EPA-YOLO: A Lightweight Pipeline Leak Detection Algorithm Based on Improved YOLOv7

This paper proposes a lightweight EPA-YOLO pipeline leakage detection model designed to address the frequent leakage issues in large chemical plant pipelines, mainly focusing on the challenges of detecting small-scale and dispersed leaks. By constructing a lightweight ELAN-P layer aggregation module, the model enhances feature extraction capabilities for pipeline leakage regions while significantly reducing computational complexity. Furthermore, an AFPN progressive pyramid network is employed for multi-scale feature fusion, effectively improving the detection of small leak targets. Lastly, the FReLU activation function is introduced into the convolutional layer, enhancing the model's sensitivity to spatial information. Experimental results demonstrate that the proposed algorithm performs well on a real-world dataset, achieving a mean Average Precision (mAP) of 90.3%, 2.9% higher than the original model, while reducing model parameters by 4.1M. The FPS rate is 39.1 (frames/s), which is better than the original YOLOv7. The model shows a significantly lower missed detection rate in complex pipeline scenarios, underscoring its practical application value.