Dual-wavelength efficient phase imaging method based on convolutional neural networks

Abstract

Traditional dual-wavelength interference techniques often require collecting multiple frames of intensity maps, followed by phase shifting and unwrapping to derive phase information. This method is not only time-consuming but also complex. To address these shortcomings, a high-precision and fast phase recovery method is proposed based on deep learning techniques. This approach leverages a large dataset of interferograms for training, testing based on U-Net. Remarkably, our method predicts phase information from a single frame interferogram. It significantly simplifies the computational steps and enhances a certain degree of generalization ability, as various types of fringe interferograms can be processed through separate training. Simulation tests reveal root mean square errors (RMSEs) of 0.0108 rad, 0.0232 rad, and 0.0465 rad for three different types of interferograms, indicating excellent phase recovery accuracy. Further robustness testing with Gaussian white noise shows minimal changes in RMSE, underscoring the method's stability. Real experimental results confirm the method's feasibility and better computational efficiency, achieving phase information retrieval in just 0.5 s.