一种无需非线性结构的光学卷积神经网络  

Optical Convolutional Neural Network for Image Recognition without Nonlinear Structures

作　　者：江奔 张薇[1] 许涛[1] Ben Jiang;Wei Zhang;Tao Xu(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai Received:Apr)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海

出　　处：《建模与仿真》2024年第3期3851-3860,共10页Modeling and Simulation

基　　金：国家重点研发计划资助(2021YFB2802300)。

摘　　要：卷积神经网络在视觉处理方面具有独特的优势。最近的一些研究使用光学的卷积神经网络来实现更为快速和低功耗的图像处理系统。我们的工作提出了一种无需额外非线性结构的光学卷积神经网络,使用绝对值函数作为激活函数,设置了特定的三层卷积神经网络,该网络对MNIST和Fashion-MNIST数据集的识别准确率与目前普遍使用的激活函数相差不大。通过对软件仿真与硬件仿真的结果进行对比,发现图像经过第一层卷积和非线性后的结果误差不超过2%,输出结果验证了我们的光学卷积神经网络对图像处理的有效性。这为实现高效可编程的光学卷积神经网络提供了可行方案。Convolutional neural networks(CNNs)have unique advantages in visual processing.Some recent studies have used optical CNNs to realize faster and low-power image processing systems.This study proposes an optical CNN without the need for additional nonlinear structures,using the absolute value function as the activation function.A specific three-layer CNN developed in this study provided recognition accuracies for the MNIST and Fashion-MNIST datasets similar to those of the commonly used activation functions currently in use.By comparing the results of the software simulation with the hardware simulation,it is found that the error of its results after the first layer of convolution and nonlinearity was not more than 2%,and the output results verify the effectiveness of our optical CNN for image processing.This provides a viable solution for realizing efficient and programmable optical CNNs.

关 键 词：光学卷积神经网络 激活函数 马赫–曾德尔干涉仪 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]