Image Denoising Using Fully Connected Network with Reinforcement Learning

Abstract

Deep reinforcement learning (DRL), where an agent learns behaviors in an environment by actions and receiving rewards, has been applied successfully in the robotics area and game controllers at a human level. However, the application of DRL in image processing is still scarce. In this paper, we present a novel approach for image denoising by combining a fully connected network with the gated recurrent unit in an asynchronous advantage actor-critic scheme. The proposed method assigns an agent to every pixel of the input image, and the agent changes the value of each pixel by selecting an action from a predefined list. The goal is to learn an optimal policy to maximize the reward at all pixels of the image. We conduct the denoising experiments on the BSD68 dataset and the results show that the proposed approach produces equivalent or higher PSNR scores compared to several state-of-the-art models based on supervised learning. Our approach is interpretable to humans by showing the agent’s actions, which is a significant difference from original CNNs.