Dual-Proxy Modeling for Masked Face Recognition

With the recent worldwide COVID-19 pandemic, almost everyone wears a mask daily, leading to severe degradation in the accuracy of conventional face recognition systems. Several works improve the performance of masked faces by adopting synthetic masked face images for training. However, such methods often cause performance degradation on unmasked faces, raising the contradiction between the face recognition system's accuracy on unmasked and masked faces. In this paper, we propose a dual-proxy face recognition training method to improve masked faces’ performance while maintaining unmasked faces’ performance. Specifically, we design two fully-connected layers as the unmasked and masked feature space proxies to alleviate the significant difference between the two data distributions. The cross-space constraints are adopted to ensure the intra-class compactness and inter-class discrepancy. Extensive experiments on popular unmasked face benchmarks and masked face benchmarks, including real-world mask faces and the generated mask faces, demonstrate our method's superiority over the state-of-the-art methods on masked faces without incurring a notable accuracy degradation on unmasked faces.