Complex-valued generative adversarial network for real-time and high-quality computer-generated holography

Real-time generation of high-quality phase-only hologram (POH) is a critical challenge in holographic display. Deep learning-based computer-generated holography (CGH) offers innovative solutions to this challenging problem. In this paper, a complex-valued generative adversarial network (CV-GAN) is proposed for the real-time generation of high-quality POHs. Firstly, a complex-valued U-shaped network (U-Net) is designed as the generator of CV-GAN, which is able to process complex-valued data directly, thereby improving the learning efficiency of the network. Secondly, a complex-valued sub-pixel convolution up-sampling strategy is proposed and an adversarial learning mechanism is established to enhance the fitting ability of the network. Finally, the band-limited angular spectrum method is embedded into the neural network, allowing the CV-GAN to be trained in an unsupervised manner. The proposed algorithm generates high-quality 1920 × 1072 resolution POHs with an average peak-signal-to-noise ratio (PSNR) of 33.68 dB and an average structural similarity index measure (SSIM) of 0.95 at a speed of 51.63 frames per second (FPS) on the DF2K validation dataset. Both numerical reconstructions and optical experiments demonstrate that CV-GAN exhibits excellent fitting ability and lightweight network architecture, capable of generating high-quality POHs efficiently.