Fourier ptychographic microscopy reconstruction with an untrained deep neural network

Fourier ptychographic microscopy (FPM) is a developed microscopic imaging technique which can break the limitation of the space-bandwidth-product and realize wide field of view, high-resolution and quantitative phase imaging. In recent years, application of deep learning in FPM effectively improved the imaging resolution of the amplitude and phase. However, these methods always require the construction of enormous datasets to train the network. To overcome the above problems, an untrained deep neural network (DNN) based on the physical model of FPM is proposed. Different from the traditional deep learning methods, this method aims to discard thousands of label data. According to the known frequency shift, the output images in DNN are virtually imaged with the forward model of FPM and a series of low-resolution images that have been simply fused are calculated to update the network by comparing errors between them and the experimental data. For a small reconstruction task, the proposed network can be treated as the iterative phase retrieval procedure, the amplitude and phase can be well retrieved with untrained parameters. The simulation results verify the feasibility of proposed physics-driven DNN. Compared with traditional deep learning methods, this method discard thousands of label data in the case of limited resolution loss.