An Enhanced Diffractive Neural Network for Metasurface Holograms With High Resolution, Low Noise, and Uniform Intensity

Abstract

In this article, an enhanced diffractive neural network is proposed for achieving metasurface holograms with high resolution, low noise, and uniform intensity. First, we prove the feasibility of Rayleigh-Sommerfeld diffraction theory on a subwavelength scale. Based on this theory, the fully connected diffraction layer is constructed to build a high-resolution diffractive neural network (HR-DNN). Due to the capability of the diffraction layer in precisely manipulating subwavelength electromagnetic (EM) waves, high-resolution holographic imaging of complex patterns can be realized. In addition, a postprocessing method is particularly designed to separate clean target images from noisy holograms without reference assistance. The metrics, such as imaging efficiency (IE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM), are defined to estimate the imaging quality of the proposed HR-DNN-based holographic imaging system. Three types of complex patterns (airplane, phrase “WORLD PEACE 0921,” Olympic rings) are performed in the full-wave simulation, as well as the imaging results are highly recognizable with low-noise and uniform-intensity features. Compared with the weighted Gerchberg-Saxton (GS) algorithm, the proposed HR-DNN gains significant improvements in IE (241.6%), PSNR (45.6%), and SSIM (44.0%). Finally, a metasurface with $30\lambda \times 30\lambda $ based on 3-D printing technology is fabricated to image the Olympic rings. The measured results are in good agreement with the simulated and target ones. Therefore, the proposed HR-DNN can provide a pathway for high-resolution metasurface holograms.