Research on diffractive optical neural networks based on 4F systems and phase-type dielectric metasurfaces and application in image recognition

Rapid and accurate image recognition holds significant application prospects in fields such as autonomous obstacle detection, satellite remote sensing, and medical imaging. The traditional image recognition requires steps including image recording, storage, and transmission, which not only incur substantial hardware and software costs, but are also slow and power-intensive. To address this issue, we propose a diffractive optical neural network that is equipped with a 4F multi-kernel convolutional front-end and four diffractive fully connected layers. The convolutional layers are implemented by a 4F system with a dielectric metasurface placed on the Fourier plane to extract the frequency-domain features, while the fully connected layers are composed of the same dielectric metasurface. With the convolutional layer, the model achieves 4.25% and 8.25% higher accuracy on the MNIST and Fashion-MNIST datasets, respectively. Each unit of the metasurface can modify the phase in the range of 0 to 2π, and can be realized by the dielectric nano-pillar array in a square lattice. The network training was carried out using the Rayleigh–Sommerfeld diffraction theory and a centrosymmetric initialization method to improve the training accuracy. After training, this diffractive optical neural network can perform highly precise image-classification tasks and yields accuracies of 98.32% for the handwritten numbers and 88.22% for the clothing images, respectively. These results indicate the potential application prospects of the designed optical neural network.