Objective Radiographs of girth welds in long-distance oil and gas pipelines are the core basis for integrity assessment and risk management. Marks on these radiographs—such as weld serial numbers, applicable standards, and weld positions—must be digitally archived. Traditional manual interpretation is labor-intensive, inefficient, costly, and prone to errors due to visual fatigue. Therefore, developing an intelligent, accurate, and lightweight recognition method is imperative.

Methods Based on YOLO (You Only Look Once) v11n as the baseline model, a YOLO-MPW (You Only Look Once for the Marks of Pipeline Weld Radiographs) model was constructed for mark detection on weld radiographs, with three key improvements implemented: (1) A ConvFormer with Convolutional Gated Linear Unit (CF-CGLU) was designed and embedded into the C3k2 module. The gating mechanism was applied to dynamically allocate feature weights, enhance responses to key character regions, and suppress background and occlusion noise. (2) A Lightweight Detection Head (LDH) was designed, in which depthwise separable convolutions were used to replace standard convolutions, resulting in significant reductions in model parameters and complexity while the accuracy was maintained. (3) The Content-Aware ReAssembly of FEatures (CARAFE) sampling operator was introduced to enhance the model’s response to important features and improve the utilization of semantic information in feature maps.

Results Using weld radiographs from a southern China oil and gas pipeline for training and validation, the YOLO-MPW model improved the mean Average Precision (mAP@0.50) by 2.5％ compared to the YOLOv11n baseline model, while reducing parameter count and computational load by 17.2％ and 18.2％, respectively. Against nine mainstream models—including RT-DETR (Real-Time Detection Transformer), YOLOv3tiny, and YOLOv5n—YOLO-MPW achieved the best performance in terms of accuracy, parameter count, and computational load, as well as lower missed-detection rates under challenging conditions such as overlapping, occlusion, and flipping, with more consistent attention to target edges and irregular shapes.

Conclusion The YOLO-MPW model achieves a coordinated breakthrough in “high accuracy + ultra-lightweight” recognition of pipeline weld radiograph marks. It supports real-time on-site detection, offers a replicable technical solution for digital pipeline integrity management, and demonstrates robust engineering applicability across industrial imaging scenarios such as oil and gas stations, refining and chemical plants, and ship weld inspections.