DGICR-Net: Dual-Graph Interactive Consistency Reasoning Network for Weld Defect Recognition With Limited Labeled Samples

Weld defect recognition has become a major research topic as it ensures the security of industrial equipment. Existing methods for this task rely on abundant labeled X-ray images, which, however, are not always available due to costly annotation and environmental complexity, leading to undesirable recognition performance. To solve the above issues, a dual-graph interactive consistency reasoning network (DGICR-Net) is proposed in this article. First, a data acquisition approach completed by X-ray radiographic inspection, film scanning, and X-ray image annotation is presented, so as to obtain weld X-ray images. Second, a dual-graph interactive consistent reasoning network is proposed, where the instance graph and the distribution graph are constructed to capture the global correlation information between multiple X-ray images, and then the two graphs are continuously updated in an interactive and consistent reasoning way, thus the label information can be aggregated from limited labeled X-ray images to unlabeled X-ray images. Finally, three losses containing the instance graph loss, the distribution graph loss, and the structure consistency constraint loss are designed to completely train the overall DGICR-Net, enabling achieving weld defect recognition with limited labeled X-ray images. Three experiments are conducted on the datasets pipeline weld dataset (PPL), the metal weld dataset (MTL), and the public weld dataset (GDXray), and the results demonstrate that the proposed DGICR-Net can achieve superior recognition performance than other state-of-the-art (SOTA) methods for limited labeled X-ray images.