A Robust Processing Chain for Face Recognition under Varying Illumination

In order to make face recognition more reliable under varying illumination, a robust processing chain is presented in this paper. Most of the illumination normalization methods treat all face images in the same way without considering the specific illumination condition of each probe image. For the nearly welllit face images, they may be misclassified after illumination normalization. But they can be correctly classified without illumination normalization. To address this problem, the illumination quality index (IQI) of face image is proposed. According to the IQI of a probe face image, it can be determined whether the illumination normalization should be applied to it. In the proposed processing chain, the probe face image needing no illumination normalization will directly be used for recognition using normalized correlation. Otherwise a new illumination normalization approach, based on the Retinex theory and the total variation under L2 norm (TVL2) constraint model, is conducted on it. The proposed illumination normalization approach utilizes the edge-preserving capability of the TVL2 model, and can effectively weaken the halo effect. Gradient direction and magnitude are extracted from the illumination normalized face image, and then fused at decision level for recognition. The experimental results on ‘Yale B+ Extended Yale B’ face database demonstrate the robustness of the proposed processing chain.