AG-Unet: a robust laser stripe extraction method in line-laser imaging

In laser triangulation measurement, high reflectivity and complex exposure environments can cause defects such as highlights, artifacts, and noise in the extracted laser stripes, affecting the accuracy of stripe center extraction. To address this problem, a U-net network structure integrated with additive attention gates is proposed. This structure uses contextual semantics to train the model to predict the weights of effective regions of the stripes, implicitly learning to suppress irrelevant areas in the stripe image. Additionally, an improved Steger algorithm is proposed, which utilizes the grayscale centroid method to filter out invalid center points in the stripe direction. Experimental results show that, compared to traditional small-sample fully convolutional networks, the attention U-net achieves higher accuracy in extracting the stripe centerline, with a mean squared error (MSE) of only 5.3544 pixel, a peak signal-to-noise ratio (PSNR) of 40.8437 dB, and a structural similarity index (SSIM) of 0.9801%. At the same time, the improved Steger algorithm effectively corrects extraction deviations at the edges of the stripe.