<p>Conveyor belts are crucial for efficient coal mining and safe production in coal mines. To address challenges, such as complex mine environments, blurry images, dense and overlapping targets, and limited computing power of equipment, a lightweight foreign object detection algorithm for coal mine conveyor belts based on CMCF-DETR is proposed. First, an efficient StarNet is introduced into the main network, which maps data to a high-dimensional nonlinear feature space through star-shaped operations, providing rich feature representations while significantly reducing network complexity. Second, an Efficient Intra-Scale Feature Interaction (EIFI) Module is proposed, which replaces the explicit key–value interaction of Transformer by simple linear transformation to achieve better global context coding. The Dynamic Multi-Scale Feature Optimization and Fusion (DFOF) Module enhances complex information representation and promotes multi-scale information interaction through dynamic group convolutions, channel reshuffling, and Transformer architecture. Subsequently, the CIoU loss function is introduced to enhance bounding box localization accuracy by incorporating centroid distance and width-to-height ratio penalties. Additionally, the dual-mode low-light image enhancement algorithm (DLEA) is proposed, which combines sparse matrices with a fast solver to improve image input quality in complex environments. The experimental results show that the accuracy of CMCF-DETR is 92.7%, mAP0.5 is 87%, and the inference speed is 97.3 FPS. Compared with the baseline model, the accuracy rate is increased by 5.4%, the mAP0.5 is increased by 3.2%, the parameter quantity is reduced by 47%, and the FLOPs are reduced by 56.6%, which significantly improves the accuracy and real-time performance of foreign body detection in coal mine conveyor belt.</p>