摘要
LeNet-5卷积神经网络(LeNet-5 Convolutional Neural Network)虽然在手写数字识别中取得了不错的成绩,但是对具有复杂纹理特征的图像进行分类时准确率不高。针对LeNet-5网络对低层次特征利用率较低的问题,引入跨连结构,将第1个池化层和第2个池化层向后传播的同时与第2个全连接层相连,充分地利用网络提取的低层次特征。针对LeNet-5网络泛化能力低的问题,采用重叠池化并在后面加上局部响应归一化操作,提高网络的泛化能力。在Fer2013、Cifar-10和Fashion-MNIST数据集上进行的实验结果表明,与LeNet-5卷积神经网络相比,改进的LeNet-5卷积神经网络在复杂纹理特征数据集上表现出了更好的分类能力。
Although the LeNet-5 convolutional neural network had achieved good results in handwritten digital recognition,the accuracy of classifying images with complex texture features was not high. In view of the low utilization rate of low-level features in LeNet-5 network,a cross-connection structure was introduced to connect the first pooling layer and the second pooling layer with the second full-connection layer while propagating backward,so as to make full use of the low-level features extracted by the network. In view of the low generalization ability of LeNet-5 network,overlapping pooling was adopted and local response normalization operation was added after the pooling layer to increase the generalization ability of the model. Experimental results carried out on Fer2013,Cifar-10 and Fashion-MNIST data sets showed that the improved LeNet-5 convolutional neural network expressed better classification ability in complex texture feature data sets compared with LeNet-5 convolutional neural network.
作者
陈恩志
王春阳
李晨晨
吴夏铭
CHEN Enzhi;WANG Chunyang;LI Chenchen;WU Xiaming(School of Electronics and Information Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报(自然科学版)》
2022年第5期74-79,共6页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技发展计划项目(20160101279JC)。
关键词
卷积神经网络
图像分类
局部响应归一化
过拟合
convolutional neural network
image classification
local response normalization
overfitting