We propose a new learning and inferring model that generates digital holograms using deep neural networks (DNNs). This DNN uses a generative adversarial network, trained to infer a complex two-dimensional fringe pattern from a single object point. The intensity and fringe patterns inferred for each object point were multiplied, and all the fringe patterns were accumulated to generate a perfect hologram. This method can achieve generality by recording holograms for two spaces (16 Space and 32 Space). The reconstruction results of both spaces proved to be almost the same as numerical computer-generated holograms by showing the performance at 44.56 and 35.11 dB, respectively. Through displaying the generated hologram in the optical equipment, we proved that the holograms generated by the proposed DNN can be optically reconstructed.

In this paper, we propose a new system for a real-time holographic augmented reality (AR) video service based on a photorealistic three-dimensional (3D) object point for multiple users to use simultaneously at various locations and viewpoints. To observe the object from all viewpoints, a camera system capable of acquiring the 3D volume of a real object is developed and is used to generate a real object in real-time. Using the normal of the object point, the observable points are mapped to the viewpoint at which the user is located, and a hologram based on the object point is generated. The angle at which the reflected light from each point is incident on the hologram plane is calculated, and the intensity of the interference light is adjusted according to the angle to generate a hologram with a higher 3D effect. The generated hologram is transmitted to each user to provide a holographic AR service. The entire system consists of a camera system comprising eight RGB-D (depth) cameras and two workstations for photorealistic 3D volume and hologram generation. Using this technique, a realistic hologram was generated. Through experiments displaying holograms simultaneously from several different viewpoints, it is confirmed that multiple users can concurrently receive hologram AR.

Much research and development have been made to implement deep neural networks for various purposes with hardware. We implement the deep learning algorithm with a dedicated processor. Watermarking technology for ultra-high resolution digital images and videos needs to be implemented in hardware for real-time or high-speed operation. We propose an optimization methodology to implement a deep learning-based watermarking algorithm in hardware. The proposed optimization methodology includes algorithm and memory optimization. Next, we analyze a fixed-point number system suitable for implementing neural networks as hardware for watermarking. Using these, a hardware structure of a dedicated processor for watermarking based on deep learning technology is proposed and implemented as an application-specific integrated circuit (ASIC).

This paper proposes a method to embed and extract a watermark on a digital hologram using a deep neural network. The entire algorithm for watermarking digital holograms consists of three sub-networks. For the robustness of watermarking, an attack simulation is inserted inside the deep neural network. By including attack simulation and holographic reconstruction in the network, the deep neural network for watermarking can simultaneously train invisibility and robustness. We propose a network training method using hologram and reconstruction. After training the proposed network, we analyze the robustness of each attack and perform re-training according to this result to propose a method to improve the robustness. We quantitatively evaluate the results of robustness against various attacks and show the reliability of the proposed technique.

In this paper, we propose a method for compressing the phase component of a full-complex hologram. The JPEG (Joint Photographic Experts Group) Pleno is undergoing standardization for compressing full-complex holograms. If the full-complex hologram is compressed in the form of amplitude and phase components, the three-dimensional information of the hologram may be better preserved. Therefore, in order to solve the disadvantages of the method of independently compressing real and imaginary parts, we propose a method for directly compressing phase information. We select the HEVC (High Efficiency Video Coding), which has the best performance in compressing holograms from previous studies, as the anchor codec, and propose an algorithm for converting the phase information into the form suitable for the HEVC. Since the phase component is very random, we propose a modified phase unwrapping technique to improve this. In addition, in order to make good use of the property of HEVC Intra coding, the phase unwrapping considering Intra prediction is applied, and the most suitable HEVC Intra coding condition is searched. Compared with the result of compressing the phase using the HEVC, the hologram was improved by 2 dB or more and the reconstruction result was improved by more than 4 dB at a compression ratio of 80:1. If the compression ratio is increased, the proposed method has better results.

This paper proposes a coding method for compressing a phase-only hologram video (PoHV), which can be directly displayed in a commercial phase-only spatial light modulator. Recently, there has been active research to use a standard codec as an anchor to develop a new video coding for 3D data such as MPEG point cloud compression. The main merit of this approach is that if a new video codec is developed, the performance of relative coding methods can be increased simultaneously. Furthermore, compatibility is increased by the capability to use various anchor codecs, and the developing time is decreased. This paper uses a currently used video codec as an anchor codec and develops a coding method including progressive scaling and a deep neural network to overcome low temporal correlation between frames of a PoHV. Since it is difficult to temporally predict a correlation between frames of a PoHV, this paper adopts a scaling function and a neural network in the encoding and decoding process, not adding complexity to an anchor itself to predict temporal correlation. The proposed coding method shows an enhanced coding gain of an average of 22%, compared with an anchor in all coding conditions. When observing numerical and optical reconstructions, the result images by the proposed show clearer objects and less juddering than the result by the anchor.