Abstract: The complex and variable environment of coal mines, coupled with the blurriness of conveyor belt images, results in challenges for accurately detecting large coal blocks and poor model generalization. In this situation, a detection method for large coal blocks based on CA-YOLOv8 is proposed. Firstly, considering the characteristics of conveyor belt images, image restoration is used to preprocess the blurred images, thereby enhancing their clarity and providing high-quality inputs for detection of large coal blocks. Then, based on the characteristics of large coal blocks, the YOLOv8 algorithm is improved, including adaptive scaling of the images at the input based on real size, embedding a coordinate attention in the network structure, and removing unnecessary detection heads, which facilitate accurate, stable, and efficient detection of large coal blocks. Experiments conducted on a dataset created from surveillance video images of four different conveyor belts in a coal mine demonstrate that the proposed method achieve a precision of over 95.6% and a recall of over 89.0% under the same environmental conditions. Compared to the original YOLOv8 algorithm, the precision is improved by over 17.0% and the recall is improved by over 15.3% under different environmental conditions.