Face Recognition Method of Mask Occlusion

Under the influence of the COVID-19, people can effectively prevent New Coronavirus infection by wearing masks in public places. However, the mask obscuration causes some face recognition systems to fail to recognize properly. Therefore, in this paper, we propose Multi-Residual Attention Network(MRANet) based on deep convolutional neural network for recognizing faces obscured by masks and improve a loss function. In our model, an attention mechanism and multiple residual layers skip connections are introduced, which allow the model to focus more on the unobscured facial information and contribute to increase the efficiency of information flow and gradient flow between each network layers. A dynamic addictive angular margin loss function, a more reasonable decision boundary function, is also proposed to improve the model’s discriminative power and convergence speed. Our algorithm can effectively identify and verify not only normal unobstructed faces, but also faces obscured by masks. We achieved a accuracy rate of 96.7% on the widely used Labeled Faces in the Wild dataset (LFW), a accuracy rate of 84.837% on the Real-world Masked Face Recognition Dataset (RMFRD), and the false accepted rate in the simulated face recognition system is as low as 0.944%.