资源描述:
《sift算法 图像检索 特征提取 必备(david glowe)》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、DistinctiveImageFeaturesfromScale-InvariantKeypointsDavidG.LoweComputerScienceDepartmentUniversityofBritishColumbiaVancouver,B.C.,Canadalowe@cs.ubc.caJanuary5,2004AbstractThispaperpresentsamethodforextractingdistinctiveinvariantfeaturesfromimagesthatcanbeusedtoperformreliab
2、lematchingbetweendifferentviewsofanobjectorscene.Thefeaturesareinvarianttoimagescaleandrotation,andareshowntoproviderobustmatchingacrossaasubstantialrangeofaffinedis-tortion,changein3Dviewpoint,additionofnoise,andchangeinillumination.Thefeaturesarehighlydistinctive,inthesens
3、ethatasinglefeaturecanbecor-rectlymatchedwithhighprobabilityagainstalargedatabaseoffeaturesfrommanyimages.Thispaperalsodescribesanapproachtousingthesefeaturesforobjectrecognition.Therecognitionproceedsbymatchingindividualfea-turestoadatabaseoffeaturesfromknownobjectsusingaf
4、astnearest-neighboralgorithm,followedbyaHoughtransformtoidentifyclustersbelongingtoasin-gleobject,andfinallyperformingverificationthroughleast-squaressolutionforconsistentposeparameters.Thisapproachtorecognitioncanrobustlyidentifyobjectsamongclutterandocclusionwhileachievingn
5、earreal-timeperformance.AcceptedforpublicationintheInternationalJournalofComputerVision,2004.11IntroductionImagematchingisafundamentalaspectofmanyproblemsincomputervision,includingobjectorscenerecognition,solvingfor3Dstructurefrommultipleimages,stereocorrespon-dence,andmoti
6、ontracking.Thispaperdescribesimagefeaturesthathavemanypropertiesthatmakethemsuitableformatchingdifferingimagesofanobjectorscene.Thefeaturesareinvarianttoimagescalingandrotation,andpartiallyinvarianttochangeinilluminationand3Dcameraviewpoint.Theyarewelllocalizedinboththespat
7、ialandfrequencydomains,re-ducingtheprobabilityofdisruptionbyocclusion,clutter,ornoise.Largenumbersoffeaturescanbeextractedfromtypicalimageswithefficientalgorithms.Inaddition,thefeaturesarehighlydistinctive,whichallowsasinglefeaturetobecorrectlymatchedwithhighprobabilityagain
8、stalargedatabaseoffeatures,providingabasisforobjectandscenerecognition.Thecostofex