Efficient sub-pixel laser centerline extraction via an improved U-Net for structured light measurement

Laser line scanning, as a non-contact 3D measurement technique, has been widely employed in industrial inspection and 3D reconstruction, where the accuracy of stripe center extraction directly affects measurement precision. This paper presents a deep learning-based method for laser stripe center extraction, aiming to improve both the accuracy and efficiency of structured light measurement systems. The proposed method builds upon the U-Net architecture, introducing depthwise separable convolutions in the encoder, which significantly reduce computational cost while preserving the spatial resolution required for narrow and continuous stripe structures. In the decoder, attention mechanisms are introduced to emphasize informative spatial regions, improving feature discrimination in cluttered or low-contrast backgrounds, while multilayer perceptron modules are incorporated to improve multi-scale feature fusion and improve stripe continuity near endpoints. Moreover, a center localization loss function is designed by integrating geometric deviation and count-based constraints, effectively guiding the network to focus on stripe center regions and enhancing segmentation performance near stripe endpoints. In the post-processing phase, polynomial fitting and moving average smoothing are applied to further improve the sub-pixel accuracy of the extracted centerline coordinates. Experimental results demonstrate that the proposed method outperforms several state-of-the-art deep learning models in terms of stripe segmentation accuracy and inference speed. Furthermore, compared to the classical Steger algorithm, our method achieves significantly higher inference efficiency while maintaining superior localization accuracy, validating its robustness under various simulated industrial noise conditions and its potential for real-world deployment in real-world industrial laser stripe extraction tasks.