面向遥感图像道路提取的多尺度上下文感知网络  

Multi-scale context-aware network for road extraction in remote sensing images

作　　者：李智杰[1] 惠爱婷 李昌华[1] 董玮 张颉[1] 介军[1] LI Zhijie;HUI Aiting;LI Changhua;DONG Wei;ZHANG Jie;JIE Jun(School of Information and Control Engineering,Xi’an University of Architectural Science and Technology,Xi'an 710055,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,陕西西安710055

出　　处：《光学精密工程》2025年第4期610-623,共14页Optics and Precision Engineering

基　　金：国家自然科学基金(No.62276207);陕西省住房城乡建设科技计划项目(No.2020-K09);陕西省教育厅协同创新中心项目(No.23JY038)。

摘　　要：针对深度神经网络在遥感图像道路提取中面临的局部特征丢失和提取精度低的问题,本文基于SwinUnet网络提出了一种面向遥感图像道路提取的多尺度上下文感知网络。首先,在编码器中设计了一个具有上下文聚合模块的分支,以增强上下文信息提取,并缓解因遮挡引起的语义模糊问题。其次,为了解决编解码器之间语义信息不匹配的问题并提高模型的空间信息提取能力,在跳跃连接中引入了空间特征提取模块,取代了SwinUnet中直接复制编码器特征的方法。最后,在下采样阶段设计了一个特征收缩模块,以减少编码器中信息丢失并增强网络的分割能力。在Massachusetts道路数据集上进行测试结果显示,该方法在F1,IoU,Pr和Re指标上分别达到了80.91%,69.40%,78.03%和65.20%。与主流方法UNet和SwinUnet相比,IoU分别提高了4.45%和2.72%,证明了所提算法通过全局建模、上下文增强和信息匹配优化,有效提升了遥感图像道路提取的精度和性能。To address the issues of local feature loss and low extraction accuracy faced by deep neural networks in remote sensing image road extraction,a multi-scale context-aware network was proposed based on the SwinUnet network for remote sensing image road extraction.Firstly,a branch with a context aggregation module was designed in the encoder to enhance the extraction of contextual information and alleviate the problem of semantic ambiguity caused by occlusion.Secondly,to solve the problem of semantic information mismatch between the encoder and decoder and to improve the model's ability to extract spatial information,a spatial feature extraction module was introduced in the skip connections,replacing the direct copying of encoder features in SwinUnet.Finally,a feature compression module was designed in the down-sampling stage to reduce information loss in the encoder and enhance the network's segmentation capability.The test results on the Massachusetts road dataset show that this method achieved F1,IoU,Pr,and Re scores of 80.91%,69.40%,78.03%,and 65.20%,respectively.In comparison with mainstream methods such as UNet and SwinUnet,the IoU improved by 4.45%and 2.72%,respectively,demonstrating that the proposed algorithm effectively improves the accuracy and performance of remote sensing image road extraction through global modeling,context enhancement,and information matching optimization.

关 键 词：遥感图像 道路提取 语义分割 SwinUnet 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]