Occlusion Aware Deep Face Recognition

Abstract

Face recognition systems achieve high accuracy under controlled conditions, but performance drops significantly with real-world occlusions such as masks, sunglasses, scarves, or hands. As deployments in security, access control, healthcare monitoring, and smart environments increase, demand grows for occlusion-robust solutions reliable in unconstrained settings. This paper systematically reviews 20 peer-reviewed studies (2021–2026) on occlusion-aware face recognition and facial landmark detection using deep learning, classifying approaches into five paradigms: CNN-based, transformer-based, hybrid dual-branch, generative reconstruction/inpainting, and YOLO-based detection. A comparative evaluation covers model architectures, training settings, benchmark datasets, performance metrics, and research gaps. Results show hybrid dual-branch designs—combining local CNN processing with global transformer attention—consistently outperform CNN-only methods across tasks, while generative reconstruction techniques show strong promise for identity-preserving occlusion completion. Key open challenges remain in cross-dataset generalization, extreme occlusion robustness, lightweight edge deployment, and demographic dataset diversity. This review synthesizes prior work, highlights gaps, and proposes future directions for effective occlusion-aware face recognition systems.