Super-resolution quantitative phase imaging of out-of-focus images based on deep learning

Quantitative phase information which can reflect the internal structure and refractive index distribution of the object is able to be obtained by diffractive and interferometry techniques. However, the phase resolution achieved by the diffraction method is lower than that of interferometry method; while the setup for interferometry method is more complex. To obtain high-resolution phase images without reference beam path, we propose an end-to-end DL based super resolved quantitative phase imaging method (AF-SRQPI) based on generative adversarial network (GAN) to transform low-resolution amplitude images into super-resolved phase images. Meanwhile, considering the inevitable out-focusing during the long hours of observing, autofocusing function is also included by the network. In the training process, out-of-focus low-resolution amplitude images are used as the inputs and corresponding super-resolved phase images obtained by structured illumination digital holographic microscopy (SI-DHM) are used as the ground truth labels. The well-trained network can reconstruct the high-resolution phase image at high speed (20fps) from a single-shot out-of-focus amplitude image. Comparing with other DL-based reconstruction schemes, the proposed method can perform autofocusing and superresolution phase imaging simultaneously. The simulation results verify that the high-resolution quantitative phase images of different biological samples can be reconstructed by using AF-SRQPI .