Multi-depth hologram generation from two-dimensional images by deep learning

Abstract

Recently, proposals have been made to use deep learning for hologram calculations to directly infer holograms from three-dimensional (3D) data. However, this approach is expensive because it requires capturing depth information using an RGB-D camera for inference. In this study, we propose a novel approach that can infer 3D holograms directly from a color two-dimensional (2D) image without requiring depth information, using deep learning. The proposed scheme comprises three deep neural networks (DNNs). The first DNN predicts the depth information from the 2D images, the second DNN generates holograms using the 2D image and the inferred depth information, and the third DNN optimizes the quality of the holograms generated by the second CNN. The inference speed was superior to a state-of-the-art graphics processing unit. We prepared a training dataset comprising pairs of holograms and 2D images. The holograms are generated from the RGB-D image using a layer-based hologram calculation. One significant benefit of our proposed approach is that the reproduced image of the final hologram contains a natural depth cue, i.e., it can represent a natural 3D reproduced image in the depth direction. In addition, conventional image sensors can be used to create input information for inference.