三层衍射神经网络实现手写数字识别  

Handwritten digit recognition by three-layer diffractive neural network

作　　者：徐平[1] 徐海东 杨拓[1] 黄海漩 张旭琳[1] 袁霞[1] 肖钰斐[1] 李雄超[1] 王梦禹 Xu Ping;Xu Hai-Dong;Yang Tuo;Huang Hai-Xuan;Zhang Xu-Lin;Yuan Xia;Xiao Yu-Fei;Li Xiong-Chao;Wang Meng-Yu(Institute of Micro-Nano Optoelectronic Technology,College of Physics and Optoelectronic Engineering,Shenzhen University,Shenzhen 518060,China;College of Big Data and Internet,Shenzhen Technology University,Shenzhen 518118,China)

机构地区：[1]深圳大学物理与光电工程学院,微纳光电子技术研究所,深圳518060 [2]深圳技术大学大数据与互联网学院,深圳518118

出　　处：《物理学报》2022年第18期209-216,共8页Acta Physica Sinica

基　　金：国家自然科学基金(批准号:61275167);深圳市高等院校稳定支持计划(批准号:20200812103045003);深圳市基础研究自由探索项目(批准号:JCYJ20180305125430954,JCYJ20170817102315892,JCYJ2017081701827765)资助的课题。

摘　　要：光学衍射神经网络(optical diffraction neural network,ODNN)以光波作为计算媒介执行神经网络的逻辑分析与运算功能,具有高速度、低功耗及并行处理的优势.本文设计了一种仅有三层相位调制的ODNN,提出了基于目标空间频率一级谱分布提升ODNN的数字识别性能的方法,经优化获得了系统最优的像素大小、衍射距离,以及最佳的三层相位分布.设计的ODNN对MNIST手写体数字集识别准确率达到了95.3%,高于文献中采用五层衍射神经网络实现的准确率91.75%(Lin X,Rivenson Y,Yardimci N T,Veli M,Luo Y,Jarrahi M,Ozcan A 2018 Science 3611004),且精简了系统结构.结合ODNN高速度、低功耗的优点,提出的基于频谱分析方法有利于提高ODNN的性能,使ODNN在边缘计算领域有巨大的应用潜力.Optical diffractive neural network(ODNN)uses light wave as a computing medium to perform the inference and prediction function of neural network,which has the advantages of high speed,low power consumption,and parallel processing.In this work,an ODNN with only three layers of phase modulation is designed,and a method to improve the recognition performance of ODNN based on the first-order spectral distribution of targets is proposed.Using this method,the parameters of a three-layer ODNN are effectively optimized and the optimal pixel size,diffraction distance,and method for image preprocessing are obtained.The three-layer ODNN designed in this work has a recognition accuracy rate of 95.3%for MNIST(handwritten digit set),compared with the 91.75%accuracy achieved by the five-layer ODNN in the reference(Lin X,Rivenson Y,Yardimci N T,Veli M,Luo Y,Jarrahi M,Ozcan A 2018 Science 3611004).In addition,the system volume is greatly reduced and the system structure is simplified.Combined with the advantages of high speed and low power consumption,it has huge potential applications in the fields such as edge computing in the future.

关 键 词：衍射神经网络 光学识别 手写数字 衍射光学元件 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]