Wrapped phase denoising: a WISE-transformer and comprehensive comparisons

Wrapped phase denoising is a fundamental task in optical metrology, with numerous traditional and deep learning methods developed in the past decades. In this work, we first propose yet another deep learning method called WISE-Transformer, which is a novel U-shaped transformer model that incorporates both windowed and inter-window self-attention to balance local structure preservation and global context integration. Through extensive experiments, we demonstrate that WISE-Transformer achieves state-of-the-art performance, comparing with other popular deep learning models across multiple evaluation metrics. During the experiments, we also verify the effectiveness of using sine and cosine phases as the network input and a dynamic data generation strategy to enhance training diversity. Next, a rigorous and comprehensive comparison is carried out among WISE-Transformer (as a representative of deep learning methods), windowed Fourier transform (WFT) (as a representative of traditional methods), and the theoretical Cramer-Rao lower bound (CRB), to clarify the capabilities, advantages and disadvantages of wrapped phase denoising methods. WISE-Transformer not only convincingly outperforms the WFT, but also interestingly breaks the CRB in simulated settings. While WISE-Transformer maintains competitive performance on real experimental data, it is shown less effective than WFT methods, underscoring the generalization challenge as a core focus of our future research.