Robust score normalization for relational approaches to face authentication

Relational approaches to pattern recognition consist in modeling the relationship between observations. In this paper, we consider two score normalization strategies based on a Bayesian framework for relational approaches to face authentication. The first one is specific to relational approaches and models the relationship between face images of different persons. The second one, which is very general and can be applied to any face authentication system, models directly impostors. These two techniques are compared from a theoretical and an experimental point of view and both comparisons hint at a superiority of the general approach.