Probabilistic Modeling for Face Recognition
Biometrics is a rapidly maturing industry in which face recognition will play a significant role. Verification of one's identity based on facial appearance has been a recurring research problem in computer vision for the last 20 years. To assess the state-of-the-art in face recognition technology, Defense Advanced Research Project's Agency (DARPA) recently set up the "FERET" program in which multiple academic and commercial groups were encouraged to participate. The goal was to provide a common objective testbed to compare different recognition strategies and algorithms. The system developed in this project was the winner of this competition in the last round of tests in September 1996. It not only advanced the envelope of performance previously established by other techniques but also introduced a new computational mechanism for face recognition which uses a probabilistic formulation in terms of dual "eigenfaces."
Background & Objective: Past techniques for face recognition can be categorized as either feature-based (geometric) or template-based (photometric), of which the latter have proved the more successful. Template-based methods use measures of facial similarity based on standard Euclidean error norms (e.g., template matching) or subspace-restricted error norms (e.g., eigenspace matching). The latter technique of "eigenfaces" has in the past decade become the "golden standard" to which other algorithms are often compared. The goal of this research was to improve on this standard benchmark while formulating a novel probabilistic similarity function for recognition.
Technical Discussion: We use a novel technique for direct visual matching of images for the purposes of face recognition and image database search. Specifically, we argue in favor of a probabilistic measure of similarity, based on a Bayesian analysis of image differences: we model two mutually exclusive classes of variation between two facial images: intra-personal (variations in appearance of the same individual, due to different expressions or lighting, for example) and extra-personal (variations in appearance due to different identity). The high-dimensional probability density functions for each respective class are then obtained from training data and used to compute a similarity measure based on the a posteriori probability of membership in the intra-personal class, which is used to find best matches in the database. The performance advantage of this probabilistic matching technique over standard nearest-neighbor eigenspace matching is demonstrated using results from DARPA's 1996 "FERET" face recognition competition, in which this system was found to be the top performer.
This probabilistic framework is particularly advantageous in that the intra/extra density estimates explicitly characterize the type of appearance variations which are critical in formulating a meaningful measure of similarity. For example, the differences corresponding to facial expression changes (which may have high error norms) are, in fact, irrelevant when the measure of similarity is to be based on identity. The subspace density estimation method used for representing these classes thus corresponds to a learning method for discovering the principal modes of variation important to the classification task. Furthermore, by equating similarity with the a posteriori probability we obtain an optimal non-linear decision rule for matching and recognition. This aspect of our approach differs significantly from recent methods which use linear discriminant analysis techniques for recognition.
| Technical Reports: | |
| Face Recognition in Subspaces | |
| Principal Manifolds and Probabilistic Subspaces for Visual Recognition | |
| A Bayesian Similarity Measure for Deformable Image Matching | |
| Bayesian Face Recognition with Deformable Image Models | |
| Bayesian Face Recognition | |
| Principal Manifolds and Bayesian Subspaces for Visual Recognition | |
| Beyond Euclidean Eigenspaces: Bayesian Matching for Visual Recognition | |
Technology Area: Computer Vision
Modification Date: September 12, 2007