基于稠密块改进LinkNet的高分遥感图像道路提取  

Road Extraction Method of High-Resolution Remote Sensing Images Based on Dense Blocks and Improved LinkNet

作　　者：王增优 张鲜化[3] 刘荣[1,2] 陈志高[1,2] 朱旺煌 WANG Zengyou;ZHANG Xianhua;LIU Rong;CHEN Zhigao;ZHU Wanghuang(School of Surveying,Mapping and Spatial Information Engineering,East China University of Technology,Nanchang 330013,China;Key Laboratory of Mine Environmental Monitoring and Treatment in Poyang Lake Region,Ministry of Natural Resources,East China University of Technology,Nanchang 330013,China;Jiangxi College of Applied Technology,Ganzhou 341000,China)

机构地区：[1]东华理工大学测绘与空间信息工程学院,南昌330013 [2]东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,南昌330013 [3]江西应用技术职业学院,赣州341000

出　　处：《航天返回与遥感》2024年第3期107-117,共11页Spacecraft Recovery & Remote Sensing

基　　金：国家自然科学基金(42266006,41806114)。

摘　　要：针对LinkNet网络模型在进行道路图像分割任务时,特征信息易丢失以及缺乏对目标特征的关注度问题,提出了一种基于改进LinkNet残差网络的高分遥感图像道路提取方法。将原本LinkNet模型中编码区的残差块(Res Block)替换为稠密块(Dense Block),密集连接的方式减少特征信息在传递过程中的损失,并在每个稠密块之后构建卷积注意力单元来提高模型对目标特征的学习能力,最后用空洞空间金字塔池化模块将编码区与解码区进行连接,扩大感受野的同时还能接受多尺度目标特征信息。实验表明,该方法在DeepGlobe数据集上的准确率、平均交并比和F1-score分为82.16%、83.21%和81.65%,均优于同类网络,通过对提取的路网结果对比,该算法对于树木遮蔽处以及建筑物阴影下的路网提取在完整性和准确性上都具有明显提升。Aiming at the problem that feature information is easily lost and lacks attention to target features when the LinkNet network model performs road image segmentation tasks,a high resolution remote sensing image road extraction method based on an improved residual network in LinkNet is proposed.Replace the residual block(Res Block)in the coding area of the original LinkNet model with a dense block(Dense Block).The dense connection method reduces the loss of feature information during the transmission process,and builds convolutional attention after each dense block.Units are used to improve the model’s learning ability of target features.Finally,the atrous space pyramid pooling module is used to connect the encoding area and the decoding area to expand the receptive field while also accepting multi-scale target feature information.Experiments show that the accuracy,average intersection ratio and F1-score of this method on the DeepGlobe data set are 82.16%,83.21%and 81.65%,respectively,which are all better than similar networks.By comparing the extracted road network results,the algorithm has significantly improved the completeness and accuracy of road network extraction under tree shelters and building shadows.

关 键 词：残差网络 道路提取 稠密块 卷积注意力 空洞空间金字塔池化 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] V19[自动化与计算机技术—控制科学与工程]