融合多尺度与边缘特征的道路提取网络  

Road extraction networks fusing multiscale and edge features

作　　者：孙根云[1,2,3] 孙超 张爱竹[1] SUN Genyun;SUN Chao;ZHANG Aizhu(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,China;Key Laboratory of Natural Resources Monitoring in Tropical and Subtropical Area of South China,Ministry of Natural Resources,Guangzhou 510700,China;Laboratory for Marine Mineral Resources,Qingdao National Laboratory for Marine Science and Technology,Qingdao 266237,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580 [2]自然资源部华南热带亚热带自然资源监测重点实验室,广东广州510700 [3]青岛海洋科学与技术试点国家实验室海洋矿产资源评价与探测技术功能实验室,山东青岛266237

出　　处：《测绘学报》2024年第12期2233-2243,共11页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42371350,42271347);自然资源部华南热带亚热带自然资源监测重点实验室开放基金(2024NRMK03)。

摘　　要：利用遥感影像提取道路对城市发展有重要意义。但是由于道路尺度多变、易被遮挡等因素,导致出现道路漏检、边缘不完整等问题。针对以上问题,本文提出了一种融合多尺度与边缘细节特征的道路提取网络(MeD-Net)。MeD-Net包括道路分割与边缘提取两部分。道路分割网络使用多尺度深层特征处理模块(MDFP),提取顾及全局与局部信息的多尺度特征,在卷积后使用组归一化优化模型训练;边缘提取网络利用细节引导融合算法提升深层边缘特征的细节信息,并利用注意力机制进行特征融合。为验证算法性能,本文利用Massachusetts道路数据集和青岛地区GF-2号道路数据集进行试验。试验表明,MeD-Net在两个数据集上交并比和F_(1)值均取得最高精度,能够提取不同尺度道路和更完整地保持道路边缘。Extracting roads using remote sensing images is of great significance to urban development.However,due to factors such as variable scale of roads and easy to be obscured,it leads to problems such as road miss detection and incomplete edges.To address the above problems,this paper proposes a network(MeD-Net)for road extraction from remote sensing images integrating multi-scale features and focusing on edge detail features.MeD-Net consists of two parts:road segmentation and edge extraction.The road segmentation network uses multi-scale deep feature processing(MDFP)module to extract multi-scale features taking into account global and local information,and is trained using group normalization optimization model after convolution.The edge extraction network uses detail-guided fusion algorithms to enhance the detail information of deep edge features and uses attention mechanisms for feature fusion.To verify the algorithm performance,this paper conducts experiments using the Massachusetts road dataset and the GF-2 road dataset in Qingdao area.The experiments show that MeD-Net achieves the highest accuracy in both datasets in terms of intersection-over-union ratio and F_(1)value,and is able to extract roads at different scales and maintain road edges more completely.

关 键 词：道路提取 语义分割 多尺度特征 边缘提取 

分 类 号：P231[天文地球—摄影测量与遥感]