Automatic seam detection of welding robots using deep learning

Welding robots are employed to improve the welding efficiency and quality of steel structures. However, the complexity and diversity of weldments hinder the ability to detect weld seams. To address this limitation, this paper presents a vision-based model that uses a deep learning network combined with the symbol-patching method to plan welding trajectories for welding robots. Semantic straight lines are detected by the stacked hourglass network, and welding paths are determined by assistant symbols with geometric information. Additionally, the image stitching algorithm is used to obtain a broad view of the seam image for subsequent welding processes. The proposed method achieves 90.6% of recall under different lighting conditions. Furthermore, comparative experimental results indicate that the proposed method is robust and accurate for seam detection and localization.