YOLO-DMSA: Improved Low-Light Object Detection via Adaptive Feature Extraction and Attention Fusion

In low-light environments, owing to the restrictions of illumination conditions, images are affected by problems such as contrast loss, noise interference, and color distortion, which notably increase the challenge of object detection. To address the aforementioned problems, this paper proposes an improved method called YOLO-DMSA. First, to address the issue of capturing global information in low-light images, a Dual-Mode Adaptive Feature Extraction module is designed, enabling the efficient extraction of the global features of objects. Second, aiming at the problems of low attention to key regions in low-light images and significant background noise interference, Manhattan-SLA module is constructed to focus attention on key regions and suppress noise interference. Finally, considering the characteristics of large object-scale variations in low-light environments, YOLO-DMSA upgrades the detection head to ASFFHead, which realizes the dynamic fusion of feature maps at diverse scales, thereby strengthening the detection robustness of multi-scale objects. Experimental outcomes on the DarkFace and ExDark datasets show that, compared with the baseline algorithm, YOLO-DMSA attains a 1.8% and 3.5% increase in mAP@0.5, respectively, verifying that the proposed method exhibits favorable detection performance. The code is available at https://github.com/Jiaqihang1/YOLO-DMSA.