Laser welding is a key joining process for large complex curved ship components, where precise real-time weld defect detection is critical to hull structural safety. However, current lightweight detection models struggle to simultaneously achieve high accuracy for tiny defects and high inference speed under industrial conditions, due to challenges such as low-contrast imaging, varying surface reflections from curved geometries, and the multi-scale nature of defect features. To address these issues, this paper proposes an improved YOLOv7-tiny model named IYOLOv7-tiny-Weld (Improved YOLOv7-tiny for Weld Defect Detection). A Decoupled Self-Attention (DSA) module is embedded to enhance local defect feature extraction, and Group Shuffle Convolution (GSConv) is adopted to offset the extra computational overhead introduced by DSA. The model further incorporates a Weighted Bidirectional Feature Pyramid Network (BiFPN) and Content-Aware Reassembly of Features (CARAFE) to achieve superior multi-scale feature fusion and reconstruction. In addition, Decoupled Detection Heads (DDH) are utilized to mitigate the conflicts between classification and regression tasks. Experiments are conducted on a self-built ship laser welding dataset containing 3,200 images, as well as on the public WELD dataset. On the self-built ship laser welding dataset, the model attains a median detection confidence of 89.75% and a median IoU of 82.16%. The model achieves AP values of 92.56%, 89.14%, and 85.43% for porosity, undercut, and lack of fusion, respectively. Under low illumination and long-distance conditions, the precision drops by only 4.17% and 3.41%, outperforming other compared models. Compared with the baseline YOLOv7-tiny, the proposed model achieves a 12.62 percentage point improvement in mAP@0.5 on the public WELD dataset. The modified YOLOv7-tiny achieves simultaneous advances in accuracy, inference speed, and environmental robustness, offering an efficient lightweight solution for real-time online ship laser welding quality monitoring.
