A centerline extraction method for laser stripes with geometric discontinuous profiles based on physics-conditioned generative adversarial networks

To address the accuracy degradation in centerline extraction for laser stripes with geometric discontinuities, an improved centerline extraction method integrating both contour segmentation and geometric decomposition based on physics-conditioned generative adversarial network (pc-GAN) was proposed. First, the causes of accuracy degradation at curvature discontinuities were analyzed, with an intensity distribution model and a mixed noise distribution model established to characterize the laser stripes. Secondly, the Gaussian-Poisson mixed noise distribution was taken as the constraint for pc-GAN to segment laser stripes with complex profiles, further decompose the laser stripes into geometric primitives according to the predicted inflection point probability map. Then, a data simulation algorithm was developed based on the imaging mechanism of line structured light, combined with real data, a semi-physical dataset was constructed. Finally, experimental results demonstrated that the proposed network achieved a 2.42% increase in IoU and a 0.037 reduction in RMSE compared with existing networks, which improved the reconstruction accuracy of complex profiles by a factor of 3.43. And the peak to valley standard deviation of 2.67 μm when applied in PCB inspection.