Image feature acquisition and identification of multi-layer and multi-pass weld seams based on region growing segmentation combined with deep learning

To address the significant challenges associated with the multi-layer and multi-pass automated welding of medium-thick plate, including uneven laser stripe brightness, local high reflection, serious noise, and complex and variable weld contour features, a weld seam feature extraction strategy of region growing image segmentation based on block division combined with the YOLOv8 deep learning algorithm was proposed for weld seam feature extraction based on structured light vision sensing. A visual sensing system with a single-stripe laser and high-resolution camera was established and utilized to collect multi-layer and multi-pass weld seam images of thin-wire submerged arc welding. Region growing segmentation based on adaptive threshold and block division, convolutional filtering, and segment fitting treatment were then applied to each layer of weld seam images to extract the feature points of each layer of weld seam. The YOLOv8 algorithm was employed to extract the pixel coordinates of the feature points within the designated area of the weld seam images, thereby constructing the welding path planning for robots. The results demonstrate that the algorithm of region growing segmentation based on adaptive thresholding and block division is capable of fully extracting the feature points of laser stripe images and multi-layer weld beads. The overall root mean square error (RMSE) of the deep learning model is 0.055 1 mm, and the identification accuracy of the feature point is 98.41%. The model demonstrates high algorithm accuracy and robustness, effectively satisfying the requirements of weld seam detection. The welding tests confirm that the morphology of a multi-layer and multi-pass weld seam with sound formation is obtained by employing this proposed algorithm, providing an innovative solution for weld seam detection in multi-layer and multi-pass welding of medium-thick plates.