Face Mask Invariant End-to-End Face Recognition

Abstract

This paper introduces an end-to-end face recognition network that is invariant to face images with face masks. Conventional face recognition networks have degraded performance on images with face masks due to inaccurate landmark prediction and alignment results. Thus, an end-to-end network is proposed to solve the problem. We generate face mask synthesized datasets by properly aligning the face mask to images on available public datasets, such as CASIA-Webface, LFW, CALFW, CPLFW, and CFP. Then, we utilize those datasets as training and testing datasets. Second, we introduce a network that contains two modules: alignment and feature extraction modules. These modules are trained end-to-end, which makes the network invariant to face images with a face mask. Experimental results show that the proposed method achieves significant improvement from state-of-the-art face recognition network in face mask synthesized datasets.