Coal mill plays a crucial role in thermal power plants, however, operational issues such as coal leakage can lead to coal accumulation, adversely affecting equipment performance and resulting in energy wastage. The challenging operational environment of coal mills, characterized by dim lighting, dust interference, and other factors, poses difficulties for coal leakage detection. To address these challenges, this paper proposes a staged coal leakage detection method based on improved YOLOv5s and improved MobileNetV3-Small. The method operates in two stages: firstly, YOLOv5s is employed to identify areas prone to coal accumulation, and secondly, MobileNetV3-Small is utilized to confirm the presence of coal accumulation in these areas. In addition, the coordinate attention mechanism is introduced to both network, which enhances the attention of the network with fewer parameters. This approach effectively enhances detection accuracy with minimal computing cost by leveraging two improved networks. Experimental results demonstrate an impressive accuracy of 93.7% in detecting coal accumulation phenomena. Moreover, the entire process is executed within 0.039 seconds, meeting the requirements for real-time detection.