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

Abstract

In large industrial pipeline transportation systems, detecting small leakage targets is significantly hindered by a cluttered background and substantial variability in leakage sizes. To solve this problem, the EP-YOLO pipeline leak detection model was proposed. Firstly, the model introduced a small target detection layer to improve the accuracy of minor leak detection. To further enhance feature extraction, the EfficientViT module, combined with the Lightweight Multiscale Attention (MSA) module, was employed to restructure the original backbone network. This allows the model to capture global context information from shallow feature maps, thereby improving the detection efficiency of leakage targets across various scales. In addition, a lightweight layer aggregation network structure based on the SimAM parameter-free attention mechanism is designed to enhance the dependence between global and feature information in the model, thereby improving the detection capability for small leakage features while reducing the model’s computational complexity. Experiments conducted on the newly established pipeline leakage dataset demonstrate that the EP-YOLO model’s Average Precision (mAP) reaches 92.3%, representing a 4.9% improvement over the original YOLOv7 model. Moreover, it outperforms other methods in detecting small leaks. Finally, a pipeline leak detection system was developed to facilitate automatic leak detection via a visual interface.

Keywords

• Pipeline
• YOLO
• leak detection
• four detection scales
• deep learning