Occlusion-Aware Face Retrieval: Identity Restoration through Generative Inpainting

Occlusions remain a major obstacle for reliable face recognition and person re-identification, as they obscure critical identity features and distort embedding representations. In this work, we present a unified and reproducible pipeline for evaluating both the impact of occlusion and the effectiveness of generative inpainting in restoring identity-preserving facial information. Using the LFW dataset, we generate realistic occlusions via semantically guided masks with Stable Diffusion, and subsequently apply state-of-the-art inpainting methods for occlusion removal. Reconstruction quality is assessed using a comprehensive set of structural and perceptual metrics, while identity preservation is evaluated through retrieval experiments across diverse embedding architectures representing both supervised and self-supervised learning paradigms. Experimental results demonstrate that inpainting can significantly mitigate the adverse effects of occlusion on retrieval performance. Furthermore, our findings reveal that conventional perceptual metrics do not fully capture identity consistency, highlighting the need for embedding-aware evaluation. This study offers a complete benchmarking framework and cross-model analysis to advance research on occlusion-robust face recognition.