Lightweight Network for Masked Face Recognition Based on Improved Dual Attention Mechanism

Abstract

More and more people are wearing masks to avoid contracting COVID-19. However, masked occlusion will lead to the loss of facial features. The universal face recognition model cannot recognize the masked face accurately and quickly. Masked face recognition (MFR) has become a very urgent challenge. Based on MobileNetV2, this paper will replace the average pooling in the attention module with depth-separable convolution, also the normalization operation is improved by replacing the Relu activation function with Prelu. The improved dual attention module is introduced to redistribute the weight parameters of bottleneck. The results show that the accuracy of mask-LFW and mask-AgeDB data sets reached 90% and 91% respectively, and the model size is reduced to one tenth of the other common model size. It is proved that the improved network can effectively reduce the occlusion interference, reduce the calculation amount, and improve the robustness of the system.