Illumination Subspaces based Robust Face Recognition

In this paper a face recognition system that is based on illumination subspaces is presented. In this system, first, the dominant illumination directions are learned using a clustering algorithm. Three main illumination directions are observed: Ones that have frontal illumination, illumination from left and right sides. After determining the dominant illumination direction classes, the face space is divided into these classes to separate the variations caused by illumination from the variations caused by different identities. Then illumination subspaces based face recognition approach is used to benefit from the additional knowledge of the illumination direction. The proposed approach is tested on the images from the illumination and lighting subsets of the CMU PIE database. The experimental results show that by utilizing knowledge of illumination direction and using illumination subspaces based face recognition, the performance is significantly improved.