基于细节自适应和浅层引导的遥感影像道路提取
打开文本图片集
中图分类号：TP751.1；TP391 文献标识码：A doi：10.37188/CJLCD.2025-0131 CSTR：32172.14.CJLCD.2025-0131
Road extraction from remote sensing images based on shallow guidance and detail-adaptive
ZHAI Baoming 1，2 ， SHAO Pan 1，2* ， XIONG Biao¹²，QI Chenwei 1，2 ， GAO Yuqian 1，2 ，LI Jingyi2 （1. Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering ， Yichang ， China; 2. College of Computer & Information Technology， Yichang ， China）
Abstract：To address two key chalenges-namely，the multi-scale and multi-directional nature of roads and the frequent loss of fine-grained details，we propose a novel remote sensing road extraction network based on detail-adaptive and shallow feature guidance. Firstly，we design a shallow feature-guided attention module，which integrates strip convolutions，atention mechanisms，and shalow feature guidance to enhance the reconstruction of detailed information.Then，a multi-scale and multi-directional feature fusion module is developed，leveraging Gabor convolutions，local pooling，and global directional pooling to improve the representation of road features at various scales and directions.Finally，a detail-adaptive enhancement module is constructed using Gaussian filtering，interpolation，serpentine convolution，and standard convolution，which adaptively enhances the detailed information. The proposed network achieves F1- scores and IoU values of 74.62% and 62.38% on the CHN6-CUG dataset and 77.86% and 65.17% on the Massachusetts dataset， outperforming eight existing methods by at least 1.41% and 1.32% and 1.04% （204 and 0.88% ， respectively. Compared with other advanced methods，the proposed model demonstrates superior performance in road extraction tasks，maintaining high extraction accuracy while having a lower parameter count，showing a good balance between performance and eficiency.
Key words： remote sensing image;road extraction; multi-scale;attention；multi-directional convolution
1引言
随着高分辨率卫星遥感技术的快速发展，道路提取已成为遥感影像智能解译的关键任务，在智慧交通[1]（如无人驾驶[2]、车辆导航[3]）、城市规划[4]等领域具有重要应用价值[5]。（剩余16474字）
