Intelligent Agent-Based Architecture for Low-Light Image Enhancement Using an A3C Framework

Low-light image enhancement (LLIE) is an important area of research as many applications such as photography, video surveillance, and security are confronted with image degradation in low-light environments. This paper presents an intelligent agent-based method for LLIE using the Asynchronous Advantage Actor-Critic (A3C) framework. The enhancement task is effectively cast in this work into the framework of a Markov Decision Process. This method enables an agent to learn a policy that successively improves image quality. In the agent, features are extracted by a Fully Convolutional Network (FCN), a policy network for choosing an action, and a value network for estimating the reward. In training, non-reference loss functions are also used to measure image quality without the availability of the reference image or ground truth images. Such functions include spatial consistency loss, exposure control loss, and illumination smoothness loss, and the approach achieves end-to-end enhancement without reference image. The experimental results on LOL and MIT-Adobe dataset also show that the proposed technique enhances image brightness, Contrast, and structure much better as compared to other state-of-the-art methods. Especially, the methodology proposed scored 25.93 PSNR, 0.932 SSIM, and 0.053 LPIPS on the LOL dataset, achieving better results than related strategies. The designed agent-based approach works under a wide range of low-light situations. This approach allows obtaining enhancement results that will be satisfactory in terms of the users’ preferences and needs of the specific applications. The findings highlight the method's robustness and flexibility, making it suitable for various practical applications. This work demonstrates that reinforcement learning agents have promising applications in improved image processing capabilities, and establishes a new record for low-light image improvement.