To address the challenges of low detection accuracy and poor real-time performance exhibited by the traditional lump coal detection algorithms for the coal conveyor belts in the complex environments, this paper proposes a novel MEB-YOLO multi-scale attention detection model. Firstly, a lightweight backbone network, MobileNetV4, is introduced to reduce the computational complexity and enhance inference speed. Secondly, EMA mechanism is integrated to significantly improve the feature fusion capability, thereby boosting detection performance across the varying target scales. Finally, the BiFPN module is incorporated to facilitate the superior fusion of multi-scale information. Experimental results demonstrate that the MEB-YOLO model achieves the higher detection accuracy and faster processing speed in complex environments compared with conventional algorithms.