Lightweight real-time detection method for coal flow blockage based on Star-YOLOv8

Stable operation of the scraper conveyor in the fully mechanized mining face is a key factor in advancing the intelligent construction of coal mines. Existing vision-based detection methods mainly focus on equipment pose monitoring and recognition, while neglecting the problem of coal flow blockage caused by large coal pieces or coal pileups. Moreover, they generally suffer from high computational cost, insufficient detection accuracy, poor robustness, and limited real-time performance. To address these limitations, a real-time lightweight coal flow blockage detection method based on Star-YOLOv8 is proposed. Without increasing hardware costs, the proposed method utilizes the existing video monitoring system of the fully mechanized mining face to achieve real-time recognition and early warning of scraper conveyor blockages, thereby ensuring safe and efficient production. The proposed approach replaces the backbone network of YOLOv8 with StarNet, which significantly enhances the feature extraction capability of underground images while reducing the number of parameters. Furthermore, a contextual star-shaped fusion module is introduced to strengthen multi-dimensional feature fusion, and a weight-sharing detection head is designed to reduce computational complexity. This ensures real-time performance while maintaining detection accuracy. To verify the effectiveness of the proposed method, experiments were conducted on the DsLMF+ underground longwall mining face dataset and real blockage images collected from a coal mine. Experimental results demonstrate that, compared with the original YOLOv8 model, the proposed Star-YOLOv8 achieves a 4.57% improvement in mAP@0.5:0.95, a 1.06% increase in detection accuracy, a reduction of 1.2 M parameters, a decrease of 2.714 GFLOPs, and a 0.075 ms improvement in single-frame detection speed under the COCO evaluation standard. These results indicate that the proposed method achieves significant improvements in both detection accuracy and model lightweighting. Compared with traditional object detection algorithms, the Star-YOLOv8 model exhibits superior overall performance and practical applicability in coal flow blockage detection tasks.