A Deep Learning Approach to Terracotta Warriors’ Facial Recovery and Quadratic
Yu Xu, Long Ma · 2022
Facial restoration and animation of degraded images of Terracotta Warriors are important for enhancing the vitality and promotion of cultural relics. In this paper, based on a deep learning theoretical framework, blind facial restoration is performed on degraded images of terracotta warriors using rich and diverse facial priors in pre-trained generative adversarial networks. The facial prior is involved in the image generation process through the spatial feature transformation layer to achieve a good balance of realism and fidelity in the terracotta warrior restoration process. The anime style migration algorithm is used to realize the animalization of the terracotta warrior restoration image by learning the style features of the anime reference image and fitting the style migration process. The experimental results show that the method in this paper achieves facial recovery of the terracotta warriors with high fidelity; the generated secondary terracotta heads have obvious anime style and the content information remains highly consistent with the original images.