Fig. 1. (a) Workflow of conventional methods based on the vanilla UNet architecture. (b) Our proposed cross-domain fusion network highlights the role of our multi-stage conversion architecture and the infinity phase mapper (IPM) in accomplishing the cross-domain transformation task. The multi-stage conversion architecture employs multiple feature maps to facilitate a gradual transformation between two domains. Concurrently, the infinity phase mapper is designed to accommodate the periodic nature of phase values, ensuring the preservation of the physical consistency.

Fig. 2. Hologram generation workflow of our proposed method and the corresponding network architecture. The feature maps representing the various conversion stages are color-coded diversely. The first sub-network CDFN1 is used to predict the initial phase and the second sub-network CDFN2 is used to predict POHs.

Fig. 3. Schematic of infinity phase mapper (IPM) M, which maps infinite phase values ϕi to their corresponding points ϕo within the [−π,π] interval.

Fig. 4. Comparison of numerically reconstructed color images. From left to right: results of double phase-amplitude coding (DPAC), stochastic gradient descent method (SGD), HoloNet, CCNN, and our proposed CDFN, respectively (PSNR in dB).

Fig. 5. Comparison of simulated reconstruction images in our ablation study (green channel). From left to right: reconstruction results of HoloNet, multi-stage architecture (MS), multi-stage architecture with infinity phase mapper (MS w IPM), multi-stage architecture with deep supervision (MS w DS), multi-stage architecture with infinity phase mapper and deep supervision (MS w IPM&DS) (PSNR in dB).

Fig. 6. Normalized phase value distribution of a randomly picked POH. From left to right: results of HoloNet, CDFN without infinity phase mapper (CDFN w/o IPM), CDFN with infinity phase mapper (CDFN w IPM). Note that we shift the value in [−π,π] to [0,2π] in this figure for the easier observation of values around π. The arrows highlight that the IPM can generate a continuous phase value distribution around the π value, where the conventional active function generates a gap.

Fig. 7. Holographic display setup. (a) Schematic diagram of the optical display system setup. (b) Photograph of the optical display system setup.

Fig. 8. Optical reconstruction images in the green channel. From left to right: results of Gerchberg–Saxton algorithm (GS), double phase-amplitude coding (DPAC), HoloNet, CCNN, and CDFN, respectively.

Fig. 9. Optical reconstruction images from 1920×1072 pixels POHs on a 1920×1080 SLM in red, green, blue, and color channels. The images are directly captured by a camera. The color image is obtained by synchronizing the three-color laser source and sequentially loading different POHs. From left to right: results in red, green, blue, and color channels, respectively.

Fig. 10. Optical reconstruction images of our method in color channels directly captured by a camera and the corresponding phase-only holograms. (a) Cropped-zoomed patch of the first color image in (b) for visualization. (b) Captured color images. (c) The corresponding phase-only holograms of (b).

Fig. 11. Results of our method applied to binary images. (a) Simulated images. (b) Zoom-in patches. (c) Experimental images. Note that target binary images are not in our training set.