资源描述:
《Distinctive Image Features from Scale-Invariant Keypoints》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、InternationalJournalofComputerVision60(2),91–110,2004c2004KluwerAcademicPublishers.ManufacturedinTheNetherlands.DistinctiveImageFeaturesfromScale-InvariantKeypointsDAVIDG.LOWEComputerScienceDepartment,UniversityofBritishColumbia,Vancouver,B.C.,CanadaLowe@cs.ubc.caReceivedJanuary10
2、,2003;RevisedJanuary7,2004;AcceptedJanuary22,2004Abstract.Thispaperpresentsamethodforextractingdistinctiveinvariantfeaturesfromimagesthatcanbeusedtoperformreliablematchingbetweendifferentviewsofanobjectorscene.Thefeaturesareinvarianttoimagescaleandrotation,andareshowntoproviderobus
3、tmatchingacrossasubstantialrangeofaffinedistortion,changein3Dviewpoint,additionofnoise,andchangeinillumination.Thefeaturesarehighlydistinctive,inthesensethatasinglefeaturecanbecorrectlymatchedwithhighprobabilityagainstalargedatabaseoffeaturesfrommanyimages.Thispaperalsodescribesanap
4、proachtousingthesefeaturesforobjectrecognition.Therecognitionproceedsbymatchingindividualfeaturestoadatabaseoffeaturesfromknownobjectsusingafastnearest-neighboralgorithm,followedbyaHoughtransformtoidentifyclustersbelongingtoasingleobject,andfinallyperformingverificationthroughleast-s
5、quaressolutionforconsistentposeparameters.Thisapproachtorecognitioncanrobustlyidentifyobjectsamongclutterandocclusionwhileachievingnearreal-timeperformance.Keywords:invariantfeatures,objectrecognition,scaleinvariance,imagematching1.Introductionexpensiveoperationsareappliedonlyatloc
6、ationsthatpassaninitialtest.FollowingarethemajorstagesofImagematchingisafundamentalaspectofmanyprob-computationusedtogeneratethesetofimagefeatures:lemsincomputervision,includingobjectorscenerecognition,solvingfor3Dstructurefrommultipleim-1.Scale-spaceextremadetection:Thefirststageof
7、ages,stereocorrespondence,andmotiontracking.Thiscomputationsearchesoverallscalesandimagelo-paperdescribesimagefeaturesthathavemanyprop-cations.Itisimplementedefficientlybyusingaertiesthatmakethemsuitableformatchingdifferingdifference-of-Gaussianfunctiontoidentifypoten-imagesofanobje
8、ctorscene.Thefeaturesarein