摘要
数字识别在邮政编码、车牌数字检测识别等场景有着广泛的应用。以Lenet-5卷积神经网络为基础,研究卷积计算量过大和预测速度慢的问题。对不同的卷积层(首层、中间层、尾层)和网络结构进行了改进和优化,采用不同组合的方案使卷积核连接数减少一定的数量,进行仿真实验。实验结果表明,保证在同一准确率的前提下,综合考虑迭代次数和预测时延,在C3层(中间层)做优化最为合适,总结出的规律和方法也可对复杂卷积神经网络优化提供参考,满足实时性要求高、大数据量的应用场景。
Digital recognition has a wide range of applications in postal code and license plate digital detection and recog-nition.The problem of over large convolution calculated amount and slow prediction speed are studied in this paper on the basis of Lenet-5 convolution neural network.The different convolution layers(first layer,middle layer,last layer)and network struc-ture are improved and optimized.Different combinational solutions are used to reduce the linking number of convolution kernels for simulation experiment.The simulation experiment results show that the C3(middle layer)is the most suitable layer for opti-mization,in the premise of ensuring the same accuracy and considering the iterations and prediction delay comprehensively.This guideline can also provide a reference for the optimization of complex convolutional neural networks and meet some applica-tion scenarios with high real-time requirements and large data volume.
作者
杜阔
李亚
DU Kuo;LI Ya(Tianjin University of Science & Technology,Tianjin 300222,China)
出处
《现代电子技术》
北大核心
2019年第16期98-103,共6页
Modern Electronics Technique
基金
国家自然科学基金青年科学基金项目(61705166)~~
关键词
数字分类器
卷积神经网络
卷积计算
数字识别
网络结构优化
仿真实验
digital classifier
convolutional neural network
convolution calculation
digital recognition
network structureoptimization
simulation experiment
