Fig. 1. Traditional computer-generated hologram generation method. (a) Direct generation and its optical reconstruction; (b) iterative generation algorithm

Fig. 2. Fundamental process for generating computer-generated hologram based on training deep learning networks

Fig. 3. Classifications of generation algorithms of computer-generated hologram based on deep learning

Fig. 4. Training process and datasets used for Tensor Holography and Holoencoder. (a) Training process of Tensor Holography^[29]; (b) dataset used for Tensor Holography^[29]; (c) training process of Holoencoder^[30]; (d) DIV2K dataset^[31]

Fig. 5. Schematic diagrams of HoloNet and Self-holo algorithms. (a) Schematic diagram of HoloNet algorithm^[32]; (b) schematic diagram of Self-holo algorithm^[33]

Fig. 6. Optimization methods for neural network structures. (a) Increasing the receptive field through Fourier modules^[34]; (b) using complex-valued neural networks^[49]; (c) incorporating dual-phase encoding structures^[51]

Fig. 7. Hologram reconstruction effects generated by neural networks trained with different loss functions and datasets. (a) Effects of multi-scale loss function^[55]; (b) effects of global loss function^[56]; (c) effects with LDI dataset^[59]; (d) depth-of-field supervised training effects^[65]

Fig. 8. Optimization algorithms combined with hardware. (a)(b) Experimental optical path and optimization effects of holograms through camera optimization^[32,66]; (c) time-multiplexed neural holography flowchart^[69]; (d) high dynamic range optimization effects^[71]

Fig. 9. Hologram generation algorithm for real-world scenes based on deep learning. (a) Combining monocular acquisition and depth estimation algorithm^[81]; (b) combining liquid lens camera acquisition algorithm^[88]

Fig. 10. Compression and adjustment of computer-generated hologram. (a) Compression and decompression of computer-generated hologram^[90]; (b) adjustment of reconstruction image position using deep learning-generated computer-generated hologram^[96]