| Summary: | Digital holography is a promising candidate for advanced display, although several obstacles remain, such as the problem of heavy time consumption in the generation of phase-only holograms. Recently, deep-learning-based methods have achieved the real-time generation of holograms while maintaining high image quality. However, the holograms created with deep neural networks can reproduce images only at a specific distance because their target depth is fixed in the training process. This paper suggested and demonstrated a deep neural network that can continuously control the depth of the phase-only hologram. The network takes a target depth and an input image and generates a phase-only hologram. We added a depth embedding block that moves the hologram latent vector depending on the target depth. Thus, we can change the location of the image plane without retraining. The numerical and optical experiments show that the network understands the relationship between the depth and the appearance of the phase-only hologram. As a result, phase-only holograms generated with the proposed network can reconstruct images with around 25-dB PSNR.
|
|---|