Bayesian image retrieval in biometric databases
B. Moghaddam, Alex Pentland · 2003
We demonstrate a Bayesian technique for direct visual matching of biometric images for the purposes of face recognition and retrieval, using a probabilistic measure of facial similarity. This technique differs significantly from simpler methods which are based on standard Euclidean norms on image feature vectors (template matching) or subspace-restricted norms (eigenspace matching). Our similarity measure is based primarily on a statistical analysis of the observed inter-image differences in the database. The performance advantage of this probabilistic matching technique over a standard Euclidean (nearest-neighbor) eigenspace technique has been demonstrated in DARPA's "FERET" face recognition competition, in which our probabilistic retrieval algorithm was found to be a top performer.