融合注意力和扩张卷积的遥感影像道路信息提取方法  

Road Information Extraction Method of Remote Sensing Image byCombining Attention and Extended Convolution

作　　者：肖振久[1] 郝明 曲海成[1] 侯佳兴 XIAO Zhenjiu;HAO Ming;QU Haicheng;HOU Jiaxing(School of Software,Liaoning Technical University,Huludao,Liaoning 125015,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125015

出　　处：《遥感信息》2024年第1期18-25,共8页Remote Sensing Information

基　　金：辽宁省高等学校基本科研项目(LJKMZ20220699);辽宁工程技术大学学科创新团队项目(LNTU20TD-23)。

摘　　要：针对高分辨率遥感影像语义分割存在地物边缘分割不连续、道路及背景特征复杂多样导致道路提取分割精度不高的问题,提出了一种融合双通道注意力和扩张卷积的遥感影像道路信息提取语义分割网络(A 2DU-Net)。首先,在特征提取部分引入坐标注意力(coordinate attention,CA)模块,捕捉道路位置、方向和跨通道信息,精确定位道路信息。其次,针对网络对细节特征丢失的敏感问题,在编码器的末端利用不同扩张率的空洞卷积构建多尺度特征融合的空洞空间金字塔池化模块(multi-scale Atrous spatial pyramid pooling module,MASPPM)来获得更大的感受野,提高网络性能。最后,为了避免U-Net中纯跳跃连接在语义上不相似特征的融合,在编码器和解码器的跳跃连接之间增加了双通道注意力机制来实现门控筛选,抑制非目标区域的特征,提高网络的分割精度。实验在公共道路数据集Massachusetts上对网络模型进行测试,OA(准确率)、交并比(IoU)、平均交并比(mIoU)和F1等评价指标分别达到98.07%、64.39%、81.20%和88.67%。与主流方法U-Net和DDUNet进行比较,mIoU分别提升了3.07%、0.22%,IoU分别提升了1.98%、0.52%。实验结果表明,所提出的方法优于所有的比较方法,能够有效提高道路分割的精确度。Aiming at the problem that the semantic segmentation of high-resolution remote sensing images has discontinuous ground edge segmentation as well as the complexity and diversity of road and background features result in low accuracy of road extraction and segmentation,a semantic segmentation network(A 2DU-Net)for road information extraction of remote sensing images integrating dual-channel attention and expansion convolution is proposed.Firstly,the coordinate attention(CA)module is introduced in the feature extraction part to capture road location,direction and cross-channel information to accurately locate road information.Secondly,aiming at the sensitive problem of network loss of detailed features,the multi-scale Atrous spatial pyramid pooling module(MASPPM)of multi-scale feature fusion is constructed by using hole convolution with different expansion rates at the end of the encoder to obtain larger receptive fields and improve network performance.Finally,in order to avoid the fusion of semantically dissimilar features of pure hop connections in U-Net,a dual-channel attention mechanism is added between the hop connections of encoder and decoder to achieve gating screening,suppress the features of non-target regions,and improve the segmentation accuracy of the network.The network model is tested on the public road dataset Massachusetts,and the evaluation indexes such as OA(accuracy),intersection-union ratio(IoU),average intersection-union ratio(mIoU)and F1 reaches 98.07%,64.39%,81.20%and 88.67%,respectively.Compared with mainstream methods such as U-Net and DDUNet,mIoU increases by 3.07%and 0.22%,and IoU increases by 1.98%and 0.52%.Experimental results show that the proposed method is superior to all comparison methods,which can effectively improve the accuracy of road segmentation.

关 键 词：语义分割 道路提取 注意力机制 U-Net 空洞空间金字塔池化 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]