关键词: 卷积神经网络, 多层特征融合, 多尺度池化, 遥感图像场景分类

Abstract: In the scene classification of moderate resolution remote sensing images, the traditional feature extraction method cannot be used to extract comprehensive features, while in the case of scene classification using convolutional neural network, the convolution kernel of the same size cannot extract features in ground objects of different sizes, which reduces the classification accuracy. In order to extract features in different sizes and improve the accuracy of classification, this paper proposes a moderate resolution remote sensing scene classification algorithm based on multi-scale feature fusion. The traditional convolutional neural network was improved to adapt to the moderate resolution remote sensing data set, and multi-scale pooling was added, and multi-level feature maps were input to the full-connection layer for classification. The experiment shows that the feature information extracted by multi-layer feature fusion method is more comprehensive and the classification effect is better than that by single-layer multi-scale pooling method. Compared with other traditional classification methods, our method yields better classification results.- DOI:10.3969/j.issn.1000-1565.2019.06.015基于多尺度特征融合的中分辨率遥感场景分类算法张军^1,2,张敏^1,2,郝小可^1,2,解鹏^1,2(1.河北工业大学 人工智能与数据科学学院,天津 300401;2.河北省大数据计算重点实验室,天津 300401)摘 要:在对中分辨率遥感图像进行场景分类时,传统的特征提取方法无法提取全面的特征,若使用卷积神经网络进行场景分类,同一大小的卷积核无法提取尺寸大小各异的地物特征,导致分类精度降低.为了提取不同尺寸的地物特征,提高分类精度,本文提出一种基于多尺度特征融合的中分辨率遥感场景分类算法.对传统的卷积神经网络进行改进以适应中分辨率遥感数据集,并在其基础上添加多尺度池化,将连接多层次的特征图谱输入到全连接层进行分类.实验表明,多层特征融合方法提取的特征信息比单层多尺度池化方法提取的特征信息更全面,分类效果更优.与其他的传统分类方法相比,本文方法获得更好的分类结果.关键词:卷积神经网络;多层特征融合;多尺度池化;遥感图像场景分类中图分类号:TP399 文献标志码:A 文章编号:1000-1565(2019)06-0666-07Moderate resolution remote sensing scene classification method based on multi-scale feature fusionZHANG Jun^1,2, ZHANG Min^1,2, HAO Xiaoke^1,2, XIE Peng^1,2(1. School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China; 2. Hebei Province Key Laboratory of Big Data Calculation, Tianjin 300401, China)Abstract: In the scene classification of moderate resolution remote sensing images, the traditional feature extraction method cannot be used to extract comprehensive features, while in the case of scene classification using convolutional neural network, the convolution kernel of the same size cannot extract features in ground objects of different sizes, which reduces the classification accuracy. In order to extract features in different sizes and improve the accuracy of classification, this paper proposes a moderate resolution remote sensing scene classification algorithm based on multi-scale feature fusion. The traditional convolutional neural network was improved to adapt to the moderate resolution remote sensing data set, and multi-scale pooling was added, and multi-level feature maps were input to the full-connection layer for classification. The experiment shows that the feature information extracted by multi-layer feature fusion method is more comprehensive and the classification effect is better than that by single-layer multi-scale pooling method. Compared with other traditional classification methods, our method yields better classification results.- 收稿日期:2019-02-15 基金项目:国家自然科学青年基金资助项目(61806071);天津市自然科学基金资助项目(16JCYBJC15600);河北省自然科学基金资助项目(F2017202145) 第一作者:张军(1976—),男,河北张家口人,河北工业大学副教授,博士,主要从事智能信息处理研究.E-mail: zhangjun@scse.hebut.edu.cn 通信作者:郝小可(1985—),男,天津市人,河北工业大学讲师,博士,主要从事机器学习、医学图像分析研究.E-mail:haoxiaoke@scse.hebut.edu.cn第6期张军等:基于多尺度特征融合的中分辨率遥感场景分类算法

Key words: convolutional neural network, multi-level feature fusion, multi-scale pooling, remote sensing image scene classification