A lightweight approach for accurate detection of pipeline weld defects

Pipeline welding is essential in oil and gas transport, where weld quality determines pipeline safety and reliability. Although deep neural networks excel at automated weld defect detection, many models are heavy and demand high computation, hindering real-time industrial use. We propose a lightweight weld defect detection framework based on an improved YOLOv8n architecture, incorporating three innovations. First, a FasterNet-inspired feature extraction and fusion module (C2f-FCG) preserves channel information while significantly reducing computational load. Second, a re-parameterisable GhostNet-based convolution module (RGE) generates redundant feature maps through cost-efficient operations, achieving substantial parameter compression and lower computation. Third, a detail-enhanced shared convolution detection head (DetectLSC) improves feature consistency and boosts detection accuracy. Experiments show that our approach reduces parameters and computational load by 63% and 51%, respectively, compared to the baseline, while increasing average detection accuracy by around 2%. Comparative tests on public datasets validate robust generalisation across diverse defect scenarios. Deployment on standard industrial GPUs confirms real-time inference speeds, highlighting its field practicality. These results demonstrate that our lightweight model effectively balances efficiency and accuracy, making it well-suited for real-time industrial weld defect detection.