WFD-YOLO: A Hybrid YOLO Architecture with Frequency-Domain Guidance for Weld Defect Segmentation

Precise segmentation of weld defects offers clearer advantages over simple localization in the modern manufacturing, which can improve reliability in high-density weld zones. In order to improve the segmentation mean Average Precision (mAP) and inference speed, we propose a hybrid WFD-YOLO that employs a wavelet-based frequency down-sampling (WFD) module, a lightweight channel-thresholding attention (CTA), and a dedicated P2 small-object layer for weld defect segmentation, where the WFD module is used for suppressing aliasing while preserving low-frequency structural details, the CTA module is used for reducing the impact of background and noise during defect segmentation, and the dedicated P2 small-object layer is used for giving explicit sensitivity to minor defects like porosity and spatters. The upgraded model improves precision by 3.5%, recall by 7.8%, mAP@0.5 by 7.3%, and mAP@0.5–0.95 by 2.7% over the original YOLO11n-seg, while achieving an inference speed of 303 FPS. The segmentation mAP for porosity and spatters, which represent the most challenging defect categories, is improved by 16% and 15.8%, respectively. These performance gains position the hybrid WFD-YOLO network as an industry-deployable tool for safety-critical weld inspection, compatible with high-speed automated welding production lines.