Line-structured light vision 3D measurement technology is widely used for its high precision and non-contact 3D reconstruction. However, existing methods face calibration coupling issues and are highly sensitive to background noise and lighting changes in complex environments, leading to reduced accuracy in stripe extraction and 3D measurements. To address these challenges, a robust 3D measurement method is proposed, which is based on convolutional neural networks (CNN). First, we design an innovative Residual U-shaped block feature pyramid network (RSU-FPN) to suppress background noise and achieve high-precision extraction of the structured light stripe center. Second, we develop a new line-structured light sensor and introduce a decoupled calibration model that separates camera and light plane calibration, enhancing system flexibility and scalability. Experimental results show that our method achieves high-precision stripe extraction with root mean square errors of 0. 005 mm, 0. 009 mm, and 0. 097 mm in the x, y, and z directions, respectively. It also provides high-precision 3D reconstruction on different surface types, demonstrating its robustness and excellent performance in real-world applications.