摘要
传统卷积神经网络存在卷积核单一、网络结构复杂和参数冗余的问题。提出一种轻量级特征融合卷积神经网络MS-FNet。在融合模块中采用多路结构以增加卷积神经网络的宽度,通过不同尺寸的卷积核对输入特征图进行处理,提高网络在同一层中提取不同特征的能力,并在每次卷积后采用批归一化、ReLU等方法去除冗余特征。此外,使用卷积层代替传统的全连接层,从而加快模型的训练速度,缓解因参数过多造成的过拟合现象。实验结果表明,MS-FNet可在降低错误率的同时,有效减少网络参数量。
The traditional Convolutional Neural Networks(CNN)suffer from single convolutional kernels,complex network structure and redundant parameters.To address the problem,a lightweight CNN named MS-FNet is designed for feature fusion.The fusion module employs a multi-branch structure to increase the width of the CNN,and different sizes of convolutional kernels to process the input feature map,which improves the ability of the network to extract different features in the same layer.And the redundant features are removed after each convolution by using BN,ReLU,etc.Convolutional layers are used to replace the traditional fully connected layer,which not only accelerates the training speed of model but also alleviates overfitting problems caused by too many parameters.The experimental results show that MS-FNet greatly reduces the number of network parameters and the error rate.
作者
陈鑫华
钱雪忠
宋威
CHEN Xinhua;QIAN Xuezhong;SONG Wei(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi,Jiangsu 214122,China;Engineering Research Center of Internet of Things Technology Applications of Ministry of Education,Jiangnan University,Wuxi,Jiangsu 214122,China;Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computer Intelligence,Jiangnan University,Wuxi,Jiangsu 214122,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2021年第11期268-275,共8页
Computer Engineering
基金
国家自然科学基金(61673193)
中国博士后科学基金(2017M621625)
江苏省自然科学基金(BK20181341)。
关键词
深度学习
卷积神经网络
特征提取
特征融合
图像分类
Deep Learning(DL)
Convolutional Neural Network(CNN)
feature extraction
feature fusion
image classification