To address the challenge of balancing detection performance and real-time capability under complex factors in coal mine environments, such as noise, multi-scale targets, and occlusion, this paper proposes a lightweight and high-precision foreign object detection method for underground conveyor belts based on an improved YOLOv11. Firstly, building upon YOLOv11, a dynamic convolution module integrating multi-path channel attention (DCNv2-Dynamic) is introduced to replace the standard convolutions in the C3k2 blocks of the backbone network, enhancing the capture of critical features in complex scenes. Secondly, a 160×160 detection layer is added to strengthen small object perception, while the redundant 20×20 output layer is removed to reduce computational load. Furthermore, a Separate and Enhance Attention Module (SEAM) is embedded into the detection heads to mitigate the information loss caused by occlusion. Experimental results show that the improved model achieves significant performance gains, achieving a 3.7% improvement in mAP50 and a 10.8% improvement in mAP50:95, alongside a 26.2% reduction in computational cost and a 21.8% reduction in model size, thereby establishing an efficient and reliable solution for foreign object detection in coal mine conveyor belts.