Diffraction model-driven neural networks trained using multiscale frequency-domain loss for layer-based high-quality computer-generated holography

Deep learning provides an efficient and feasible solution for computer-generated holography (CGH), and learning-based CGH shows great potential in realizing real-time and high-quality holographic display. However, due to the difficulty of convolutional neural networks (CNNs) in learning cross-domain tasks, most existing learning-based algorithms still struggle to produce high-quality holograms. Here, we propose a diffraction model-driven neural network (M-Holo) that uses multi-scale frequency domain loss to train network parameters to produce high-quality phase-only holograms(POHs). M-Holo embedded multi-receptive-field (MRF) modules into complex amplitude-generating network encoders designed to improve the receptive field of neural network. In addition, the multi-scale frequency domain loss (MSFL) is also increased in the training process of M-Holo, and the abstract feature of multiple levels of the target image are learned in the frequency domain, which further restricts the spatial domain loss insensitive information. The generalization effect of M-Holo is verified by numerical simulation and optical experiment of grayscale and 3D images. M-Holo can effectively improve the quality of reconstructed images and suppress image artifacts.