Face Recognition with Harmonic De-lighting

Evaluations of the state-of-the-art of both academic face recognition algorithms and commercial systems have shown that performance of most current technologies degrades due to the variations of illumination. We propose a novel technique for face recognition under generic illumination in this paper, namely, calibrating an input face image to an image under canonical illumination, named as face de-lighting, to reduce the negative effect of non-ideal illumination in the original image. The canonical illumination is defined as an illumination environment in which light is constant in every direction. Inspired by the low dimension effect of light on Lambertian surface and the compact representation of the canonical illumination in spherical frequency space (only the DC component needed), face de-lighting is achieved with spherical harmonics. Experiments show the effectiveness of the proposed method in both lighting estimation and face recognition.