基于CBAM VGG16-UNet语义分割模型的建筑物提取研究
Research on building extraction based on CBAM VGG16-UNet semantic segmentation model
赵兴旺 1吴治国 1刘超 1刘春阳 1陈健1
作者信息
- 1. 安徽理工大学空间信息与测绘工程学院,矿山采动灾害空天地协同监测与预警安徽普通高校重点实验室,矿区环境与灾害协同监测煤炭行业工程研究中心,安徽淮南 232001
- 折叠
摘要
针对在遥感影像建筑物提取中常常出现"漏检""错检""空洞"等问题,提出了融合双注意力机制的CBAM VGG16-UNet网络,用于建筑物提取研究.基于U-Net网络架构,在下采样部分,用VGG16前5个卷积块代替U-Net网络的编码器部分,在上采样的每个特征融合时引入双注意力机制CBAM,并用双线性插值代替U-Net的转置卷积.使用WHU建筑物数据集以及贵阳建筑物数据集进行了模型验证,与Mobile-UNet、U-Net、VGG16-UNet 3种建筑物提取网络模型进行对比分析.实验表明,CBAM VGG16-UNet在WHU建筑物数据集上精准率、召回率、F1-score、IoU 达到了94.90％,95.46％,95.18％,90.80％,在贵阳建筑物数据集上精准率、召回率、F1-score、IoU达到了77.53％,84.46％,80.85％,67.85％,均优于3种对比模型.为解决建筑物提取常见问题提供了新思路,具有一定的工程应用价值.
Abstract
Aiming at the problems of"missed detection","false detection"and"holes"in remote sensing image building extraction,this study proposes a CBAM VGG16-UNet network with dual attention mechanism for building extraction research.In the downsampling part,the first five convolution blocks of VGG16 are used to replace the encoder part of the U-Net network.The dual attention mechanism CBAM is introduced in the fusion of each feature of the upsampling,and the transposition convolution of U-Net is replaced by billinear interpolation.In this study,the WHU building dataset and Guiyang building dataset are used to verify the model,and compared with three building extraction network models such as Mobile-UNet,U-Net and VGG16-UNet.Experiments show that the precision,recall rate,F1-score and IoU of CBAM VGG16-UNet on WHU building dataset reach 94.90％,95.46％,95.18％and 90.80％.On the Guiyang building dataset,the accuracy,recall.F1-score and IoU reach 77.53％,84.46％,80.85％and 67.85％,which are better than the three comparison models.This study provides a new idea for solving common problems in building extraction,which has certain engineering application value.
关键词
U-Net/VGG16/CBAM/建筑物提取/WHU建筑物数据集Key words
U-Net/VGG16/CBAM/building extraction/WHU building data set引用本文复制引用
基金项目
安徽省自然科学基金(2208085MD101)
安徽省自然科学基金(2108085QD171)
安徽省自然科学基金(2108085MD130)
安徽省高等学校科学研究项目(2022AH050849)
出版年
2024
NETL
NSTL
万方数据