Automated inspections using unmanned aerial vehicles (UAV) and computer vision have revolutionized complex infrastructure maintenance, like port and power system. Yet existing methods face significant challenges in low-light environments crucial for nighttime monitoring. To address this gap, we propose SCINet-YOLOv8, an enhanced target detection framework that integrates a Self-Calibrating Illumination Network (SCINet) into YOLOv8's architecture. Our approach combines a hybrid dataset—1,200 synthetic low-light images generated via pixel-level adjustments and 200 real low-light UAV-captured scenes—to train a model that adapts to diverse lighting conditions. Experimental results demonstrate a 3.6% improvement in mean average precision (mAP) over the baseline, achieving 98.4% mAP overall, with critical categories like tension machines (ylj) reaching 99.4% average precision under extreme darkness. Despite integrating SCINet, the model retains real-time efficiency at 45 frames per second and lightweight deployment with only 3.2 million parameters, ensuring practicality for UAV-based applications such as nighttime grid inspections and post-disaster recovery. This work bridges the gap between theoretical advancements and real-world low-light detection demands.