Research on an Improved YOLO-Based Algorithm for Underwater Weld Seam Detection

To address blurred boundaries, pronounced texture variations, and strong background interference in images of weld seams produced by local dry underwater MIG welding after cooling, this study proposes an improved YOLO11 weld seam detection model, termed YOLO-UWI. First, a high-resolution P2 detection branch is introduced into the neck to preserve weld-seam edges and local spatial details. Second, C3k2-Strip modules are incorporated at the P3 and P4 stages of the backbone to enhance extraction of elongated morphology, continuous texture, and directional features. Finally, an Efficient Channel Attention (ECA) mechanism is added to the P2 detection branch to adaptively recalibrate channel weights, emphasize weld-related features, and suppress background interference. Ablation and comparative experiments are conducted using the original YOLO11 as the baseline. The improved model achieves Precision, Recall, and mAP@0.5 values of 82.2%, 64.1%, and 74.3%, respectively, improving by 0.6, 6.5, and 4.0 percentage points over the original YOLO11. Compared with YOLOv8, YOLOv10, YOLO12, and YOLO26, YOLO-UWI also achieves higher overall detection accuracy. The results demonstrate that the proposed model effectively improves detection accuracy and localization for weld seams produced by local dry underwater MIG welding and provides a reliable technical solution for weld-quality inspection.