Purpose This study aims to enhance pipeline defect detection by addressing key challenges in current methods, including low accuracy in detecting small-scale defects, limited multi-scale fusion capability and excessive computational complexity. Ensuring reliable defect detection is vital for pipeline safety and efficient maintenance. Design/methodology/approach A Dynamic Multi-scale Lightweight-YOLO (DML-YOLO) algorithm is proposed, built upon the YOLOv11 framework. The approach introduces three innovations: dynamic convolution (DyC3k2) to improve small-defect feature extraction, DAS-PAN for more effective multi-scale fusion and a HybridLite Head to reduce computational redundancy while maintaining detection accuracy. Findings Experiments on a pipeline defect data set demonstrate that DML-YOLO achieves superior performance compared with YOLOv5, YOLOv8, YOLOv10, YOLOv11 and YOLOv12. The model reached a precision of 0.904, recall of 0.789, mAP@0.5 of 0.870 and an F1-score of 0.839, while reducing FLOPs to 18.1 G. Physical robot experiments further validated its real-time defect detection capability. Originality/value This research provides a lightweight yet high-accuracy detection framework that balances computational efficiency with robustness. By combining dynamic convolution, efficient feature fusion and a streamlined detection head, DML-YOLO demonstrates strong potential for deployment on embedded robotic platforms, contributing both theoretical insights and practical value for industrial pipeline inspection.
Guan et al. (2026) studied this question.