Face Reconstruction from Profile to Frontal Evaluation of Face Recognition

Abstract

One of the main challenges in face recognition is handling extreme variation of poses which may be faced for images collected in labs and in the wild. Recognizing faces in profile view has been shown to perform poorly compared to using frontal view of faces. Indeed, previous approaches failed to capture distinct features of a profile face compared to a frontal one. Approaches to enhance face recognition on profile faces have been recently proposed following two different trends. One trend depends on training a neural network model with big multi-view face datasets to learn features of faces by handling all poses. The second trend generates a frontal face image (face reconstruction) from any given face pose and applies feature extraction and face recognition on the generated face instead of profile faces. Recent methods for face reconstruction use generative adversarial networks (GAN) learning model to train two competing neural networks to generate authentic frontal view of any pose preserving person’s identity. For the work described in this paper, we trained a feature extraction neural network model to learn representation of any face pose which is then compared with each other using Euclidean distance. We also used two recent face reconstruction techniques to generate frontal faces. We evaluated the performance of using the generated frontal faces against the posed counterparts. In the conducted experiments, we used three face datasets that contain several challenges for face recognition having faces in a variety of poses and in the wild.