采用融合卷积网的图像分类算法  

A novel fusion DCNN for image classification

作　　者：李聪 潘丽丽[1] 陈蓉玉 周燕 邵伟志 LI Cong;PAN Li-li;CHEN Rong-yu;ZHOU Yan;SHAO Wei-zhi(College of Computer Science and Information Technology,Central South University of Forestry&Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学计算机与信息工程学院

出　　处：《计算机工程与科学》2019年第12期2179-2186,共8页Computer Engineering & Science

基　　金：国家自然科学基金面上项目（61772561）;湖南省重点研发计划（2018NK2012）;中南林业科技大学研究生教育教学改革课题（2018JG005）;中南林业科技大学教育教学改革项目（20180682）

摘　　要：目前,卷积神经网络已成为视觉对象识别的主流机器学习方法。有研究表明,网络层数越深,所提取的深度特征表征能力越强。然而,当数据集规模不足时,过深的网络往往容易过拟合,深度特征的分类性能将受到制约。因此,提出了一种新的卷积神经网络分类算法:并行融合网FD-Net。以网络融合的方式提高特征的表达能力,并行融合网首先组织2个相同的子网并行提取图像特征,然后使用精心设计的特征融合器将子网特征进行多尺度融合,提取出更丰富、更精确的融合特征用于分类。此外,采用了随机失活和批量规范化等方法协助特征融合器去除冗余特征,并提出了相应的训练策略控制计算开销。最后,分别以经典的ResNet、InceptionV3、DenseNet和MobileNetV2作为基础模型,在UECFOOD-100和Caltech101等数据集上进行了深入的研究和评估。实验结果表明,并行融合网能在有限的训练样本上训练出识别能力更强的分类模型,有效提高图像的分类准确率。Deep Convolutional neural networks(DCNNs)have become the dominant machine learning approach for visual object recognition.Research studies have shown that the deeper the network layers are,the stronger the representational ability of the deep feature is.However,very deep neural networks would be affected by the overfitting problem and the classification performance of deep features will be restricted when the dataset scale is insufficient.Therefore,this paper proposes a new convolution neural network classification algorithm:Fusion-Double Network(FD-Net),which improves the expression ability of features by network fusion.The network first organizes two identical subnets to extract image features in parallel,then uses a well-designed feature fusion device to fuse the subnet features in multiple scales to extract richer and more accurate fusion features for classification.Furthermore,methods such as dropout and batch normalization are used to assist the feature fusion to remove redundant features,and the corresponding training strategy is proposed to control the computational expense.Finally,the classical ResNet50,InceptionV3,DenseNet121 and MobileNetV2 were used as subnets to construct the FD-Net,and the datasets of UECFOOD-100 and Caltech101 were studied and evaluated.Theoretical analysis and experimental results demonstrate that FD-Net improves the performance of convolutional neural networks with the limited training samples and achieves promising accuracy for classification.

关 键 词：深度学习 图像分类 卷积神经网络 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]