Combined SubspaceMethod sing lobaland ocalFeatures forFaceRecognition
Chunghoon Kim · 2005
Thispaperproposes acombined subspace method using bothglobal andlocal features forfacerecognition. The global andlocal features areobtained byapplying theLDA- basedmethodtoeither thewholeorpartofa faceimage, respectively. Thecombined subspace isconstructed withthe proection vectors corresponding tolarge eigenvalues ofthe between-class scatter matrix ineachsubspace. Itisbased onthe factthattheeigenvectors corresponding tolarger eigenvalues havemorediscriminating power.Thecombined subspace is evaluated inviewoftheBayeserror, whichshowshowwell samples canbeclassified. Thecombined subspace gives small Bayes error thanthesubspaces composed ofeither theglobal or local features. Comparative experiments arealso performed using theColor FERETdatabase offacial images. Theexperimental results showthatthecombined subspace methodgives better recognition ratethanother subspace methods.