26th IEEE Signal Processing and Communications Applications Conference (SIU), İzmir, Turkey, 2 - 05 May 2018
Face frontalization increases accuracies of face and gesture recognition applications. In this paper, we propose a 2D patch warping based face frontalization method which that has a simple but efficient flow due to its lower computation cost. We partition the human face into 23 nearly planar regions that are constituted by 68 landmark points to form a frontal face model and used for warping process. Planar places warped by using homography unlike other affine transform based methods. Warping rectangle regions with homography preserve global structure of face as well as it decreased the computational cost of frontalization as againts situations that work with a lot of triangular region like Delaunay triangulation. In order to test recognition performance, every test sample frontalized with respect to average face model computed as the average of all train samples. Test sets created by the pose angles of samples, tested separately to measure the contribution of proposed method to recognition and we compare the proposed method to another state of art frontalization method in literature.