Abstract: To address false detections and missed detections of underground safety equipment, such as helmets, self-rescuers, and miner’s lamps, caused by poor illumination, heavy dust, and other complex environmental factors in coal mines, a small-object enhanced detection network, namely LSD-YOLO, is constructed based on the YOLOv11n framework. First, an illumination-aware spatial-channel adaptive modulation module (LASCAM) is introduced into the neck network to perform channel-wise affine compensation and spatial saliency modulation on feature responses in low-light and backlit scenes. Meanwhile, a frequency-aware small-object pyramid module (FSPM) is designed to enhance the detail representation of small objects through multi-scale frequency decomposition and high-frequency modulation. During training, a low-light small-object detection friendly loss (LSD-Loss) is incorporated to strengthen effective learning signals, and a scale-adaptive task-aligned assignment strategy (SATAD) is introduced to adapt the positive-sample matching process to object scales, thereby improving training stability. Experimental results show that, compared with YOLOv11n, the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 3.7% and 2.2%, respectively, and achieves more stable detection performance in scenarios with low light, backlighting, and a high proportion of small objects.