Detection and Recognition of the Seam for Gap-Free Butt Joints Based on Scheimpflug Imaging and Deep Learning

Due to the limited camera depth of field, intense arc interference, and occlusion by fixed weld points, the accurate detection and recognition of the seam for gap-free butt joints confronts substantial technical difficulties. These difficulties are further exacerbated by the limited generalization capability of traditional image processing methods under such noisy and occluded conditions. In response, a vision sensor was designed based on the Scheimpflug imaging principle, utilizing a tilted lens structure to extend the range of clear imaging. The camera calibration process was simplified by modeling the Scheimpflug transformation as specific distortion parameters. An improved lightweight YOLOv3 network was developed, employing a MobileNetV1 backbone combined with an improved Bi-FPN structure, significantly enhancing inference speed while maintaining detection accuracy. Training results demonstrate that the proposed method achieves a mean Average Precision (mAP) of 96.74% for detecting weld seams and fixed points with an inference speed of 63.45 FPS. In the context of thin-sheet, gap-free butt welding with tack weld interference, the proposed method exhibits satisfactory robustness and practical utility. Furthermore, a trajectory fitting algorithm for the weld seam center was developed based on the grayscale distribution characteristics within the detection bounding boxes, enabling effective marking of the weld trajectory and thereby laying the foundation for weld seam tracking.