摘要
将传统卷积神经网络应用于小数据集上,LeNet模型准确率低并且收敛速度慢,VggNet等模型存在过拟合问题.针对小数据集提出一种改进LeNet模型,该模型在LeNet基础上使用ReLU函数替换sigmoid来提高收敛速度,加入1*1卷积增加模型深度并利用其改变维度的特点来提高识别准确率,通过分解卷积和提出改进Dropout方法减少过拟合.结果表明:改进LeNet模型分类自制小龙虾数据集,比LeNet收敛速度快6000步并且准确率提高约15%,比VggNet和ResNet过拟合程度明显减少;将改进LeNet模型推广应用于开源数据集MNIST和Fashion-MNIST上,改进模型也有良好的表现.
The results of applying traditional convolutional neural networks to small datasets showed that LeNet model had the problems of low accuracy and slow convergence,other models like VggNet appeared over-fitting matters.An improved LeNet model for small datasets is proposed,which was mainly changed as follows.Firstly,sigmoid is replaced with ReLU to improve the convergence speed.Secondly,1*1 convolution is added to increase the depth of the model and improve recognition accuracy.Finally,an improved dropout method is proposed to reduce over-fitting.In the crayfish project,the proposed model converged 6000 steps faster and had a 15%higher accuracy than LeNet,while the degree of over-fitting was significantly reduced compared to VggNet and ResNet.The proposed LeNet model was also applied to the open source datasets MNIST and Fashion-MNIST,and the results show good performance.
作者
舒军
杨露
陈义红
杨莉
邓芳
SHU Jun;YANG Lu;CHEN Yihong;YANG Li;DENG Fang(Hubei Collaborative Innovation Center for High-efficiency Utilization of Solar Energy,Hubei University of Technology,Wuhan 430068,China;Wuhan Sintec Optronics Co.,LTD,Wuhan 430205,China;College of Computer Hubei University of Education,Wuhan 430025,China)
出处
《中南民族大学学报(自然科学版)》
CAS
2019年第4期605-612,共8页
Journal of South-Central University for Nationalities:Natural Science Edition
基金
湖北省科技厅重大专项(2017ACA105)