Mixed Distorted Image Restoration Based on Residual Double Deep Q Network

In this paper, we investigate an image restoration method based on a deep reinforcement learning framework, aiming to recover low quality images to high quality images. The general deep learning-based approach trains a single large network to accomplish a specific task, which is difficult to handle when facing mixed distorted images. To address this problem, we propose the residual Double DQN algorithm, which introduces the idea of residuals into the deep reinforcement learning framework. The agent learns a policy to select appropriate actions from the action set to gradually restore the quality of mixed distorted images. The framework uses the residual blocks to improve the feature extraction ability of the agent, so as to guide it to adaptively select the appropriate action for image recovery. In addition, based on the the new reward function which is designed based on human-eye inspiration, the framework can handle the mixed distortion of images containing noise, blur, and JPEG compression at the same time. The experimental results show that our proposed model has low complexity and is superior to the existing methods in processing mixed distortion images.