Multi-pose Face Recognition Based on TP-GAN

In recent years, a large number of smart products have been widely used and become an indispensable part of people's life. At the same time, the identity authentication by means of face recognition has become familiar. Although the traditional face recognition system can achieve a high recognition rate in most environments, as the external factors change, such as lighting, blocking, and posing, the performance of the system degrades. In this paper, we propose a new deep network method based on TP-GAN and investigate the behavior of the proposed system when there is a change in the angle pose of the face. In the generation part, we propose a deeper convolutional neural network to extract the pose-invariant face features and synthesize the virtual pose, simultaneously. The deeper network is divided into multiple overlapping local networks, each of which was trained to synthesize a small pose change; the joint training local network synthesizes the front face from the non-positive pose in a progressive manner. By stacking multiple local networks, we can extract more robust pose-invariant features and generate multiple virtual poses in front of the synthetic front. Face recognition with different postures is achieved by combining pose-invariant features and virtual postures. Experimental results demonstrate that our method has achieved superb results in pose-invariant face recognition.