Face Restoration Based on GANs and NST

Abstract

Image restoration technology is a research hotspot in the field of deep learning computer vision. However, due to the complexity of facial textures, the existing face restoration algorithms lack visual coherence. We propose a face restoration algorithm (GNST) based on Generative Adversarial Networks (GANs) and Neural Style Transfer (NST), which uses the generative network to repair the damaged facial contents and then uses the style transfer to adjust the overall style. At the stage of repairing content, we design four losses from different aspects. We add two total variation losses (tv loss) besides the traditional content loss and adversarial loss. The first tv loss can make the generated image smoother and cleaner. The second tv loss can effectively prevent the generator's gradient collapse when generator "cheat" the discriminator. In addition, it is used to highlight facial important features. Experiments show that G-NST achieves better results than existing methods.