Aiming to address the issues of limited robustness and low accuracy in traditional laser stripe centerline extraction algorithms, a novel method for centerline extraction is proposed based on improved U2-Net network. Firstly, in order to perform precise pixel-level segmentation and ef-fectively remove noise in the image, the transformer-self-attention (TSA) and transformer-cross-attention (TCA) modules in the network model are introduced. Secondly, according to the application scenario, the Steger method is improved to obtain high-precision extraction of the laser stripe centerline. Finally, the reliability evaluation method is applied to assess the preci-sion of the extracted center point. The experimental results show that the reliability of the pixel center point extracted by our proposed method is 1.9 times higher than that of Steger's algorithm, with higher extraction accuracy and better anti-noise performance, meeting the demand for high-precision industrial measurement applications.