基于UNet模型的遥感影像建筑物变化检测研究
Research on Building Change Detection in Remote Sensing Image Based on UNet Model
王盼盼 1刘超 1孙健飞 2樊亚 3刘佳祥 3董亮2
作者信息
- 1. 安徽理工大学空间信息与测绘工程学院,232001,安徽,淮南
- 2. 江苏省地质矿产局第六地质大队,222000,江苏,连云港
- 3. 中国建筑材料工业地质勘查中心贵州总队,550009,贵阳
- 折叠
摘要
UNet是一个典型的对称U型网络,针对该网络无法准确地捕捉到建筑物的边界及细节信息的问题,提出了一个改进的UNet网络模型,即在UNet网络模型的跳跃连接中加入可以提高网络感知力的scSE注意力模块,同时将模型中的编码器更换为能够更好地捕捉图像细节和纹理的VGG19,对公开数据集LEVIR-CD进行建筑物变化检测实验.实验结果表明,该方法较于原方法虽然精确率下降 0.66%,但是召回率和F1 分数分别提高了 13.18%和 5.17%,说明该方法有效提升了UNet网络模型对建筑物边界及细节的识别,使建筑物变化检测的精度得到有效提升.
Abstract
UNet is a typical symmetric U-shaped network,for the problem that this network cannot accurately capture the boundary and detail information of buildings.In this paper,we propose an improved UNet network model,we add the scSE attention module,which can improve the network perception,to the jump link of UNet network model,and at the same time,we replace the encoder in the model with VGG19,which is able to better capture the image details and textures,to conduct building change detection experiments on the publicly available dataset LEVIR-CD.The experimen-tal results show that the method improves the recall by 13.18%and F1 by 5.17%compared to the original method although the precision rate decreases by 0.66%.This shows that the method effec-tively improves the detection of building boundaries and details by the UNet network model,so that the precision of building change detection is effectively improved.
关键词
建筑物变化检测/注意力机制/UNet/遥感影像/编码器Key words
building change detection/attention mechanism/UNet/remote sensing image/encoder引用本文复制引用
基金项目
安徽省高等学校科研项目(2022AH050849)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据