The rapidly growing computational demands of deep learning are increasingly constrained by the performance limitations of conventional electronic computing hardware. Optical Diffractive Neural Networks (ODNNs) emerge as a promising solution to this challenge by harnessing the unique advantages of light waves, including high-speed propagation and ultralow power consumption, for data processing. In this paper, we derive a theoretical framework based on optical diffraction principles and design a five-layer phase-modulated ODNN architecture. Through systematic analysis of key factors such as input image dimensions, pixel resolution, inter-layer distance and the number of modulation layers, we optimize the recognition system. The proposed model achieves 97.1% accuracy on the MNIST handwritten digit dataset, demonstrating successful simulation and improved performance in optical image recognition tasks. These results validate the significant potential and practical value of ODNNs in computer vision applications.
Chen et al. (Fri,) studied this question.