DL-SI-DHM: a deep network generating the high-resolution phase and amplitude images from wide-field images

Abstract

Structured illumination digital holographic microscopy (SI-DHM) is a high-resolution, label-free technique enabling us to image unstained biological samples. SI-DHM has high requirements on the stability of the experimental setup and needs long exposure time. Furthermore, image synthesizing and phase correcting in the reconstruction process are both challenging tasks. We propose a deep-learning-based method called DL-SI-DHM to improve the recording, the reconstruction efficiency and the accuracy of SI-DHM and to provide high-resolution phase imaging. In the training process, high-resolution amplitude and phase images obtained by phase-shifting SI-DHM together with wide-field amplitudes are used as inputs of DL-SI-DHM. The well-trained network can reconstruct both the high-resolution amplitude and phase images from a single wide-field amplitude image. Compared with the traditional SI-DHM, this method significantly shortens the recording time and simplifies the reconstruction process and complex phase correction, and frequency synthesizing are not required anymore. By comparsion, with other learning-based reconstruction schemes, the proposed network has better response to high frequencies. The possibility of using the proposed method for the investigation of different biological samples has been experimentally verified, and the low-noise characteristics were also proved.