Face learning using a sequence of images

We present a face matching algorithm used in the face learning stage. For this, several images of the same person are acquired under head movement constraints, the user rotating his head from left to right. After a precise eyes location, the normalization of the face under rotation and scale is achieved, setting the right and the left eyes at a fixed position. For the recognition, we represent a face as a set of weighted salient points attributed with a local feature vector, giving an invariant description of the image (signal) around each point. In order to compute the variability of the component of the feature vector, as well as the weight associated to each point, denoting its relative importance in the face representation, we have developed a matching algorithm throughout the same sequence. It is formalized as a maximal matching in a bipartite graph and is approximated, due to a planarity constraint, by a probabilistic relaxation scheme.