资源描述：
《外文翻译---特征空间稳健性分析：彩色图像分割》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、附录2：外文翻译RobustAnalysisofFeatureSpaces:ColorImageSegmentationAbstractAgeneraltechniquefortherecoveryofsignificantimagefeaturesispresented.Thetechniqueisbasedonthemeanshiftalgorithm,asimplenonparametricprocedureforestimatingdensitygradients.Drawbacksofthecurrentmethods(includingrob
2、ustclustering)areavoided.Featurespaceofanynaturecanbeprocessed,andasanexample,colorimagesegmentationisdiscussed.Thesegmentationiscompletelyautonomous,onlyitsclassischosenbytheuser.Thus,thesameprogramcanproduceahighqualityedgeimage,orprovide,byextractingallthesignificantcolors,apr
3、eprocessorforcontent-basedquerysystems.Acolorimageisanalyzedinlessthan10secondsonastandardworkstation.Graylevelimagesarehandledascolorimageshavingonlythelightnesscoordinate.Keywords:robustpatternanalysis,low-levelvision,content-basedindexing1IntroductionFeaturespaceanalysisisawid
4、elyusedtoolforsolvinglow-levelimageunderstandingtasks.Givenanimage,featurevectorsareextractedfromlocalneighborhoodsandmappedintothespacespannedbytheircomponents.Significantfeaturesintheimagethencorrespondtohighdensityregionsinthisspace.Featurespaceanalysisistheprocedureofrecoveri
5、ngthecentersofthehighdensityregions,i.e.,therepresentationsofthesignificantimagefeatures.Histogrambasedtechniques,Houghtransformareexamplesoftheapproach.Whenthenumberofdistinctfeaturevectorsislarge,thesizeofthefeaturespaceisreducedbygroupingnearbyvectorsintoasinglecell.Adiscretiz
6、edfeaturespaceiscalledanaccumulator.Wheneverthesizeoftheaccumulatorcellisnotadequateforthedata,seriousartifactscanappear.TheproblemwasextensivelystudiedinthecontextoftheHoughtransform,e.g..Thus,forsatisfactoryresultsafeaturespaceshouldhavecontinuouscoordinatesystem.Thecontentofac
7、ontinuousfeaturespacecanbemodeledasasamplefromamultivariate,multimodalprobabilitydistribution.Notethatforrealimagesthenumberofmodescanbeverylarge,oftheorderoftens.Thehighestdensityregionscorrespondtoclusterscenteredonthemodesoftheunderlyingprobabilitydistribution.Traditionalclust
8、eringtechniques,canbeusedforfeaturespace
