Local invariants and local constraints for face recognition

We propose a multi-scale face recognition algorithm, robust to the geometrical distortions of the images and to the changes of lighting conditions. A face is represented by a balanced set of points and each point is attributed with a local feature vector, based on the orthogonal Zernike moments, extracted at several scales. The similarity measure between two faces is a measure of the best multiscale matching between those feature points of the two faces, each matching being obtained by dynamic programming. Experimental results shows the robustness of the method, under changes of expressions, lighting conditions, and inaccurate geometrical normalization