Digital holographic reconstruction based on deep learning framework with unpaired data

Abstract

Convolutional neural network (CNN) has great potentials in holographic reconstruction. Although excellent results can be achieved by using this technique, the number of training and label data must be the same and strict paired relationship is required. Here, we present a new end-to-end learning-based framework to reconstruct noise-free images in absence of any paired training data and prior knowledge of object real distribution. The algorithm uses the cycle consistency loss and generative adversarial network to implement unpaired training method. It is demonstrated by the experiments that high accuracy reconstruction images can be obtained by using unpaired training and label data. Moreover, the unpaired feature of the algorithm makes the system robust to displacement aberration and defocusing effect.