Physics-driven Deep Neural Network for Fourier Phase Retrieval

Fourier Phase retrieval (PR), aiming at recovering a complex-valued signal from its Fourier intensity measurements, has attracted widespread attention due to its importance in many optical imaging applications. Recently, deep learning-based approaches were developed that achieved some success. These approaches require only a single Fourier intensity measurement without the need to impose any additional constraints on the measured data. However, the quality of the reconstructed images still has much room to improve. Besides, many of these approaches follow the traditional iterative estimation framework and require a lengthy computation process. In this paper, a novel physics-driven multi-scale DNN structure dubbed PPRNet is proposed. Similar to other deep learning-based PR methods, PPRNet requires only a single Fourier intensity measurement. It is physics-driven and has a multi-scale structure such that the network is guided to follow the Fourier intensity measurement at different scales to enhance the reconstruction accuracy. Since the process is non-iterative and the network is end-to-end trained, it is much faster and more accurate than the traditional physics-driven PR approaches. Extensive simulations were conducted on two datasets. The results demonstrate the superiority of the proposed PPRNet over the traditional PR methods.