Self-supervised attention-based domain adaptation for robust occluded face recognition

Abstract

Cross-domain face recognition remains a challenging problem when datasets exhibit significant variations in data distribution, imaging conditions, and occlusion patterns. These challenges are particularly critical in practical scenarios where faces are partially occluded by masks, glasses, or other objects. Existing domain adaptation methods often treat occlusion robustness and domain alignment as separate problems, which limits their ability to generalize across domains under complex occlusion conditions. This paper introduces SelfAttDANet, a unified domain adaptation framework designed to jointly address occlusion robustness and domain shift for cross-domain face recognition. Instead of relying on independent modules, the proposed framework integrates self-supervised representation learning, attention-guided feature refinement, and adversarial domain alignment within a coherent learning strategy. The model first employs self-supervised learning objectives to learn domain-agnostic feature representations that remain stable under occlusion variations. These representations are then refined through an attention-guided mechanism that emphasizes discriminative facial regions while suppressing occlusion-induced noise. Finally, an adversarial domain alignment mechanism is incorporated to encourage domain-invariant feature distributions between source and target domains. Extensive experiments conducted on CASIA-to-LFW and CASIA-to-AR cross-domain adaptation scenarios demonstrate that the proposed framework achieves improved recognition performance under both synthetic and real occlusion conditions while maintaining computational efficiency suitable for real-world deployment.