A physics-guided memory-enhanced semi-supervised approach for detecting weld defects in radiographic images

Deep learning-based weld defect detection in industrial X-ray imaging is often constrained by scarce annotated data, high labeling costs, and poor generalization to unseen scenarios. To address these challenges, this paper proposes a physics-guided, memory-enhanced semi-supervised defect detection model (PMSDM). The model constructs normal pattern representations from easily accessible defect-free images via a localization memory module, enhances anomaly learning through artificial defects generated based on X-ray imaging principles to better define classification boundaries, and achieves fine-grained classification using a small number of labeled defect samples through a classification memory module. Localization and segmentation of tiny defects are performed via difference-based comparison with the memory bank, further refined by a spatial attention module that fuses multi-scale feature information. Experiments on the self-built dataset and the publicly available GDXray dataset show that PMSDM consistently outperforms traditional and deep learning baselines in terms of Dice, IoU, Accuracy, and mAP, while maintaining strong generalization under domain shifts. PMSDM offers an efficient and scalable solution for weld defect detection in data-scarce industrial environments.