Occlusion face recognition based on improved attention mechanism

Abstract

Due to the new crown and other epidemic diseases that make people wear masks to travel, the accuracy of the original face recognition system is affected. To address this challenge, a mask-wearing face recognition system based on an improved attention mechanism is proposed. First, Adding a maximum pooling operation to the CA (Coordinate Attention) attention module, then, placing attention module in the residual unit to form a feature extraction network. LResNet18E-IR is selected as the backbone network. Finally, the ArcFace loss and occlusion probability loss are combined to establish a multi-task network, which further promotes the accuracy of occluded face recognition. The results demonstrate that the system effectively increases the recognition accuracy of masked face and maintains almost the same accuracy as the original model on the unmasked dataset.