Dual-multiplexed coaxial holograms reconstruction based all-optical diffraction deep neural network

Abstract

The all-optical diffractive deep neural network (D2NN) utilizes passive diffraction layers to perform machine learning, achieving complex functions at the speed of light, akin to traditional computer-based neural networks. This paper explores a dual-multiplexed coaxial hologram reconstruction technique based on an all-optical D2NN. In this approach, the input holograms are processed by two sets of transmissive layers trained in parallel. By exploiting the inherent parallelism of optical systems, we divide the optical path into two jointly trained diffractive networks that work in parallel, reducing crosstalk and optical signal coupling between the two images. The results show that the dual-multiplexed coaxial holograms can be simultaneously reconstructed by both sets of layers, effectively eliminating twin image artifacts. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) of the reconstructed image improve by 14.29% and 25%, respectively, compared to reconstructions of a single hologram using all-optical D2NN. Additionally, we assess the network’s performance with noisy and partially occluded holograms, demonstrating that, unlike conventional methods, this approach significantly enhances image quality, even under salt-and-pepper noise or partial occlusion. These findings offer new insights into the real-time reconstruction of dual-multiplexed digital holograms.