Learning Diffusion Functions for Image Restoration

Joel Valdivia Ortega, Manuel Haas, Alexander Effland

Abstract: Anisotropic diffusion models play a major role in numerous image restoration tasks. A key ingredient for these models is the diffusion function, which is normally an a priori fixed function. In this paper, we advocate a novel approach to learning the diffusion function, which is represented as a Fields of Experts (FoE) function or a U-Net. In several numerical experiments, we prove our technique outperforms both the classical models and state-of-the-art algorithms. The generalization to other datasets/restoration problems is also discussed.