Neural-network-based methods in digital and computer-generated holography: a review

Abstract

Subject of study. An overview of modern neural-network-based methods for digital and computer-generated holography is presented. Relevant works on phase and amplitude reconstruction, media characterization, phase unwrapping, computer hologram generation, and related topics are discussed. Aim of study. The study investigates modern neural-network-based methods for digital and computer-generated holography. Method. The methods discussed in this review are based on neural networks and are developed for specific tasks in holography. Training datasets for supervised learning typically contain a set of input images (e.g., digital holograms, wrapped phase, 3D-scenes) and a set of target images (e.g., reconstructed scenes, unwrapped phase, 3D-holograms). The network learns to generate the target images from the input images during training. In unsupervised learning, there is no need to prepare target images; the training is based solely on the input images and the transformations applied to them. Main results. This review provides an overview of the applications of neural networks in holography, focusing on state-of-the-art works. It discusses the neural network architectures most commonly used in holography, organizing cited works based on their particular applications. The review highlights the most significant results and achievements of neural-network-based methods for digital and computer-generated holography. Practical significance. This review may interest researchers specializing in holography or adjacent fields. It familiarizes readers with modern neural-network-based methods used in computer hologram generation and holographic image reconstruction, as well as the specifics of their practical applications. The data presented demonstrate that neural-network-based methods can sometimes offer advantages in speed and/or quality compared with conventional methods.

Topics & Concepts