HFRA-Net:基于混合特征和残差注意力的遥感影像建筑物提取方法  

HFRA-Net:a method for hybrid feature fusion and residual attention building extraction from remote sensing images

作　　者：薛辉 徐雯佳[2] 牛棚辉 XUE Hui;XU Wenjia;NIU Penghui(The 54th Research Institute of CETC,Shijiazhuang Hebei 050081,China;Hebei Prospecting Institute of Hydrogeology and Engineering Geological(Hebei Remote Sensing Center),Shijiazhuang Hebei 050021,China;School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]中国电子科技集团公司第五十四研究所,河北石家庄050081 [2]河北省水文工程地质勘查院(河北省遥感中心),河北石家庄050021 [3]河北工业大学人工智能与数据科学学院,天津300401

出　　处：《河北省科学院学报》2024年第4期31-39,共9页Journal of The Hebei Academy of Sciences

基　　金：中国电子科技集团公司第五十四研究所空间信息综合应用平台建设及应用示范项目(SYX20103T001);河北省自然资源科技项目(454-0601-YBN-IBBM)。

摘　　要：针对基于深度学习的遥感影像建筑物提取方法存在覆盖范围大的建筑物信息提取不完整和部分边界细节信息的丢失的问题,本文提出了一种新的采用混合特征融合和残差注意力机制的网络方法HFRA-Net。该方法引入混合特征融合模块,以并行结构捕获全局和局部特征,提高建筑物分割的完整性。同时,通过引入自上而下和自下而上双向注意力反馈的残差注意力机制,自适应地捕获不同尺度下的边界细节信息。最后,在两个公共数据集上的实验结果证明了提出方法对建筑物提取的有效性。Building extraction from remote sensing images based on deep learning faces two challenges:incomplete extraction with large coverage and the loss of partial boundary details.To address these issues,a novel network model HFRA-Net using hybrid feature fusion and residual attention mechanism is proposed.The model introduces a hybrid feature fusion module to capture global and local features in a parallel structure,enhancing the integrity of building segmentation.At the same time,the residual attention mechanism of top-down and bottom-up bidirectional attention feedback is introduced to adaptively capture the boundary detail information at different scales.Finally,experimental results on two public datasets demonstrate the effectiveness of the proposed method for building extraction.

关 键 词：遥感影像 建筑物提取 特征融合 残差注意力 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]