摘要
为解决图像分类中深度卷积神经网络(Convolutional neural networks,CNN)中较为复杂的人工网络设计与调参问题,提出基于ResNet模块的进化卷积、神经网络(Evolutionary convolutional neural network,ECNN)的自动设计方法,并将其运用到图像分类中.该方法基于ResNet模块与2D卷积层,采用进化算法(Evolutionary algorithm,EA)对网络结构及参数进行优化.在NLM官方发布的疟疾数据集下进行实验,不同比例的测试集划分可以达到95.6%的分类准确率,文中算法与AlexNet、VGG16、Xception等人工设计的深度学习分类算法进行了比较,实验结果表明,其准确率提升了约1%.在斯坦福大学发布的Stanford cars车辆图像数据集中进行了算法泛化验证,结果表明,文中算法在不同比例数据的测试中准确率均在94.5%以上,将该算法与深度学习分类算法VGG16进行比较,准确率效果相当,模型测试图像分类耗时仅为VGG16耗时的1/13,且训练参数量较少.两组对比测试实验数据表明,相比人工设计的深度学习算法,本文方法具有较好的图像分类性能与较快的图像分类速度.
In order to solve the problems of network design and the manual adjustment of parameters in the deep convolution neural network of image classification,an automatic design network method using the ResNet module evolutionary convolution neural network is proposed and applied to image classification.Based on ResNet module and 2D convolution layer,the parameters and network structure are optimized by evolutionary algorithm.In the experiment of malaria data set published by NLM,95.6% classification accuracy can still be achieved with different proportion of test set division.The ECNN-1 algorithm model is compared with AlexNet,VGG16,Xception and other manually designed deep learning classification algorithms,and the experimental results show that the accuracy effect is improved by about 1%.The experimental results of Stanford cars vehicle image data set classification published by Stanford University show that the accuracy of ECNN-2 algorithm is over 94.5 % in the test of different scale data.Comparing the ECNN-2 algorithm with VGG16 algorithm,the accuracy effect is equivalent.The image classification time of model test is only 1/13 of VGG16 and the amount of training parameters is less.Two groups of experimental data show that compared with the artificially designed deep learning algorithm,this method has better image classification performance and faster speed.
作者
马永杰
刘培培
MA Yong-jie;LIU Pei-pei(College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,Gansu,China)
出处
《西北师范大学学报（自然科学版）》
CAS
北大核心
2020年第3期55-61,134,共8页
Journal of Northwest Normal University(Natural Science)
基金
国家自然科学基金资助项目(41461078)。
关键词
卷积神经网络
ResNet
进化卷积神经网络
图像分类
convolutional neural networks
ResNet
evolutionary convolutional neural network
image classification
作者简介
马永杰(1967—),男,甘肃灵台人,教授,博士,硕士研究生导师.主要研究方向为进化算法、人工智能和测控技术.E-mail:myjmyj@163.com。
