A model of illumination variation for robust face recognition

We recently introduced a novel approach to face recognition which consists in modeling the set of possible transformations between face images of the same person. While our previous work focused on geometric transformations to model facial expressions, in this article we consider feature transformations as a means to compensate for illumination variations. Although this approach requires to learn the set of possible illumination transformations through a training phase, we will show experimentally that the trained parameters are very robust. Even in the challenging case where the databases used to train the transformation model and to assess the performance of the system are very different, the proposed approach results in large improvements of the recognition rate.