摘要
卷积神经网络由于其出色的性能,在计算机视觉领域被广泛使用。但是由于卷积神经网络其自身特性所限制,常常出现训练所需数据量大、模型训练困难等问题。为了达到模型轻量化的目的,文章改进了网络的基本模块,并将卷积核进行分解,使用卷积层代替全连接层,以达到减少参数量。实验证明所提出的模型分类正确率为90.5%,而且提出的模型在与ResNet18分类正确率相差无几的情况下,大幅度减少参数量和计算量,具有一定的应用价值。
Convolutional neural network is widely used in the field of computer vision because of its excellent performance.However,the convolutional neural network is limited by its own characteristics.Network models often have problems such as large amount of data required and difficult model training.In order to achieve the purpose of model lightweight,this paper redesigns the basic modules of the network,decomposes the convolution kernel,and uses the convolution layer to replace the full connection layer and reduce the number of parameters.The classification accuracy of this model is 90.5%.When the classification accuracy of this model is almost the same as that of ResNet18,the number of parameters and calculation are greatly reduced,which is of a certain practical value.
作者
黄承宁
李娟
朱玉全
HUANG Chengning;LI Juan;ZHU Yuquan(College of Computer and Communication,Engineering,Nanjing Tech University Pujiang Institute,Nanjing 211222,China;College of Computer Science and Telecommunications Engineering,Jiangsu University,Zhenjiang 212013,China)
出处
《微型电脑应用》
2023年第6期25-28,33,共5页
Microcomputer Applications
基金
国家自然科学基金(61702229)
江苏省高等学校自然科学研究项目(18KJD520001)
南京工业大学浦江学院人才培养工程计划项目(njpji2019-2-01)。
关键词
卷积神经网络
卷积核分解
全卷积网络
图像分类
模型轻量化
convolution neural network
convolution kernel decomposition
fully convolutional network
image classification
model lightweight
