To address the issues of missed and false detections caused by occlusions and scale variations of foreign objects such as large gangue stones and anchor rods on underground coal mine conveyor belts, an improved detection model, SDGW-YOLOv11, is proposed. First, to achieve effective detection of occluded objects through multi-perspective feature fusion and consistency regularization, and to extract features from multiple positions and scales, the SEAM attention mechanism is introduced into the neck network of YOLOv11. This mechanism reduces the interference caused by occlusion during detection. Second, to enhance the model′s adaptability to the size variations of objects, both occluded and unoccluded, the C3k2_DCN module is designed and integrated into the backbone network of YOLOv11, improving the model′s local perception capability for objects. Finally, to prevent the attention mechanism from significantly increasing the model size and affecting detection speed, the model is optimized by replacing some conventional convolutional layers with GhostConv to reduce the number of parameters and adopting the WIoU loss function to replace the original loss function, thereby accelerating convergence.Experimental results show that the SDGW-YOLOv11 model achieves a detection accuracy of 86.1%, representing a 4.6% improvement over the original model. The optimized model achieves a detection speed of 82 fps second (FPS), fully meeting the requirements for real-time detection of conveyor belt foreign objects. The improved model outperforms Faster R-CNN, SSD, YOLOv3, YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10, and YOLOv11 in both precision and mAP@0.5, effectively reducing missed and false detections caused by occlusion and scale variation. It is better suited for foreign object detection in underground coal mine conveyor belt scenarios.