Analgorithmforfacedescriptionandrecognitionbasedonmulti-resolutionwithmulti-scalelocalbinarypattern(multi-LBP)featuresisproposed.Thefacialimagepyramidisconstructedandeachfacialimageispidedintovariousregionsfromwhichpartialandholisticlocalbinarypatter(LBP)histogramsareextracted.AllLBPfeaturesofeachimageareconcatenatedtoasingleLBPeigenvectorwithdifferentresolutions.ThedimensionalityofLBPfeaturesisthenreducedbyalocalmarginalignment(LMA)algorithmbasedonmanifold,whichcanpreservethebetween-classvariance.Supportvectormachine(SVM)isappliedtoclassifyfacialimages.ExtensiveexperimentsonORLandCMUfacedatabasesclearlyshowthesuperiorityoftheproposedschemeoversomeexistedalgorithms,especiallyontherobustnessofthemethodagainstdifferentfacialexpressionsandposturesofthesubjects.
