Improving the performance of one-shot face recognition by using data augmentation

Abstract

For a past few years, the revolution of deep learning techniques has emerged and launched several state-of-the-art models, for instance, the breakthroughs of DeepFace and DeepID to face recognition in 2014. The face recognition in CCTV systems commonly encounters a few obstacles coming from practical conditions, such as ambient light, the diverse positions and angles of cameras, face masks, face poses, and so on. In addition, people who are monitored by the CCTV systems lack photos and typically have only one photo. These problems lead to face recognition reported with unstable performance and difficult to be successfully used in practice. To tackle these problems, this paper proposes an approach, namely ISE, to face augmentation which interpolates multiple samples from an original photo. Particularly, the samples produced by ISE contain real characteristics of cameras in the CCTV systems. By practically deploying a CCTV system at the Bach Khoa Dormitory, ISE indicates that it can boost the performance of face recognition up from 72%, 46% to 84%, 64% in daytime and day-and-nighttime, respectively.