Generating real-scene hologram through light field imaging and deep learning

Abstract

Abstract Addressing the challenge of acquiring holograms from real-world scenes, this study introduces a novel approach leveraging light field cameras to capture light field data, which is subsequently transformed into authentic scene holograms. This methodology integrates light field imaging technology with a pre-trained deep neural network. To compensate for the limitations inherent in camera hardware, a super-resolution algorithm is employed. The conversion of light field information into RGB-D data facilitates its input into the deep neural network, enabling the inference of corresponding real-world scene holograms. Empirical evidence demonstrates that the system is capable of inferring high-resolution (1920 × 1080) real-world scene holograms within a timeframe of 5 s, utilizing hardware comprising an NVIDIA RTX 3060.