Old Photo Restoration Using Adaptive Energy Transition Of Latent Spaces

Recent old photo restoration works have achieved significant improvement using generative adversarial networks.However, the restoration quality under this generative framework is inevitably affected by the encoded properties of latent spaces, which reflect pivotal semantic information in the recovery process.Exploring the regularities in latent spaces is an important issue in restoration task.In this paper, we propose a novel Energy-Transition Generative Adversarial Network (ET-GAN) for old photo restoration.Based on the fact that an energy function can capture the regularities in the data effectively, we propose to learn an energy-based prior model in latent spaces to improve the expressive power of GAN model.Moreover, we employ the Markov Chain Monte Carlo method to sample a batch of latent vectors from the prior distribution as the regularization.By linearly mixing with a regularized latent vector, the original latent vector is transformed along the principal semantic direction and a better latent space can be derived to improve the generation capability of ET-GAN to restore damaged photos with the multiple degradations.Extensive experiments demonstrate that our model can achieve the state-of-the-art performance in terms of numerical metrics and visual effects.