    ;accuracy.Thisprovidesastrategyofretrievingthe ;videowithoutannotationinrealtime.

    ;4.Eventsemantictemplate

    ;Thispapertakestheeventsemanticastheresultof ;semanticobjectsevolvingwithsomecertainrules, ;structure,ormotionrelation.Soobjectsemanticisthe ;kernelofconstructingEventSemanticTemplate(EST), ;theconstructingprocessofwhichissimilartoOST. ;Wecouldtransitlow-levelfeaturetohighlevel

    ;semanticbystages.Forexample,severalsemantic ;objects,theirunstabletimespacetopologyrelation,

    ;objectstemporalsequencerelation,transferringof ;obiectmotionstateandsoon,couldrepresentcomplex ;eventsemantic.

    ;Eventsemanticisinclinedtodescribecertain ;interestedobjectregions,butignoreotherregions, ;namelyobservingdifferentregionsindifferentdegrees. ;Besides,eventsemanticisalwaysrelativetotimeand ;hasthetimedurative.

    ;ForexpedientlyconstructingEST,wedivide ;eventsemanticintounitaryandmultiunitevent ;semantic.Thefirstispossibletorefertooneortwo ;interestedobjects,butexistonlyonemainlyobserved ;semanticobject.Anotheriscombinedwithlotsof ;unitaryeventswithcertainrelationships.Theinterested ;objectofeachunitaryeventcontainedinmultiunit ;eventislikelytobedifferent.

    ;(1)Recognizingunitaryeventsemantic

    ;Unitaryeventsemanticislookedastheminimal ;unitdescribingeventsemantic.Ifavideosegment ;containsoneinterestedobject,theobjectwilldirectly ;betakenasobservedobject.Onthecontrary,ifit ;containsseveralinterestedobjectswithdifferent ;motionstate,itisnecessarytoclassifythem: ;a.Thoseinterestedobjectswithsimilarmotion ;statewillbeassignedtothesameclass;

    ;b.Constructtheminimalconvexpolygonof

    ;interestedobjectclass,andtakeitasthedeputyofthe ;class;

    ;c.Selecttworandomclassestostructurea

    ;descriptionunitofunitaryevent.Regardoneofthemas ;themainobservationclassandanotherasreference ;class.Inotherwords,nclassesarepossibletobe ;dividedinton(n1)/2unitaryeventsatmost.

;Unitaryeventisactuallyusedtodescribe

    ;semanticobjects(classes),theirsownmotionstatesand ;possiblyexistingactionrelationbetweentwoobjects ;(classes).Whenaclassisdescribedbyanobject ;semanticconcept,ifthereisanothersemanticobject ;withdifferentconceptintheclass,wesubstituteto ;describetheclassbytheclosestparentconcept.For ;example,therearetwodifferentconcepts”tiger‟‟and

    ;“horse”inaclass.theirparentconcept”animal‟‟willbe

    ;5

    ;

    ;EventsemanticrecognizingbasedonMarkovchain ;usedtodescribetheclass.Whenthelnterested?

Report this document

For any questions or suggestions please email
cust-service@docsford.com