Masked Face Recognition Based on ConvNeXt With Coordinate Attention and AdaCos Loss

Compared with traditional identity authentication methods, biometric recognition technologies offer superior security, reliability, and accuracy, among them, face recognition is of great research value. However, the outbreak of the COVID‐19 epidemic, as well as the cold and hazy weather conditions have led to a widespread preference for wearing masks. The masks will cause the failure of traditional face recognition methods. This paper develops a deep learning‐based biometric recognition method for masked faces. A simulated masked face (SMF) dataset and a real masked face dataset are constructed first. The images are preprocessed with masked face detection and mask edges are identified to segment the face area. A ConvNeXt network with coordinate attention (CA) is proposed and AdaCos is employed for comprehensive feature extraction. The results of extensive experiments demonstrate a remarkable 99% recognition rate on the SMF dataset and 94% on the real masked face dataset. It shows that our method can effectively process face recognition task with masks, with very high performance in terms of accuracy.