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(
2、includingrobustclustering)areavoided.Featurespaceofanynaturecanbeprocessed,andasanexample,colorimagesegmentationisdiscussed.Thesegmentationiscompletelyautonomous,onlyitsclassischosenbytheuser.Thus,thesameprogramcanproduceahighqualityedgeimage,orprovide,byextractingall
3、thesignificantcolors,apreprocessorforcontent-basedquerysystems.A512512colorimageisanalyzedinlessthan10secondsonastandardworkstation.Graylevelimagesarehandledascolorimageshavingonlythelightnesscoordinate.Keywords:robustpatternanalysis,low-levelvision,content-basedindex
4、ing1IntroductionFeaturespaceanalysisisawidelyusedtoolforsolvinglow-levelimageunderstandingtasks.Givenanimage,featurevectorsareextractedfromlocalneighborhoodsandmappedintothespacespannedbytheircomponents.Significantfeaturesintheimagethencorrespondtohighdensityregionsin
5、thisspace.Featurespaceanalysisistheprocedureofrecoveringthecentersofthehighdensityregions,i.e.,therepresentationsofthesignificantimagefeatures.Histogrambasedtechniques,Houghtransformareexamplesoftheapproach.Whenthenumberofdistinctfeaturevectorsislarge,thesizeofthefeat
6、urespaceisreducedbygroupingnearbyvectorsintoasinglecell.Adiscretizedfeaturespaceiscalledanaccumulator.Wheneverthesizeoftheaccumulatorcellisnotadequateforthedata,seriousartifactscanappear.TheproblemwasextensivelystudiedinthecontextoftheHoughtransform,e.g..Thus,forsatis
7、factoryresultsafeaturespaceshouldhavecontinuouscoordinatesystem.Thecontentofacontinuousfeaturespacecanbemodeledasasamplefromamultivariate,multimodalprobabilitydistribution.Notethatforrealimagesthenumberofmodescanbeverylarge,oftheorderoftens.Thehighestdensityregionscor
8、respondtoclusterscenteredonthemodesoftheunderlyingprobabilitydistribution.Traditionalclusteringtechniques,canbeusedforfeatur
