Low-Light Image Quality Enhancement through Bayesian Optimization using Gaussian Processes

摘要

Enhancing image quality under low-light conditions is a crucial challenge for autonomous robotic systems, especially in tasks such as navigation, inspection, and search-and-rescue operations, where visual data reliability directly impacts decision-making and performance. Low-light environments often degrade image quality, compromising the ability of robots to detect, classify, and interact with objects effectively. This paper proposes a novel approach that improves low-light image quality by combining brightness, contrast, and noise reduction operations, tailored for robotic vision applications. The optimal parameters for these operations are determined through Bayesian Optimization using Gaussian Processes, ensuring robust and efficient processing. Experiments conducted on the Low Light paired dataset (LOL) demonstrate the effectiveness of the approach, with image quality assessed using reference-based metrics (PSNR and SSIM) and a no-reference metric (NIQE). The results show that the proposed method outperforms state-of-the-art techniques, delivering high-quality enhanced images that are well-suited for use in robotics and other computer vision tasks.