Reinforced Diffusion for Image Denoising

This paper introduces a novel approach to image denoising by combining anisotropic diffusion with reinforcement learning. It addresses the limitations of traditional diffusion methods by learning a sequence of diffusion actions using deep Q-learning. The core contribution lies in the adaptive nature of the learned diffusion process, allowing it to better handle complex image structures and outperform existing diffusion-based and even some CNN-based methods. The use of reinforcement learning to optimize the diffusion process is a key innovation.