Systematic evaluation and benchmarking of YOLOv11, YOLOv12, and YOLOv13 for pipeline weld defect detection via Magnetic flux Leakage

Magnetic Flux Leakage (MFL) inspection is critical for pipeline safety, yet no systematic benchmark exists for evaluating the latest YOLO architectures (v11–v13) on this task. To fill this gap, we present the first comprehensive benchmark for pipeline weld defect detection using MFL data. We propose a data segmentation strategy based on instance density and class distribution, and conduct a multi-dimensional evaluation of accuracy, efficiency, robustness and stability. Our results reveal four main findings. (1) Nano models offer superior accuracy and efficiency: YOLOv12n achieves 93.2% mAP@0.5 and 51.6% mAP@0.5:0.95, with computational efficiency about 3.3 times that of small models and accuracy difference within 1%. (2) Model scaling yields marginal gains: increasing parameters by 3.6 to 3.9 times improves mAP by less than 1%. (3) In dense, imbalanced data (89.6% dominant class), YOLOv11n is the most robust: its mAP@0.5 drops only 2.9 percentage points (vs. 7.5 for YOLOv12n) and minority‑class detection (abnormality in girth weld) is 12.7 points higher than that of YOLOv12n. (4) Data distribution decisively impacts performance – minority‑class detection falls up to 27.3 points on the dense set. This study establishes a reproducible benchmark for intelligent MFL inspection and provides practical guidance for model selection in industrial vision tasks.