Face pose normalization for identity recognition using 3D information by means of neural networks

In this paper, an approach is proposed in order to solve the problem of pose variations for face recognition using single image per person by means of neural network models. Usually, face recognition systems lose their efficiency in excessive pose variations. We used some neural network models for pose detection, identity recognition and depth map reconstruction. In this approach, each facial image is first converted to its frontal view pose and then recognized. First, pose of each image is estimated by a pose classifier model. Then, using the estimated pose code besides the 2D image information, depth map is reconstructed. Finally, synthesized frontal view image, which is used for identity recognition, is provided using the estimated depth map and pose code. Our main purpose is structural modification of neural networks for the task of face recognition using single image per person. Adding top down connections from the first context representation provided by the bottom up process in the bidirectional depth reconstruction model makes the model more adaptive to new unknown variations. For evaluation of the proposed models, we used Bosphorus data. The proposed flowcharts are compared with other methods, and the best identity recognition accuracy is achieved after rotating images to their frontal view by means of the depth reconstruction bidirectional model. The best proposed flowchart shows 18.3% improvement compared with simple identity recognition and 3.57% improvement compared with the second best method.