Improving the YOLOv8-based Intelligent Detection Method for Abnormalities in Coal Mine Roadway Pipelines

To address the issues of existing abnormal pipeline identification algorithms having a large number of parameters that make it difficult to efficiently deploy inspection robots, and the complex variation in the scales of abnormal pipeline targets in detection scenarios leading to low recognition accuracy, this paper proposes an intelligent pipeline recognition method for coal mine tunnels based on an improved YOLOv8. This method introduces an improved backbone feature network using MobileNetV4 on the YOLOv8n model to reduce the number of parameters and enhance feature extraction efficiency, introduces an inverted residual efficient multi-scale attention mechanism (iEMA) to strengthen multi-scale target feature capture, uses Omni-dimensional Dynamic Convolution (ODConv) in the neck network to replace standard convolution to improve local fine-grained feature fusion and recognition accuracy, and introduces the Focal-Eiou loss function to address issues of imbalance between small targets and slender pipeline samples as well as model convergence performance. To tackle the problem of the model being unable to recognize fallen pipelines, a dual-indicator determination method based on pipeline anchor-box and ground overlap rate, as well as pipeline tilt angle, is proposed. Ablation experiments show that the improved model, while reducing the number of parameters by 19%, increases average recognition precision by 3.9%, recall by 4.7%, and mAP@0.5 by 3.3%. Comparative experiments indicate that the improved YOLOv8 model has superior overall performance, with recognition accuracy improvements over YOLOv8n, YOLOv10, YOLOv8-Ghost, YOLOv8-MLCA, Faster-RCNN, RT-DETR, and SSD by 3.3%, 5.7%, 8.3%, 1.7%, 2%, 1.3%, and 2.2%, respectively. The model has only 2.5×106 parameters and 7.7×109 GFLOPs, demonstrating optimal overall performance in model lightweight design. The average detection time is 6.2 ms, with an FPS of 86 frames per second, meeting the real-time and accurate recognition requirements of inspection robots for abnormal pipelines.