A novel method for extracting weld seam by fusing semantic segmentation and point cloud features

Abstract

Manual weld inspection suffers from low efficiency and difficulties in quantification, while 3D point cloud-based detection often encounters data redundancy and high computational costs. To address these issues, this paper proposes an automated weld seam extraction method that synergistically integrates 2D semantic segmentation with 3D point cloud feature processing. The process begins by employing an efficient, lightweight YOLOv11-Seg model to perform semantic segmentation on RGB images of welded workpieces, thereby locating the weld region of interest (ROI). The initial segmentation mask is then compensated and optimized using morphological operations to ensure complete coverage of the weld area. Subsequently, a single-mask dual-region phase-constraint strategy is introduced to restrict the point cloud computation strictly to the weld seam ROI, effectively suppressing background point cloud generation at the source and significantly improving reconstruction efficiency. Finally, an enhanced RANSAC algorithm, termed Normal-Guided and Dual-Threshold RANSAC (NGDT-RANSAC), which incorporates normal guidance and a dual-threshold inlier determination mechanism, is developed to achieve precise separation between the base-metal planes and the weld point cloud. Experimental results demonstrate that the proposed method achieves efficient and accurate extraction and segmentation of weld point clouds, with precision and Intersection over Union (IoU) reaching 98.5% and 97.7%, respectively, and an average edge segmentation error of 0.2 mm. The entire process, from acquisition to segmentation of point clouds comprising tens of thousands of points, is completed within 4 s, confirming its high efficiency and reliability for practical engineering applications.