基于超像素和全卷积网络的极化SAR图像分类  

PolSAR Image Classification Based on Superpixel and Fully Convolution Network

作　　者：陈彦桥 张小龙[1] 陈金勇[1,2] 高峰[1,2] 柴兴华[1,2] CHEN Yanqiao;ZHANG Xiaolong;CHEN Jinyong;GAO Feng;CHAI Xinghua(The 54th Research Institute of CETC,Shijiazhuang 050081,China;CETC Key Laboratory of Aerospace Information Applications,Shijiazhuang 050081,China)

机构地区：[1]中国电子科技集团公司第五十四研究所,河北石家庄050081 [2]中国电子科技集团公司航天信息应用技术重点实验室,河北石家庄050081

出　　处：《无线电工程》2020年第12期1024-1029,共6页Radio Engineering

基　　金：航天信息综合应用关键技术研究(2019)(SXX19629X060)。

摘　　要：基于深度学习的方法在极化合成孔径雷达(Synthetic Aperture Rader,SAR)图像分类中取得了不错的分类结果,尤其是全卷积网络(Fully Convolution Network,FCN),得益于其端到端、点到点的分类架构,在极化SAR图像分类中有很大的应用前景。之前的基于FCN的极化SAR图像分类方法中,没有使用超像素去修正分类结果,影响了其分类结果的进一步提升。提出了基于超像素和FCN的极化SAR图像分类方法,得到了很好的分类结果。Nowadays,the methods based on deep learning have achieved good results in PolSAR image classification.Especially Fully Convolution Network(FCN),thanks to its end-to-end,point-to-point classification architecture,it is very promising in PolSAR image classification.However,superpixel is not used in the previous image classification methods based on FCN,which limits the further development of their classification results.So the PolSAR image classification based on superpixel and FCN is proposed and the classification results are better.

关 键 词：深度学习 超像素 全卷积网络 极化SAR 图像分类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]