Research on partial occluded face recognition based on diffusion models and deep feature completion

Abstract

Partial facial occlusion caused by masks, sunglasses, scarves, and other objects significantly degrades the performance of conventional face recognition systems due to the loss of discriminative identity information. To address this challenge, this paper proposes a Diffusion-driven Deep Feature Completion Network (DDFC-Net) for robust partial occluded face recognition. The proposed framework integrates an Identity-Preserved Diffusion Restoration Module (IPDRM) and a Multi-scale Feature Completion Transformer (MFCT) to jointly recover missing facial structures and reconstruct incomplete identity representations. Specifically, identity-aware constraints are incorporated into the diffusion process to preserve discriminative facial characteristics during restoration, while a transformer-based feature completion mechanism exploits contextual dependencies to infer missing semantic information in latent space. Furthermore, an Occlusion-Aware Adaptive Fusion Module (OAFM) dynamically combines restored image features and completed deep features according to occlusion severity. Extensive experiments conducted on LFW, CFP-FP, RMFRD, MFR2, and CelebA-HQ-OCC datasets demonstrate that the proposed method consistently outperforms several state-of-the-art approaches in terms of recognition accuracy and restoration quality under various occlusion conditions, confirming its effectiveness and robustness for practical occluded face recognition applications.