Boosting denoisers with reinforcement learning for image restoration

Learning-based image restoration approaches typically learn to map distorted images to clean images. To remove multiple combined distortions with unknown mixture ratios, most of the existing methods have focused on the development of different deep neural network architectures and novel loss functions. Although these methods have proved their effectiveness on image restoration tasks, they require expensive training data and produce results in a noninterpretable way. In this work, we present a deep reinforcement learning (DRL) based method to restore the distorted images, which casts an image restoration Problem as a Partially Observable Markov Decision Process (POMDP) where actions are defined as multiple pixel-wise image denoising operations. In our method, each agent possesses a pixel, the agent learns to adjust the corresponding pixel value by determining the proper combination of the actions. We also develop a novel exploration scheme such that similar actions have similar value, thereby avoiding overfitting in state-action value estimation. Through extensive experiments, we show that our method can restore images with multiple combined distortions and our DRL approach performs comparable or better performance against previous learning-based approaches. By visualizing the process of weighting multiple pixel-wise operations, we can identify what combination of operations is employed for each pixel at each stage. We believe our work takes a step toward the explainability and interpretability of learning-based image restoration methods.