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,更多相关内容在工程资料-天天文库。
1、附录2:外文翻译RobustAnalysisofFeatureSpaces:ColorImageSegmentationAbstractAgeneraltechniquefortherecoveryofsignificantimagefeaturesispresented.Thetechniqueisbasedonthemeanshiftalgorithm,asimplenonparametricprocedureforestimatingdensitygradients.Drawbacksofthecurrentmethods(includi
2、ngrobustclustering)areavoided.Featurespaceofanynaturecanbeprocessed,andasanexample,colorimagesegmentationisdiscussed.Thesegmentationiscompletelyautonomous,onlyitsclassischosenbytheuser.Thus,thesameprogramcanproduceahighqualityedgeimage,orprovide,byextractingallthesignificant
3、colors,apreprocessorforcontent-basedquerysystems.A512512colorimageisanalyzedinlessthan10secondsonastandardworkstation.Graylevelimagesarehandledascolorimageshavingonlythelightnesscoordinate.Keywords:robustpatternanalysis,low-levelvision,content-basedindexing1IntroductionFeatu
4、respaceanalysisisawidelyusedtoolforsolvinglow-levelimageunderstandingtasks.Givenanimage,featurevectorsareextractedfromlocalneighborhoodsandmappedintothespacespannedbytheircomponents.Significantfeaturesintheimagethencorrespondtohighdensityregionsinthisspace.Featurespaceanalys
5、isistheprocedureofrecoveringthecentersofthehighdensityregions,i.e.,therepresentationsofthesignificantimagefeatures.Histogrambasedtechniques,Houghtransformareexamplesoftheapproach.Whenthenumberofdistinctfeaturevectorsislarge,thesizeofthefeaturespaceisreducedbygroupingnearbyve
6、ctorsintoasinglecell.Adiscretizedfeaturespaceiscalledanaccumulator.Wheneverthesizeoftheaccumulatorcellisnotadequateforthedata,seriousartifactscanappear.TheproblemwasextensivelystudiedinthecontextoftheHoughtransform,e.g..Thus,forsatisfactoryresultsafeaturespaceshouldhaveconti
7、nuouscoordinatesystem.Thecontentofacontinuousfeaturespacecanbemodeledasasamplefromamultivariate,multimodalprobabilitydistribution.Notethatforrealimagesthenumberofmodescanbeverylarge,oftheorderoftens.Thehighestdensityregionscorrespondtoclusterscenteredonthemodesoftheunderlyin
8、gprobabilitydistribution.Traditionalclusteringtechniques,canbeusedforfeatur