简介：Inthispaper,wepresentasimplebutpowerfulensembleforrobusttextureclassification.Theproposedmethodusesasingletypeoffeaturedescriptor,i.e.scale-invariantfeaturetransform(SIFT),andinheritsthespiritofthespatialpyramidmatchingmodel(SPM).Inaflexiblewayofpartitioningtheoriginaltextureimages,ourapproachcanproducesufficientinformativelocalfeaturesandtherebyformareliablefeaturepondortrainanewclass-specificdictionary.Totakefulladvantageofthisfeaturepond,wedevelopagroup-collaborativelyrepresentation-basedstrategy(GCRS)forthefinalclassification.Itissolvedbythewell-knowngrouplasso.Butwegobeyondofthisandproposealocality-constraintmethodtospeedupthis,namedlocalconstraint-GCRS(LC-GCRS).Experimentalresultsonthreepublictexturedatasetsdemonstratetheproposedapproachachievescompetitiveoutcomesandevenoutperformsthestate-of-the-artmethods.Particularly,mostofmethodscannotworkwellwhenonlyafewsamplesofeachcategoryareavailablefortraining,butourapproachstillachievesveryhighclassificationaccuracy,e.g.anaverageaccuracyof92.1%fortheBrodatzdatasetwhenonlyoneimageisusedfortraining,significantlyhigherthananyothermethods.