基于轻量化人工智能方法的手写体识别研究  

Research on handwriting recognition based on lightweight artificial intelligence method

作　　者：刘世泽 李紫蕊 范书瑞 贾颖淼 张晓旭 LIU Shize;LI Zirui;FAN Shurui;JIA Yingmiao;ZHANG Xiaoxu(School of Software,Beihang University,Beijing 100191,China;School of Electronics and Engineering,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]北京航空航天大学软件学院,北京100191 [2]河北工业大学电子信息工程学院,天津300401

出　　处：《现代电子技术》2022年第10期165-170,共6页Modern Electronics Technique

基　　金：国家自然科学基金面上项目(42075129);河北省高等学校科学技术研究重点项目(ZD2019010);河北省研究生创新资助项目(CXZZSS2020031)。

摘　　要：文中利用边云协同的概念,解决边缘设备运算资源与存储资源不足的问题。在服务器端训练模型并进行性能测试。挑选出轻量级的MobileNet模型后将其部署在边缘设备,并对手写体识别结果进行分析。结果表明:与其他经典模型相比,在PC端利用深度可分离卷积的MobileNet模型耗时可减少近50%;部署到嵌入式端后,其运算时间变为2 ms,运算速度提升几十倍,可解决人工智能部署到嵌入式端的难题。The concept of edge cloud collaboration is used to solve the problems of insufficient computing and storage resources of edge devices. The model is trained on the server and its performance is tested. The lightweight MobileNet is selected and then deployed on the edge device, and the handwriting recognition results are analyzed. The results show that, in comparison with other classical models,the time consumption of MobileNet using deep separable convolution on the PC side can be reduced by nearly 50%. After being deployed on the embedded side,the computing time becomes 2 ms,and the operation speed is increased by dozens of times,which can solve the problem of artificial intelligence deployment to the embedded end.

关 键 词：人工智能 边缘计算 手写体识别 端侧轻量化 Tengine EMNIST数据集 MobileNet模型 模型部署 

分 类 号：TN915-34[电子电信—通信与信息系统] TP18[电子电信—信息与通信工程]