Efficient face recognition with variant pose and illumination in video

A novel algorithm to face recognition based on video is presented in this paper. A framework is designed to work for face recognition from video sequence, which is robust to large-scale changes in facial pose and lighting conditions. Two approaches to improve the robustness of the algorithm are presented, a 2D-to-3D face model and self-PCA (principal component analysis) method based on bit-plane feature fusion. In the training stage, the basic input for recognition systems is a single frontal face image, from which an integrated 3D face model can be constructed. Then the virtual face samples which cover different pose are generated by rotating the resultant 3D face model. After that, a bit planes feature fusion approach is applied to construct a new virtual face to effectively reduce the sensitivity to illumination variances. In the recognition stage, an unknown face video sequence is adopted to find the virtual face and the self-PCA is performed. The results clearly show the potential of the combination of 2D-to-3D face model and bit planes-based self-PCA recognition towards pose and illumination variant face recognition in video.