Improved Low-Light Image Feature Matching Algorithm Based on the SuperGlue Net Model

The SuperGlue algorithm, which integrates deep learning theory with the SuperPoint feature extraction operator and addresses the matching problem using the classical Sinkhorn method, has significantly enhanced matching efficiency and become a prominent research focus. However, existing feature extraction operators often struggle to extract high-quality features from extremely low-light or dark images, resulting in reduced matching accuracy. In this study, we propose a novel feature matching method that combines multi-scale retinex with color restoration (MSRCR) and SuperGlue to address this challenge, enabling effective feature extraction and matching from dark images, successfully addressing the challenges of feature point extraction difficulties, sparse matching points, and low matching accuracy in extreme environments such as nighttime autonomous navigation, mine exploration, and tunnel operations. Our approach first employs the retinex-based MSRCR algorithm to enhance features in original low-light images, followed by utilizing the enhanced image pairs as inputs for SuperGlue feature matching. Experimental results validate the effectiveness of our method, demonstrating that both the quantity of extracted feature points and correctly matched feature points approximately doubles compared to traditional methods, thereby significantly improving matching accuracy in dark images.