Fig. 1. Fundamental principle of the proposed network architecture.

Fig. 2. Schematic representation of the proposed CHMFNet.

Fig. 3. (a) Architecture of the HCLNA module for local and non-local feature integration. (b) Architecture of the HCPLF module for complex feature enhancement.

Fig. 4. Comparative analysis of numerical reconstruction results from Holo-encoder, HoloNet, CCNN, and CHMFNet.

Fig. 5. (a) Average PSNR and SSIM values of all four algorithms. (b) Average MSE of the validation dataset during training.

Fig. 6. (a) Detailed architectural parameters of each evaluated model. (b) Average computation time per 100 CGH generation cycles using the DIV2K training dataset. (c) Comparative computation time metrics for 100 CGH generation cycles with the Flickr2K dataset training. (d) Scalability analysis showing CHMFNet’s computation time across varying dataset sizes (N = 100–400).

Fig. 7. (a) Schematic representation of the experimental setup. (b) Photograph of the implemented optical configuration.

Fig. 8. Optical reconstruction results of CGH rendered by different models. (a) The CHMFNet method, (b) the CCNN method, (c) the HoloNet method, and (d) the Holo-encoder method.