Abstract: To address the problems of large scale variation, low contrast, and diverse overlapping of welding defects in complex backgrounds, an improved YOLO11 algorithm for weld surface defect detection was proposed. A feature extraction dilation-wise residual (DWR) module was introduced into the backbone network and fused with the C3k2 module to enhance the information extraction capability of the model. A large separable kernel attention (LSKA) module was introduced to improve the multi-scale feature fusion SPPF module, enabling the model to focus on primary features. The Inner-SIoU loss function was adopted to improve the bounding box regression accuracy. The results indicate that the improved model can detect defects quickly and accurately. Its precision, recall, and P_mAP50 reach 92.5%, 92.2%, and 95.3%, respectively, which are 2.1%, 3.0%, and 3.5% higher than those of the baseline model. The model parameters and floating-point operations are reduced by 0.127 M and 0.3 G, respectively, and the inference speed reaches 293 frames/s. This indicates that the improved model is significantly enhanced with fast inference speed, effectively reducing the false detection and missed detection rates. It performs particularly well in small-sized defect detection, possesses the advantages of both high accuracy and lightweight, and is more suitable for efficient weld defect detection applications in industrial scenarios.