基于改进密集连接网络的遥感图像场景分类  

Remote sensing image scene classification based on improved dense connectivity networks

作　　者：宋中山[1,2] 彭丹 郑禄 帖军[1,3] 龙吕佳 SONG Zhongshan;PENG Dan;ZHENG Lu;TIE Jun;LONG Lyujia(College of Computer Science,South-Central Minzu University,Wuhan 430074,China;Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprises,Wuhan 430074,China;Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]湖北省制造企业智能管理工程技术研究中心,武汉430074 [3]农业区块链与智能管理湖北省工程研究中心,武汉430074

出　　处：《激光杂志》2023年第10期71-78,共8页Laser Journal

基　　金：国家民委中青年英才培养计划(No.MZR20007);湖北省科技重大专项(No.2020AEA011);武汉市科技计划应用基础前沿项目(No.2020020601012267);新疆维吾尔自治区区域协同创新专项(No.科技援疆计划)(No.2022E02035);中南民族大学研究生创新基金(No.3212022sycxjj334)。

摘　　要：针对遥感图像场景分类任务从复杂背景下准确提取出地物信息困难和普通卷积提取特征容易产生冗余特征的问题,提出一种基于改进密集连接网络(Ghost-Densenet)的分类模型。该模型利用SoftPool对MaxPool和AveragePool进行替换,最大程度上保留了遥感图像的主要特征;利用Ghost模块通过简单线性变化生成特征图的特性,有效增强模型特征提取能力的同时减少了网络瓶颈层的冗余特征和网络的参数量与计算量。实验结果表明,该模型在UC Merced_Land Use数据集上的平均准确率为92.76%,相较于Densenet121,模型大小减少26.57%,计算量降低32.99%,准确率提高1.17%。通过在Aerial Image Dataset、WHU-RS19 Date Set、RSSCN7 Date Set、SIRI-WHU Date Set四个数据集上进行实验,验证了模型的有效性和鲁棒性,对遥感图像场景分类任务具有良好的应用价值。Aiming at the difficulties of accurately extracting object information from complex backgrounds in remote sensing image scene classification tasks and the problems that ordinary convolutional extraction tends to produce redundant features,a classification model based on an improved dense link network(Ghost-Densenet)is proposed.The model uses SoftPool to replace MaxPool and AveragePool to retain the main features of remote sensing images to the maximum extent.It utilizes the feature of Ghost module to generate feature maps by simple linear variation,which can effectively enhance the feature extraction ability of the model and also reduce the number of redundant features in the bottleneck layer of the network and the number of parameters and computation of the network.The experimental results show that the model has an average accuracy of 92.76%on the UC Merced_Land Use dataset.Compared with Densenet121,its model size is reduced by 26.57%,computation is reduced by 32.99%,and accuracy is improved by 1.17%.By conducting experiments on four datasets,namely,Aerial Image Dataset,WHU-RS19 Date Set,RSSCN7 Date Set,and SIRI-WHU Date Set,the effectiveness and robustness of the model are verified,and it has good application value for remote sensing image scene classification tasks.

关 键 词：场景分类 密集连接网络 GhostNet Softpool 深度学习 

分 类 号：TN911[电子电信—通信与信息系统]