A Bayesian similarity measure for deformable image matching

We propose a probabilistic similarity measure for direct image matching bas ed on a Bayesian analysis of image deformations. We model two classes of va riation in object appearance: intra-object and extra-object. The probabilit y density functions for each class are then estimated from training data an d used to compute a similarity measure based on the a posteriori probabilit ies. Furthermore, we use a novel representation for characterizing image di fferences using a deformable technique for obtaining pixel-wise corresponde nces. This representation, which is based on a deformable 3D mesh in XYI-sp ace, is then experimentally compared with two simpler representations: inte nsity differences and optical Row. The performance advantage of our deforma ble matching technique is demonstrated using a typically hard test set draw n from the US Army's FERET face database. (C) 2001 Elsevier Science B.V. Al l rights reserved.