The integration of neural network and physical reconstruction model for Fourier ptychographic microscopy

Abstract

Fourier ptychographic microscopy (FPM) is a recently developed computational imaging method for high resolution and wide field-of-view. In this paper, we propose a new FPM method based on an untrained neural network (FuNN), which integrates a physical reconstruction model for FPM into a convolutional neural network, to solve the FPM reconstruction problem. The weights and bias of FuNN are optimized based on the interaction of the network and the physical reconstruction model so that the requirement for thousands of training datasets is eliminated. To reduce the size of the network, a set of low-resolution images are pre-processed to synthesize a complex object field before fed as the input to FuNN. Pupil recovery is also included and the final input of FuNN is a 3D image tensor that consists of synthesized field and coherent transfer function. The output of FuNN is modeled as reconstructed high-resolution complex object field and pupil function. The function of the pupil recovery is achieved mainly by the physical reconstruction model, so that the main structure of the neural network does not need to be adjusted according to the pupil recovery process as other iterative-based FPM neural network does. The performance of FuNN is evaluated by both simulation and experiment. Compared with other iterative-based FPM neural network, which just utilize network to calculate the gradient, FuNN can improve the quality of reconstructed complex object field without a large number of training datasets.