结合非局部注意和多层残差的遥感图像建筑物提取方法
Method for Building Extraction from Remote Sensing Images Based on Non-local Attention and Multi-layer Residuals
刘炜清 1贾赫成1
作者信息
- 1. 复旦大学 信息科学与工程学院,上海 200433
- 折叠
摘要
随着城市化和遥感技术的发展,高分辨率遥感图像地物提取任务也越来越具有挑战性.针对现有的方法无法捕捉图像中长距离的空间关系,以及遥感图像存在误检漏检等问题,提出了结合基于非局部注意力的多层残差遥感图像建筑物提取方法(NAMR-Net).在改进后的U-Net的结构基础上,引入了自适应非局部注意力模块(ANAB),以及多层残差学习模块(MRLB).因此,网络可以从不同的卷积层中融合长距离像素间的特征,并通过2阶段的训练,有效地提升建筑物的分割质量,并在2个公开数据集WHU、Massachusetts上进行了实验.结果表明:NAMR-Net可以实现遥感图像中建筑物目标的高质量分割,并优于近年来几种较先进的方法.
Abstract
With the development of urbanization and remote sensing technology,the tasks of extracting objects from high-resolution remote sensing images have become increasingly challenging.To address the limitation in existing methods,e.g.,the inability to capture long-range spatial relationships and false positives and negatives in remote sensing images,in this paper,a method for building extraction from remote sensing images based on non-local attention and milti-layer residuals is proposed,which is also called the non-local attention guided multi-layer residual net(NAMR-Net).Built upon the refined U-Net architecture,the NAMR-Net incorporates an adaptive non-local attention block(ANAB)and a multi-layer residual learning block(MRLB).Consequently,the network can integrate features from distant pixels at different convolutional layers,and effectively enhance the segmentation quality of buildings through a two-stage training process.Experiments are conducted on two publicly available datasets,i.e.,WHU and Massachusetts.The results demonstrate that the NAMR-Net achieves high-quality segmentation of building targets in remote sensing images and outperforms several state-of-the-art methods.
关键词
高分辨率遥感图像/建筑物提取/深度学习/残差学习/非局部注意力Key words
high resolution remote sensing image/building extraction/deep learning/residual learning/non-local attention引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据