无线电工程2025，Vol.55Issue(3)：475-483,9.DOI:10.3969/j.issn.1003-3106.2025.03.003
一种改进DeepLabV3+的SAR图像建筑分割方法
An Architectural Segmentation Method of SAR Images Based on Improved DeepLabV3+
摘要
Abstract
Compared with optical images,Synthetic Aperture Radar(SAR)images have certain advantages such as penetration ability and all-weather continuous monitoring ability,which are suitable for more applications.Building segmentation images are important for urban planning,environmental monitoring and disaster assessment.For the problem of insufficient feature extraction ability and low segmentation accuracy of building segmentation algorithms in SAR images,an improved semantic segmentation model CFNet for DeepLabV3+is proposed.CFNet first modifies the backbone network Xception of traditional DeepLabV3+to be MobileNetV2 backbone network in order to reduce the number of model parameters and improve computing speed;secondly,a new cross-attention mechanism that combines the channel attention mechanism and the spatial attention mechanism is proposed in order to extract shallow and deep features;finally,the fusion of shallow and deep features extracted from the network is improved,and shallow and deep features are introduced as auxiliary features for fusion,which makes the best use of shallow and deep features in the network.The shallow and deep features in the network are utilized to enhance feature extraction capability of the algorithm.Experimental results on the SARBuD 1.0 dataset show that the mean Intersection over Union(mIoU)of CFNet is 80.69%,the precision rate is 87.99%,the recall rate is 92.05%,and the F1 factor is 89.86%,which shows that CFNet has a certain increase in precision of architectural segmentation of SAR images compared with other segmentation networks.关键词
DeepLabV3+模型/合成孔径雷达图像/深度学习/语义分割/特征融合Key words
DeepLabV3+model/SAR images/deep learning/semantic segmentation/feature fusion分类
计算机与自动化引用本文复制引用
张文武,龙伟军,陈虹廷,陈逸飞..一种改进DeepLabV3+的SAR图像建筑分割方法[J].无线电工程,2025,55(3):475-483,9.基金项目
国家自然科学基金(62071440)National Natural Science Foundation of China(62071440) （62071440）
