Digital holography is one of the next-generation displays attracting attention, but there are still many obstacles impeding commercialization. Among them, a key problem is that the amount of computation required to generate a hologram is too large, and the computation time is very long. For example, it takes tens of minutes to convert a 1-minute 1K resolution image into a hologram. This makes it quite difficult to reproduce video on a holographic display. Recently, methods for generating real-time holograms while maintaining high quality through deep learning-based methods have been proposed, and they have a speed that is fast enough to process 30 frames per second. However, the hologram generated by the deep neural network was able to reproduce the image only at a specific distance because the target depth was fixed during the learning process. In other words, if it is necessary to adjust the target depth of the holographic image, the deep learning model must be learned from the beginning using the new target depth. Since this is very inefficient, a way of adjusting the target depth without re-learning is required.
In this thesis, a deep neural network that can continuously control the depth of a phase-only hologram is proposed and implemented. The network receives the target depth and the input image and generates a phase-only hologram, which can restore the hologram to excellent quality. A depth embedding block has been added that moves the holographic latent vector according to the target depth, and thus the position of the image plane can be changed without re-learning. It was shown that the proposed network understands the relationship between the depth and shape of the phase-only hologram, and th proposed method was verified through a computer simulation and optical reproduction. The deep learning model received an image and a target depth and generated a corresponding phase-only hologram within 35 ms. In addition, it was shown that high-quality holograms could be generated even at depths not used for learning, such as generating holograms at longer distances using networks learned from 20 cm to 30 cm. It has been demonstrated that holographic images can be reproduced from a long distance without re-learning because the suggested approach in this thesis can have generality about depth. The proposed method has high utility as it can be applied together with previously proposed deep learning networks.
Since how to adjust the target depth by adding a depth embedding block is learned, the next task is to propose a deep learning model that can reconstruct a holographic image for multiple planes rather than a single plane. In particular, in the case of a phase-only hologram, there was no deep learning model capable of reconstructing holographic images in multiple planes. Here, first, a plurality of fixed target depths are used to train a deep learning model that reconstructs images in multiple planes. Ultimately, it should be possible to adjust the target depth, but an approach to find an appropriate loss function and hyperparameter is used while fixing the target depth. The target depth was then input and a phase-only hologram corresponding to the target depth was generated. In order to achieve stable learning even in the state of having multiple target depths, a new target function has been added to suit the learning goals. Using the characteristic that the hologram is diffracted through space, the diffraction pattern was included in the loss function to learn, and it was newly calculated each time by modeling the way light travels in space. When the diffraction pattern calculated by simulation was included in the learning, it was confirmed that the learning progressed stably. As a result of testing on the test dataset, it was confirmed that the PSNR increased by 5.3 dB and the SSIM by 0.11.