X-ray Digital Image Weld Defect Detection Method Based on Improved YOLOv7 Network and Transfer Learning

Abstract

The traditional approach to detecting weld defects in industrial films requires sophisticated equipment and extensive manual labor. Not only is this process time-consuming, but it also poses risks to human vision, and the results are highly dependent on the operator’s level of technical expertise. To address this issue, an intelligent detection method based on X-ray digital image characteristics, data preprocessing, and optimized YOLOv7 network configurations was developed. A proprietary dataset of X-ray digital images containing weld defects was constructed, covering five common flaw types: porosity, slag inclusion, and lack of fusion, among others. In the multi-classification task of weld defect recognition, the trained model achieved the expected performance on the test set, demonstrating its ability to efficiently detect and classify weld defects. Under small-sample conditions, the YOLOv7 model delivered satisfactory results, with an average accuracy of 87.9% and a recall rate of 76.5%. The experimental findings confirm the research value of the self-collected and labeled weld defect dataset, while also offering an intelligent and efficient inspection solution that significantly reduces the workload of inspection personnel.