特征空間穩(wěn)健性分析彩色圖像分割畢業(yè)論文外文翻譯

1、（1）外文文獻(xiàn)原文RobustAnalysisofFeatureSpaces:ColImageSegmentationAbstractAgeneraltechniqueftherecoveryofsignificantimagefeaturesispresented.Thetechniqueisbasedonthemeanshiftalgithmasimplenonparametricprocedurefestimatingdensit

2、ygradients.Drawbacksofthecurrentmethods(includingrobustclustering)areavoided.Featurespaceofanynaturecanbeprocessedasanexamplecolimagesegmentationisdiscussed.Thesegmentationiscompletelyautonomousonlyitsclassischosenbytheu

3、ser.Thusthesameprogramcanproduceahighqualityedgeimageprovidebyextractingallthesignificantcolsapreprocessfcontentbasedquerysystems.A512512colimageisanalyzedin?lessthan10secondsonastardwkstation.Graylevelimagesarehledascol

4、imageshavingonlythelightnesscodinate.Keywds:robustpatternanalysislowlevelvisioncontentbasedindexing1IntroductionFeaturespaceanalysisisawidelyusedtoolfsolvinglowlevelimageunderstingtasks.Givenanimagefeaturevectsareextract

5、edfromlocalneighbhoodsmappedintothespacespannedbytheircomponents.Significantfeaturesintheimagethencrespondtohighdensityregionsinthisspace.Featurespaceanalysisistheprocedureofrecoveringthecentersofthehighdensityregionsi.e

6、.therepresentationsofthesignificantimagefeatures.HistogrambasedtechniquesHoughtransfmareexamplesoftheapproach.Whenthenumberofdistinctfeaturevectsislargethesizeofthefeaturespaceisreducedbygroupingnearbyvectsintoasinglecel

7、l.Adiscretizedfeaturespaceiscalledanaccumulat.Wheneverthesizeoftheaccumulatcellisnotadequatefthedataseriousartifactscanappear.TheproblemwasextensivelystudiedinthecontextoftheHoughtransfme.g..Thusfsatisfactyresultsafeatur

8、espaceshouldhavecontinuouscodinatesystem.Thecontentofacontinuousfeaturespacecanbemodeledasasamplefromamultivariatemultimodalprobabilitydistribution.Notethatfrealimagesthenumberofmodescanbeverylargeofthederoftens.searchin

9、gftheminimalvolumeellipsoidcontainingatleasthdatapoints.Themultivariatelocationestimateisthecenterofthisellipsoid.Toavoidcombinatialexplosionaprobabilisticsearchisemployed.Letthedimensionofthedatabep.Asmallnumberof(p1)tu

10、pleofpointsareromlychosen.Feach(p1)tuplethemeanvectcovariancematrixarecomputeddefininganellipsoid.TheellipsoidisinatedtoincludehpointstheonehavingtheminimumvolumeprovidestheMVEestimate.BasedonMVEarobustclusteringtechniqu

11、ewithapplicationsincomputervisionwasproposedin.Thedataisanalyzedunderseveralresolutions“byapplyingtheMVEestimatrepeatedlywithhvaluesrepresentingfixedpercentagesofthedatapoints.Thebestclusterthencrespondstothehvalueyieldi

12、ngthehighestdensityinsidetheminimumvolumeellipsoid.Theclusterisremovedfromthefeaturespacethewholeprocedureisrepeatedtillthespaceisnotempty.TherobustnessofMVEshouldensurethateachclusterisassociatedwithonlyonemodeoftheunde

13、rlyingdistribution.Thenumberofsignificantclustersisnotneededaprii.Therobustclusteringmethodwassuccessfullyemployedftheanalysisofalargevarietyoffeaturespacesbutwasfoundtobecomelessreliableoncethenumberofmodesexceededten.T

14、hisismainlyduetothenmalityassumptionembeddedintothemethod.Theellipsoiddefiningaclustercanbealsoviewedasthehighconfidenceregionofamultivariatenmaldistribution.ArbitraryfeaturespacesarenotmixturesofGaussiansconstrainingthe

15、shapeoftheremovedclusterstobeellipticalcanintroduceseriousartifacts.Theeffectoftheseartifactspropagatesasmemeclustersareremoved.Furthermetheestimatedcovariancematricesarenotreliablesincearebasedononlyp1points.Subsequentp

16、ostprocessingbasedonallthepointsdeclaredinlierscannotfullycompensatefaninitialerr.Tobeabletocrectlyrecoveralargenumberofsignificantfeaturestheproblemoffeaturespaceanalysismustbesolvedincontext.Inimageunderstingtaskstheda

17、tatobeanalyzediginatesintheimagedomain.Thatisthefeaturevectssatisfyadditionalspatialconstraints.Whiletheseconstraintsareindeedusedinthecurrenttechniquestheirroleismostlylimitedtocompensatingffeatureallocationerrsmadeduri