Face Restoration Via Generative Adversarial Networks

The objective of face restoration or de-occlusion is to eliminate occluded regions in facial images that are caused by objects or disguises. This study proposes an innovative approach to address the challenge of facial image de-occlusion. Specifically, the proposed method employs a Deep Convolutional Generative Adversarial Network (DCGAN) to restore blocked regions of facial images, resulting in enhanced visual clarity and completeness. We utilize GANs which are trained on both occlusion-free and occluded images, enabling the restoration of corrupted facial regions. Furthermore, we achieve effective model training through mixed precision training. Our method shows promising results on testing images, successfully restoring facial images with up to 50% blocked regions.