Image Inpainting-based Face Recognition Algorithm for Mask-occluded Faces

Since the COVID-19 pandemic, mask-wearing has become a norm, which poses challenges to face recognition due to occlusion. However, existing face inpainting methods often fail to restore sufficient accuracy for recognition. To address this issue, this paper proposes a mask-occluded face recognition algorithm based on image inpainting. First, a Multi-level Interactive Siamese Filtering (MISF) module is introduced into the U-Net architecture to enhance contextual information transmission and preserve more fine-grained details. Second, a joint loss function combining content loss, style loss, and adversarial loss is designed to ensure that the restored images maintain the original facial features and structure, and that the repaired regions blend more naturally with surrounding areas. Third, the SwinIR network is employed for super-resolution restoration, improving image clarity and recognition accuracy. Finally, experiments conducted on the SMFRD dataset show that the proposed algorithm improves recognition accuracy by 3.4% compared to existing methods.