Abstract
Line structured light scanning is widely used for the three-dimensional (3D) measurement of precision parts, essentially relying on the optical sensing of laser stripe information by a camera. However, for metal parts, the large curvature and specular reflection of their surfaces can introduce significant noise, reducing the stability of laser stripe perception and thus hindering the accurate extraction of the laser centerline. In order to overcome these challenges, a deep learning-driven algorithm is proposed. An attention mechanism is incorporated into the original U-Net architecture, and the encoder is modified to achieve precise laser stripe segmentation. Subsequently, the region of interest (ROI) is extracted, followed by the application of the Steger algorithm for laser stripe centerline extraction, with the extracted results further refined through non-maximum suppression (NMS), interpolation, and smoothing techniques. The experimental results indicate that, under complex perception environments, a maximum standard deviation (SD) of 0.006 mm, a maximum root mean square error (RMSE) of 0.006 mm, and a maximum absolute percentage error (APE) of 0.069% are achieved by the proposed method. These findings validate the high accuracy and stability of the scheme, especially under high curvature, reflective surfaces, and strong ambient lighting conditions.
Funding Information
  • Key Research and Development Program of Shandong Province, China (2024ZLGX09)
  • Taishan Scholar Program of Shandong Province, China (tstp20230614)
  • Taishan Scholar Program of Shandong Province, China (tsqnz20221124)
  • Young Elite Scientists Sponsorship Program by China Association for Science and Technology (2022QNRC001)
  • National Natural Science Foundation of China (52575647)
  • Municipal Science Foundation of Qingdao (25-1-1-203-zyyd-jch)

This publication has 35 references indexed in Scilit: