Quantitative pure-phase object reconstruction under single-shot Fourier measurement via deep learning

Abstract

The phase retrieval problem, usually known as reconstructing an estimate of phase object from its diffraction intensity patterns, exists widely in the optical imaging field. In this paper, a deep convolutional neural network (CNN) framework is developed to accomplish the phase reconstruction given a single-shot Fourier measurement, without the requirement of prior knowledge (like nonnegativity and support constraints) or time-consuming iterations. A receptive field and input enhance learning-based design principle is applied to establish the neural network architecture. And a new loss function that combines the structural similarity (SSIM) index and the mean square error (MSE) is taken to further improve the network training. Quantitative (the average SSIM and relative MSE are ~ 0.9900 and 0.0012, respectively) and qualitative (SSIM, ranging from 0.87 to 0.95) phase reconstructions are achieved in the numerical simulation and experiment, respectively. Once the proposed network is well trained, it is convenient for the users to perform the phase reconstruction at a remarkably fast speed (~ 0.425 s). Moreover, our network is considerably robust to different noises and could be trained to additionally function as a denoiser in the phase retrieval process. In practice, the proposed method is able to build an accurate propagation model of the practical imaging system, and significantly enhance and improve the spatial resolution. With the great convenience and high accuracy of phase recovery, the proposed deep CNN framework may be helpful to solve the phase retrieval problem in Fourier imaging systems, like coherent diffraction imaging (CDI), X-ray imaging, and imaging through diffusers, etc.