Abstract: Oversized coal blocks can easily cause poor coal flow, blockage, and coal stacking when transported on a belt conveyor. However, the differences in appearance and color between large coal blocks and ordinary coal blocks are small, and there are obstructions and stacking between coal blocks. Existing coal block detection methods are not precise enough to distinguish between large coal blocks and ordinary coal blocks, which can easily lead to missed or false detections. In order to solve the above problems, a modified YOLOv5 model is proposed for detecting large coal blocks in belt conveyors. The model uses parallel dilated convolution modules to replace some ordinary convolution modules in the YOLOv5 backbone network. It expands the receptive field, improves multi-scale feature learning capability, and better distinguishes large coal blocks from ordinary coal blocks. The joint attention module is added to the neck network to better integrate contextual information and improve the positioning capability for large coal blocks. The model uses the trained improved YOLOv5 model to detect real-time coal transportation videos captured by the camera, and links PLC alarms in real-time based on the quantity information of large coal blocks. The experimental results show that compared to the original YOLOv5 model, the improved YOLOv5 model has improved recall and average precision by 3.4% and 2.0%, respectively. PLC can operate corresponding indicator lights and buzzers to alert based on the quantity of large coal blocks detected by the improved YOLOv5 model. The improved YOLOv5 model is applied to actual coal transportation videos in coal mines, with a detection precision of 97.0% for large coal blocks, effectively avoiding missed and false detections.