Abstract: In view of the complex operational environment of conveyor belts in coal mine belt conveyors and the difficulty in distinguishing foreign objects from background noise, and the existed problems of severe missed and false detections, and cannot meet the practical requirements of multi-scale and strong interference in the small-sized foreign object detection of existing universal models (such as YOLOv8s). An improved detection network YOLOv8-FEDNet for coal scenes is constructed. This model introduces the Coal Background Suppression Attention Module (CBAM-C) into the backbone network to enhance the feature expression in foreign object areas and effectively suppress large-scale coal background interference; Small foreign object detection branches with high-resolution are added in the feature fusion stage to improve the detection capability for small-sized and irregular foreign objects; Meanwhile, the foreign object difficult sample is introduced to perceive loss function (FO-Focal Loss), guide the model to pay more attention on foreign object targets that difficult to be distinguished, and solve the problems of small amount and prone to be overlooked of foreign object samples. The experiments show that on the self built coal foreign object dataset, YOLOv8-FEDNet has an increase of 2.11 percentage points in mAP@0.5 and an improvement of 1.19 percentage points in mAP@0.5:0.95 compared to the baseline model YOLOv8s, with a significant improvement in the accuracy of small-sized foreign object detection, verifying the effectiveness of the detection branches with high-resolution in the aspect of small target identification. This method has good engineering application value while ensuring real-time inference performance, providing an efficient solution for foreign object detection of conveyor belts in coal mine belt conveyors and helping to improve the safety and reliability during the process of coal transportation.