Masked face recognition based on knowledge distillation and convolutional self-attention network

Face recognition has significantly improved with the development of deep learning technology. However, in the case of a viral epidemic like COVID-19, wearing masks reduces the risk of infection significantly but results in losing crucial face features and increasing intra-class divergence, which decreases the effectiveness and accuracy of face recognition. To deal with this issue, a novel masked face recognition (MFR) method based on knowledge distillation and convolutional self-attention network is proposed. Specifically, a knowledge distillation framework is constructed to transmit knowledge from the teacher network to the student network, which enables the student network to focus on unmasked face areas and generate more effective face embeddings. Moreover, a convolutional self-attention network including the shallow feature extraction module (SFEM) and the convolution-Transformer dual-branch module (CTDM) is proposed to extract local- and global-range features for MFR. Experimental results on multiple public datasets demonstrate that our method has superior MFR performance to state-of-the-art methods.