Physics-informed generative network for phase retrieval from a single intensity measurement

This study presents a deep learning algorithm for solving the inverse problem of phase retrieval in the optical near-field (Fresnel) regime using a single intensity measurement (in-line hologram). The algorithm was developed in a self-learning fashion based on generative adversarial networks (GANs) for tomographic reconstruction (GANrec). In this paper, the original GANrec algorithm is adapted to solve the phase retrieval problem. From a single hologram, GANrec can recover both phase and amplitude of the unpropagated exit wave field. Compared with other state-of-the-art deep learning approaches for phase retrieval, it does not need any training data. GANrec combines the advantage of an optimization-based reconstruction with the ideas of machine learning and generative models. It iteratively minimizes the difference between the measured hologram and generated hologram. The results of the validation show good quality of the phase recovery. The idea can work universally for many ill-posed inverse problems if the forward model is well defined.