Patch-Wise Normalization for Pose-Invariant Face Recognition from Single Sample

One of the knotty problems in face recognition is that the accuracy drops a lot in the case of recognizing faces captured in varieties of poses. In this paper, a patch-wise normalization method for non-frontal face images is proposed which is practical, effective, and can work well in continuous poses. Firstly, the correlation between 3D face model and the input 2D face image is constructed by the five facial landmarks detected from the input image. And then, a grid of key-points which are manually labelled on the 3D model are mapped onto the probe face image. Finally, for a pair of patches located at the positions which have the same semantic meaning in face images, a warp is estimated by calculating the homography matrix. The reconstructed frontal face image is obtained by stitching all patches sequentially. Recognition experiments on the FERET database proves that the proposed method is effective.