Metasurface-Driven Diffractive Deep Neural Networks for Handwritten Digit Recognition

Abstract

This paper proposed an optical diffractive deep neural network based on metasurfaces for efficient handwritten digit recognition tasks. The network achieves precise control of the light field by using the flexible phase modulation capability of Pancharatnam-Berry metasurface, thus constructing an all-optical diffractive neural network. Simulation results show that the proposed D2NN achieves over 90% classification accuracy for handwritten digits "0", "1", "2", and "4" on blind test datasets. Additionally, the designed metasurface unit has a transmission efficiency as high as 95% at 1 THz, with digital images produced by metal masks being focused in designated areas after passing through the three-layer hidden layer of the metasurface. This study provides new insights into optical computing based on metasurfaces, offering potential application in machine vision, image processing, and real-time object recognition.