TNUnet: U-Shaped Phase Feature Extraction Network for Phase Unwrapping in Optical Metrology

Due to the bias of convolution operation, current 2-D spatial phase unwrapping (PU) methods in optical metrology based on convolutional neural networks (CNNs) struggle to accurately capture the global context and fail to model long-range capabilities effectively. The Transformer-based PU method suffers from deep degradation due to its over-reliance on stacked layers for information interaction. Most two-step PU methods based on segmentation models tend to restore to the original resolution directly after downsampling and extracting fine features, resulting in significant loss of depth features. The unwrapping accuracy in many optical measurement scenarios has been compromised by these factors. This article introduces a new U-shaped network called TNUnet to tackle these challenges. TNUnet leverages TransNeXt as the fundamental feature learning component, effectively integrating the strengths of aggregated attention (AA) and the convolutional gated linear unit (GLU). This compensates for the limitations of CNNs in capturing long-range dependencies and mitigates the deep degradation of the Transformer. TNUnet, a robust phase feature extraction network, is ideal for regression and segmentation models. Extensive experiments demonstrate that our TNUnet achieves state-of-the-art performance on both models. When the wrapped phase is disturbed by noise or discontinuity, the unwrapping accuracy of the regression model-based TNUnet exceeds 92%, and its FLOPs are nearly 67% lower than the most competitive method, U2-Net. The parameters and FLOPs of the segmentation model-based TNUnet are reduced by nearly 88% and 83% compared to TransUNet, respectively, while achieving an ultrahigh unwrapping accuracy of 94%. The code is publicly available at https://github.com/zzi-heng/TNUnet-PU.