A novel feed-forward neural network-based method for fast hologram generation

Abstract

The enormous computing time is a challenge for computer-generated hologram (CGH) calculation in a holographic display. A learning-based method, hologram generation network (HGN), is proposed to accelerate CGH calculation. The method is a unique combination of point-source model and recent deep learning technique, showing how to obtain high-quality CGH quickly. HGN is a feed-forward neural network synthesized from different function blocks. The input of the network is a displacement tensor consisting of point clouds and hologram plane coordinates. The output is a holographic matrix whose column vector represents the coordinates of a single point hologram. The RBF network is then trained and tested by the numerical samples in the bounded field, thus the reconstruction quality of the CGH can be guaranteed strictly. Numerical simulation results show that HGN runs faster than the traditional method with high reconstruction accuracy. The optical experiments are performed to demonstrate its feasibility.