Single-frame noisy interferogram phase retrieval using an end-to-end deep learning network with physical information constraints

Phase retrieval from a single-frame interferogram is a meaningful approach in dynamic and real-time interferometry. However, existing demodulation methods are susceptible to real noise impressions, resulting in low accuracy. This paper proposes an accurate end-to-end learning-based method for one-shot phase retrieval that doesn't require pre-processing (e.g., denoising, normalization) and post-processing (e.g., phase unwrapping). Compared to existing deep learning methods, we incorporate more physical information constraints, contributing to high accuracy and robustness. A realistic noise model that incorporates the speckle noise and the imaging process is introduced to generate training datasets. Furthermore, a subnetwork is adopted for noise estimation, along with a subnetwork for frequency domain decomposition to extract multi-channel physical information. Additionally, a physical information fusion block is established to integrate relevant data, and a reconstruction network based on NAFNet is implemented for comprehensive end-to-end phase retrieval. We validate the accuracy and robustness of our approach through simulations and experiments. Our results demonstrate that the proposed method achieves a phase retrieval accuracy of 0.01 lambda for noisy interferograms, superior to several existing methods that we have reproduced.