A Three-Stage Guided Learning Occluded Face Restoration Scheme Based on GAN

Facial recognition systems often face challenges with partial occlusions, which hinder feature identification and degrade model performance. Although GAN-based methods have shown promise, their reliance on synthetic datasets limits real-world applicability due to the lack of dataset realism. In this paper, we propose a novel three-stage framework that uses occlusion region prediction, facial component segmentation, and face restoration based on the outputs from the initial stages. This approach effectively handles diverse occlusions and bridges the gap between synthetic and real-world scenarios. Experimental results show the superiority of our method, achieving over 91 % on a similarity measure and outperforming existing methods.