Foreign objects on belt conveyors in coal mines pose significant safety hazards, including potential belt tears and blockages. To address challenges such as the varied characteristics of foreign objects, low ambient lighting conditions, the inefficiency of manual inspections, Given hardware limitations, the present research proposes a streamlined and high-efficiency extraneous item recognition technique developed specifically for low-illumination contexts: YOLO_SFA. This methodology is engineered to facilitate exact on-the-fly identification of non-native objects positioned on conveyor systems amid poorly lit operating conditions. Based on the YOLOv8n framework, we introduce a Self-Attention Mixed Aggregation Feedback Block (SMAFB-CGLU) to the network’s backbone, expanding the receptive field of shallow feature extractors. This helps capture richer semantic information, compensating for the limitation of traditional CNNs, where shallow layers have abundant positional information but lack semantic depth. Additionally, we propose a lightweight detection head, Faster-Head, leveraging shared parameters to reduce model complexity and computation while maintaining accuracy. Moreover, to address background integration and adaptability to varying input sizes, we introduce the AMSPPF module, which bolsters the detection and localization of various extraneous items by preserving both high-frequency features and low-frequency cues. Lastly, the Inner-GIoU loss function is integrated to further improve detection efficacy. Extraneous items on coal mine conveyor belts present significant safety risks, including potential belt damage or operational jams.The experimental results demonstrate that YOLO_SFA improves mAP by 2.9% compared to YOLOv8,while substantially enhancing the identification of extraneous items in dimly illuminated environments