Deep-learning based reconstruction in optical scanning holography

Optical scanning holography (OSH) can be used to record holograms of large three-dimensional (3-D) objects, based on a two-dimensional (2-D) optical scan. For semi-transparent objects, diffraction waves from all the sections can be recorded in the hologram. Numerical reconstruction of a 3-D volumetric image from an optical scanned hologram is a difficult task. The main problems are the intensive computational load, and the heavy blurring of each reconstructed section with the defocused noise from other sections. In this paper, we propose a deep-learning network for high quality image reconstruction from the optical scanned holograms. Within the framework, a U-net structure is adopted to learn the mapping between a collection of holograms, and their reconstructed volumetric images. We use a two-pupil optical heterodyne scanning system to obtain the training data where a five-fold cross validation method is used to prevent from overfitting and produce enough images in the dataset. The deep-learning based OSH can eliminate the defocus noise and generate high quality reconstruction results from an unknown hologram. Our proposed method is significantly faster than conventional OSH reconstruction algorithms, and hence suitable for processing large holograms that are captured by OSH. The feasibility of our approach is demonstrated with numerical simulations and optical experiments. The deep-learning reconstruction method proposed in the present paper is also applicable to other digital holograms obtained from conventional digital holographic systems.