Fast extraction of weld seam features using a laser stripe contour in line-structured light vision

Stable feature point extraction is essential for realizing intelligent guidance and tracking in robotic welding. However, the presence of intense spatter, arc light, and smoke during the welding process poses significant challenges to the accuracy of feature extraction. To address these issues, a feature point extraction method based on a deep neural network architecture for weld seam image denoising and contour tracking algorithms is proposed. First, a deep neural network enhanced with an attention mechanism is introduced to effectively suppress strong background noise and automatically extract laser stripes from the images. Subsequently, a contour tracking algorithm based on a 16-neighborhood structure is proposed to achieve high-precision extraction of welding seam feature points from the extracted laser stripes. Furthermore, to verify the robustness and adaptability of the proposed contour tracking algorithm in different scenarios, additional experiments are conducted on weld seam images acquired during the pre-welding positioning stage. The experimental results demonstrate that the proposed algorithm achieves high accuracy and rapid performance in weld feature point extraction, with average absolute errors of 0.84 and 0.57 pixels along the <i>X</i>-axis and <i>Y</i>-axis, respectively, and an average processing time of only 0.26 ms.