Single Image Dehazing via Reinforcement Learning

Single image dehazing aims at restoring the clean image from single hazy one. A novel Reinforcement Learning (RL) based image dehazing method is proposed in this paper to handle the dehazing task with interpretable ability and extensibility. We model the dehazing problem as a Markov Decision Process (MDP) with several existing simple traditional image processing operations and prior knowledge-based dehazing methods as actions. Furthermore, a deep Q-learning network is established to learn the value function for image dehazing. Finally the learned network can iteratively choose the correct action during the processing sequence to produce dehazing results. Extensive results on real hazy images have been conducted to verify the proposed method. Also the learned sequence of image dehazing can provide considerable guidance for human.