Welding Defect Detection Using X-Ray Images Based on Deep Segmentation Network

Automated detection of welding defects in X-ray images is critical for ensuring the quality of welded structures. However, accurately locating welding defects remains challenging because of various complex factors, including low contrast, weak texture, and the limitation of small-scale datasets. In this research, we developed a framework that combines deep convolutional neural network and data augmentation techniques for semantic segmentation of welding defects with high accuracy. Specifically, we develop a modified U-Net network that can effectively and efficiently identify defects in X-ray images of weld seams. The proposed network is trained using a small-scale dataset that is augmented online to improve its robustness and generalization ability. Experiments on the public dataset GDXray demonstrate that the proposed method performs better than other models, with an F1-score of 0.85 and a mIoU of 0.75. Additionally, the proposed model’s simple structure allows for faster referencing and hardware space savings, making it suitable for practical industrial applications.