Phase Retrieval for Fourier THz Imaging with Physics-Informed Deep Learning

With the continuous progress of terahertz (THz) generating and sensing technology, THz imaging is promising for a wide application range. However, 3D scene reconstruction at THz bands faces with challenges due to the difficulty of accurate phase data recording. In this paper, we demonstrate a phase retrieval method for THz Fourier imaging based on a deep learning (DL) algorithm. This scheme incorporates Fraunhofer diffraction as the physical origin of Fourier imaging as prior knowledge. The classification and reconstruction tasks are achieved based on the dataset transplanted from MNIST. The process is composed of two steps. The supervised DL based on the prepared data for classification, and phase reconstruction and the unsupervised DL network (PhysenNet) performing de-noising. The developed DL method achieves 94% accuracy for the classification task and 0.03 mean square error for the reconstruction task, which means that complex and easily-disturbed heterodyne detection can be replaced to a certain extent.