Beyond eigenfaces: probabilistic matching for face recognition

Abstract

We propose a novel technique for direct visualmatching of images for the purposes of facerecognition and database search. Specifically,we argue in favor of a probabilistic measure ofsimilarity, in contrast to simpler methods whichare based on standard L2 norms (e.g., templatematching) or subspace-restricted norms (e.g.,eigenspace matching). The proposed similaritymeasure is based on a Bayesian analysis of imagedifferences: we model two mutually exclusiveclasses of variation between...