To address challenges in welding defect detection such as multi-scale variations, complex background interference, and low accuracy, this paper proposes a lightweight object detection algorithm called SFL-YOLO, based on the YOLOv11n framework. The model includes three key improvements: (1) a lightweight backbone network, StarNet, to enhance feature extraction and fusion; (2) a new C3k2–FasterFusion module that replaces the original C3k2, effectively reducing background noise and improving contextual feature representation; and (3) a parameter-sharing LeanHead detection head to cut computational redundancy and improve inference efficiency. We conducted extensive ablation and comparison experiments on a self-built welding defect dataset. The results show that SFL-YOLO achieves a mean Average Precision (mAP@0.5) of 94.96% while lowering computational complexity from 6.3 GFLOPs to 3.4 GFLOPs, a 46% reduction. Despite the lower computation, the model maintains a real-time inference speed of 100.64 FPS, well above the industrial requirement of 30 FPS. SFL-YOLO outperforms other mainstream lightweight models such as YOLOv8n, YOLOv9t, and RT-DETR in both parameter count and computational cost, while preserving industrial-level detection accuracy. These results demonstrate the proposed algorithm’s suitability for high-precision, real-time welding defect detection in resource-limited environments. Future work will explore model pruning and knowledge distillation techniques to further optimize the model for broader industrial applications.