Aiming at the problems of insufficient detection accuracy and susceptibility to missed detection in existing X-ray weld defect detection methods, an improved YOLOv10n-based model named SFE-YOLOv10n is proposed; SPD-Conv spatial-to-depth convolution replaces the conventional convolution in the backbone network to enhance the model"s ability to capture feature information of small target defects in X-ray welds; A lightweight C2f-Faster module is added to replace the C2f module in both the backbone and neck networks, reducing model computational complexity and achieving model lightweighting; The EMA attention mechanism is introduced based on the C2f-Faster module in the neck network to improve weld defect detection accuracy; Experiments were conducted on five typical types of X-ray weld defects, with results showing that the SFE-YOLOv10n model achieved accuracy, recall, mean average precision (mAP), and detection frame rates of 89.5%, 86.2%, 92.8%, and 80.8f/s respectively, representing improvements of 3.6%, 1.5%, 4.3%, and 1.7% over the original YOLOv10n model, while the model parameter count slightly decreased; Compared to existing methods, the SFE-YOLOv10n model accurately detects various X-ray weld defects while maintaining lightweight characteristics, and experimental testing verified that it meets the requirements for X-ray weld defect detection in industrial production。