利用高判别低维深度特征的SAR船只分类  

Classification of SAR ships using depth features of low dimension andhigh discrimination

作　　者：寻兴青 万剑华[1] 郎海涛[2] 徐永杰 刘善伟[1] XUN Xing-qing;WAN Jian-hua;LANG Hai-tao;XU Yong-jie;LIU Shan-wei(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,China;College of Mathematics and Physics,Beijing University of Chemical Technology,Beijing 100029,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580 [2]北京化工大学数理学院,北京100029

出　　处：《计算机工程与设计》2021年第1期212-219,共8页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2017YFC1405600);国家自然科学基金项目(41776182);2017珠海引进创新创业团队基金项目(ZH01110405170027PWC)。

摘　　要：为充分利用深度特征的判别信息,提高船只分类准确率,提出利用低维度高判别的深度特征进行SAR船只分类的方法。采用ImageNet数据库预训练的VGG16卷积神经网络作为特征提取器,提取船只样本的深度特征;对深度特征进行t-SNE可视化,计算每类船只的类间分离度,选择对于每类船只样本来说类间分离度最大的深度特征;对选择的深度特征进行降维,采用基于KNN的级联二分类方法进行船只分类。利用高分辨率SAR船只数据集验证该方法,实验结果表明,相比传统的船只分类方法,其分类性能有明显提高。To make full use of the discrimination information of depth features and improve the accuracy of ship classification,a method of SAR ship classification based on the depth features of low dimension and high discrimination was proposed.The pre-trained VGG16 network in ImageNet database was used as feature extractor to extract the depth feature of ship sample data.The depth feature was visualized by t-SNE,and the separability of each ship type was calculated using the distance between classes.The depth feature with the largest separation degree for each ship type sample was selected.The dimensionality of the selected effective features was reduced,and the cascade two-classification based on KNN was used to classify ships.The proposed method was validated using high resolution SAR ship data set.Experimental results show that the classification performance of the proposed method is significantly improved compared with that of the traditional ship classification method.

关 键 词：卷积神经网络 深度特征 降维 船只分类 合成孔径雷达 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]