IDWS-Net A Robust Swin Transformer Architecture with Hierarchical Attention for Internal Weld Defect Segmentation

Ultrasonic testing (UT) technology, owing to its advantages in defect detection efficiency and imaging quality, enables quantitative characterization of defects through ultrasonic images and plays a critical role in weld defect assessment. However, constrained by the presence of boundary blurring and high-intensity acoustic artifacts, existing segmentation methods struggle to meet the precision requirements necessary for quality analysis. To address these limitations, this study proposes a robust Swin Transformer architecture with hierarchical attention for internal weld defect segmentation (IDWS-Net). Two modules are introduced: the edge-aware enhancement module (EAEM), which improves boundary accuracy by hierarchically integrating multiscale edge features and contour attention mechanisms, and the multiscale context-aware gating (MCAG) module, which suppresses false defects caused by noise by utilizing global–local feature contrast. Experiments were conducted on 4000 collected and labeled B-scan images. Compared with other classic network structures, IDWS-Net demonstrated excellent segmentation performance across various defect types and sizes.