High-precision shear parameters calculation method in lateral shearing interferometry via transfer learning-enhanced TransUNet

Abstract

In lateral shearing interferometry, the discretely distributed edges and weak gradient variations in shearing interferograms can adversely affect the accurate calculation of shear parameters, consequently degrading wavefront reconstruction precision. To address this issue, this study proposes a TransUNet-based transfer learning method for shear parameter calculation. First the model pre-trained with simulated interferograms to learn fringe characteristics, then the model fine-tunes with limited real experimental data to overcome the scarcity of real interferograms. The model can achieve precise four-class segmentation of interferograms (background, two non-interference regions, and interference fringes), building upon the segmentation results, least-squares circle fitting algorithm is applied to simultaneously accomplish shear parameter calculation and wavefront center localization. Experimental validation demonstrates that the proposed method maintains excellent computational precision and automation levels even under noisy and weak-gradient conditions, while the shear parameters calculated by our method achieve significantly higher accuracy in wavefront reconstruction than existing mainstream algorithms.