Occlusion Robust Face Recognition Based on Mask Learning with Attention Mechanism

ABSTRACT

Wearing masks in public places is an efficient strategy to prevent infection and slow the spread of COVID-19. However, masked face recognition is challenging due to a lack of information about facial features in the masked area. We present a unique Switching Replacement Attention Network (SRAN) for robust face identification based on attention mechanism, which is inspired by the human visual system, which exclusively focuses on non-occluded facial areas. Firstly, a replacement module is established by training a segmentation network to segment the location of the occlusion item. To exclude the corrupted feature components from recognition, we multiply the occluded object’s segmentation mask with the original image features. To determine when the replacement module is applied, we use a lightweight switch module that is both fast and accurate. The proposed technique outperforms state-of-the-art systems on a variety of occluded and non-occluded facial recognition datasets, according to test results. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.