基于层级尺度交互的U-Net遥感影像建筑物提取方法  

Hierarchical-scale Interaction-based U-Net for Remote Sensing Image Building Extraction

作　　者：余快 宋宝贵 邵攀 余翱 YU Kuai;SONG Baogui;SHAO Pan;YU Ao(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering(China Three Gorges University),Yichang Hubei 443002,China;College of Computer and Information,China Three Gorges University,Yichang Hubei 443002,China)

机构地区：[1]水电工程智能视觉监测湖北省重点实验室(三峡大学),湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《广西师范大学学报(自然科学版)》2025年第2期121-132,共12页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金青年项目(41901341);湖北省自然科学基金(2024AFB867)。

摘　　要：针对U-Net及其改进网络在跳跃链接中因忽略多层级特征间相互作用而导致对特征的表征能力不足问题,本文提出一种基于层级尺度交互的U-Net遥感影像建筑物提取方法。首先,在U-Net网络跳跃连接中设计层级尺度交互模块,实现多层级特征的交互增强,提升对特征的表征能力;然后,通过改进空洞空间金字塔池化模块,提出一种多尺度特征提取模块,并将其应用到最高层级特征,来提升网络提取多尺度特征的能力;最后,将自校准卷积引入到解码过程,促进浅层与深层特征更好地融合。在公开建筑物提取数据集WHU和Inria上,将本文方法与6种遥感影像建筑物提取方法进行对比,实验结果表明,本文方法的IoU分别为91.26%和79.23%,均优于对比方法。Aiming at the problem that U-Net and its improved network have insufficient feature characterisation ability due to ignoring the interactions between multi-level features in jump links,a building extraction method based on hierarchical scale interactions is proposed for U-Net remote sensing images.Firstly,a hierarchical scale interaction module is designed in the jump link of U-Net network to achieve the interaction enhancement of multilevel features and improve the characterisation ability of features.Then a multi-scale feature extraction module is proposed by improving the null-space pyramid pooling module and applying it to the highest level features to enhance the ability of the network to extract multi-scale features.Finally,self-calibrating convolution is introduced into the decoding process to promote better fusion of shallow and deep features.The method of this paper is compared with six remote sensing image building extraction methods on two publicly available building extraction datasets,WHU and Inria.The experimental results show that the IoU of the proposed method is 91.26%and 79.23%,respectively,which are better than the comparison methods.

关 键 词：遥感影像 建筑物提取 U-Net 层级尺度交互 多尺度 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]