Unsupervised deep neural network for fast computer-generated holography with continuous depth control

In computer-generated holography, leveraging deep learning for holographic display with continuous depth poses challenges in balancing depth cues and runtime. Sufficient data are typically not available for existing deep neural networks based on point-based datasets. To address this, we propose an unsupervised learning neural network named Polygon-based Holo, based on an efficient diffraction model that enables calculations from a reference plane to tilted plane to generate the target field wavefront. Polygon-based Holo is trained using the proposed diffraction model through an unsupervised process, employing mid-holograms superimposed with a small number of tilted planes. Polygon-based Holo can accurately predict wavefronts from input RGB-D images with continuous depth, displaying exceptional generalization across applications. The numerical simulations and optical experiments validate our approach, emphasizing its feasibility and potential for facilitating arbitrary depth control and dynamic display of holography.