Real-time multi-depth holographic display using complex-valued neural network

Abstract

Computer-generated holography (CGH) has made significant advancements and is considered a leading approach for near-eye 3D displays. Recent learning-based CGH methods address the time-quality trade-off of traditional approaches but often face challenges related to efficiency and computational demands, especially with real-valued networks in multi-depth settings. To overcome these issues, this study proposes a residual block-based complex-valued convolutional neural network (ResC-CNN) structure, integrated into a symmetric dual-network framework driven by a diffraction model, for real-time generation of multi-depth holographic displays. This approach enhances the network's ability to handle complex domain calculations in CGH, making the learning process more efficient. A layered depth image (LDI) dataset is also incorporated to improve scene information prediction accuracy. Numerical and optical experiment results indicate that our proposed framework significantly increases the real-time generation frame rate of holograms and enhances the fidelity of displayed details, offering a practical solution for high-quality, real-time multi-depth holographic displays in applications such as augmented reality.