YOLOv11 for Radiographic Non-Destructive Testing of Welded Joints

This study presents the application of the YOLOv11 neural network model in radiographic inspection. Purpose: to substantiate the feasibility of employing YOLOv11 for the automatic detection and visualization of defects in radiographic images of welded joints. Methods: the analysis of contemporary tools and techniques, including neural networks, alongside a representative dataset of approximately 100 annotated images, featuring various defect types and image quality indicators, supplemented by augmentation and pre­filtering of the images. Results: it has been demonstrated that the utilization of an enhanced Backbone with C3k2 modules and a C2PSA spatial attention block can enable the model to achieve high values of mAP@0.5–0.95, reliably detecting both large and low­contrast small defects while maintaining acceptable per­image processing times. Practical significance: the results indicate the potential for integrating YOLO11 into existing non­destructive testing systems for welded joints.