A Multistage Defect Detection and Classification Model for Girth Weld Radiographs in Oil and Gas Pipelines

Radiographs of girth welds in oil and gas pipelines exhibit substantial variations in imaging conditions, while small defects are prone to missed detection and visually similar defect classes remain difficult to distinguish. Existing studies often focus on a single detection or classification task and rarely address the integrated multistage processing of full radiographs. To address these issues, this study proposes a multistage intelligent defect detection method consisting of weld region localization, defect instance detection, and defect type classification. YOLOv10 (You Only Look Once, version 10), an end-to-end object detection model, is first employed to localize and extract the weld region from each radiograph. Overlapping image slicing, bounding box fusion, and interference region filtering are then incorporated to improve defect localization, particularly for small defects and those with low contrast. To enhance the classification of highly similar defect classes, knowledge distillation with adaptive class temperature (KD-ACT) is introduced to dynamically adjust the distillation temperature according to the classification difficulty of each class. A dataset comprising 4396 digitized X-ray radiographs collected from engineering practice was used for multistage evaluation, with annotation review and result verification performed by expert radiographic interpreters. The proposed method achieved a weld region detection rate of 99.67%, a defect detection rate of 98.6%, and end-to-end recall of 91.8% across nine defect classes.