Accurate and real-time weld seam recognition is critical for automated welding systems in intelligent manufacturing. However, existing deep learning-based models often suffer from high computational complexity and limited real-time performance, which restrict their deployment in embedded and industrial environments. To address these challenges, this paper proposes a lightweight weld seam segmentation framework based on an optimized SGD-YOLO (Segmentation-guided Ghost Dynamic YOLO) architecture, aiming to achieve a favorable balance between accuracy and efficiency. By redesigning the network structure and enhancing feature extraction capability, the proposed model significantly reduces computational cost while maintaining high detection precision. Experiments demonstrate that the proposed method achieves a 36.5% reduction in floating-point operations and a 29.4% decrease in parameter size compared with conventional models, enabling stable real-time performance under industrial conditions. Furthermore, feature point extraction experiments show that the pixel localization error is controlled within 5 pixels and the mean depth error remains below 0.5 mm, indicating high robustness and measurement accuracy. These results confirm the effectiveness of the proposed framework in precise weld seam perception and geometric feature extraction. Overall, the proposed lightweight weld seam segmentation approach provides a practical and efficient solution for real-time welding automation, promoting the broader application of deep learning techniques in intelligent manufacturing and industrial robotics.
Li et al. (Thu,) studied this question.