基于密集卷积网络(DenseNets)的遥感图像分类研究  

Classification of remote sensing images based on densely connected convolutional networks

作　　者：李达 李琳[1] 李想 Li Da;Li Lin;Li Xiang(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;School of mechanical and electrical engineering,Suzhou university)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]苏州大学机电工程学院

出　　处：《计算机时代》2018年第10期60-63,67,共5页Computer Era

摘　　要：遥感图像空间分辨率低,如何更好地提取图像特征成为提升分类性能的关键。文章提出了一种基于密集卷积网络(DenseNets)的遥感图像分类方法,针对遥感图像样本少,采用迁移学习方法,在ImageNet上进行预训练,获得初始模型,利用预训练模型在(UCM_LandUse_21)上训练,更新训练策略获得最佳模型。结果表明,该方法比BOVW+SCK和SVM_LDA方法在分类精度上提高10%,比传统CNN提升了约7%,比MS_DCNN提升5%。因此,该方法对于遥感图像场景分类具有一定的价值。The spatial resolution of remote sensing images is low, so how to better extract image features has become the key to improve the classification performance. In this regard, this paper proposes a remote sensing image classification method based on densely connected convolutional networks （DenseNets）. For the small number of remote sensing image samples, transfer learning method is adopted to conduct pre-training on ImageNet and obtain the initial model. And using the initial model conducts training on UCM_ LandUse _21 with the training policy updated to obtain the best model. The results show that the method is 10% higher than BOVW＋SCK and SVM LDA in classification accuracy, 7% higher than traditional CNN and 5% higher than MS DCNN. Therefore, the method proposed in this paper has certain value for remote sensing image scene classification.

关 键 词：遥感图像分类 密集卷积网络 迁移学习 场景分类 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]