CAHVAE: generating CGHs with complex amplitude hologram variational autoencodern

Deep learning-based generative models for computer-generated hologram (CGH) have proven effective in overcoming the challenge of limited holographic data. However, most existing models, including VAE, GAN-Holo, and diffusion models, use real-valued convolution kernels, which fail to fully capture the complexity of optical field waves. To address this issue, we propose the complex amplitude hologram variational autoencoder (CAHVAE), what we believe to be a novel approach specifically designed for synthesizing complex amplitude holograms. Unlike traditional methods, CAHVAE directly processes complex optical field data using a complex-valued encoder and decoder, bridging the gap between real-valued kernels and the true nature of complex amplitude holograms. By modeling the latent space with a complex multivariate Gaussian distribution, CAHVAE allows for efficient random sampling of complex-valued latent variables, enabling the generation of high-quality new complex amplitude holograms, which do not exist in the original datasets. Simulations and optical experiments demonstrate the effectiveness of CAHVAE in generating new complex field waves, reconstructing high-quality color holograms, and preserving fine details. Notably, CAHVAE effectively reduces speckle noise and enhances the quality of reconstructed images, making it a promising solution for high-fidelity holographic display applications.