Scholarlink

In the field of image recognition, Diffractive Deep Neural Networks (D2NNs) utilize multiple diffractive layers to modulate light signals and perform complex computational tasks. This paper presents a five-layer D2NN architecture designed to classify images from the MNIST and Fashion-MNIST datasets, employing both phase-only and amplitude-only modulation techniques. We compare the performance of the D2NN under these different modulation modes. The five-layer D2NN with amplitude-only modulation achieves accuracies of 96.30% and 81.77% on the MNIST and Fashion-MNIST datasets, respectively, while the five-layer D2NN with phase-only modulation achieves accuracies of 96.97% and 85.87%. The results demonstrate a positive correlation between the number of diffractive layers and test accuracy, indicating that an increase in the number of diffractive layers enhances the network’s expressive capacity and recognition performance. This study provides a foundation for the practical application of diffractive deep neural networks in future research.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diffractive Deep Neural Network Based on Modulation of Amplitude or Phase

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yu Wang,
  • Feng Qi,
  • Qi Sha

摘要

In the field of image recognition, Diffractive Deep Neural Networks (D2NNs) utilize multiple diffractive layers to modulate light signals and perform complex computational tasks. This paper presents a five-layer D2NN architecture designed to classify images from the MNIST and Fashion-MNIST datasets, employing both phase-only and amplitude-only modulation techniques. We compare the performance of the D2NN under these different modulation modes. The five-layer D2NN with amplitude-only modulation achieves accuracies of 96.30% and 81.77% on the MNIST and Fashion-MNIST datasets, respectively, while the five-layer D2NN with phase-only modulation achieves accuracies of 96.97% and 85.87%. The results demonstrate a positive correlation between the number of diffractive layers and test accuracy, indicating that an increase in the number of diffractive layers enhances the network’s expressive capacity and recognition performance. This study provides a foundation for the practical application of diffractive deep neural networks in future research.