基于迁移学习的小样本SAR船只目标识别  

The Target Recognition of Small Sample SAR Ship Based on Transfer Learning

作　　者：王程丽 赵静[1] 杨攀攀 李姗 WANG Chengli;ZHAO Jing;YANG Panpan;LI Shan(PLA 92020 Group,Qingdao 266000,China;Systems Engineering Research Institute,China State Shipbuilding Corporation Limited,Beying 100094,China)

机构地区：[1]中国人民解放军92020部队,山东青岛266000 [2]中国船舶工业系统工程研究院,北京100094

出　　处：《移动通信》2022年第4期22-27,共6页Mobile Communications

摘　　要：合成孔径雷达(SAR)目标识别的主流手段是神经网络,其依赖于大数据量的训练,但SAR船只样本量少,且传统识别方法提取的特征又具有很强的易变性,分类效果不佳。针对SAR船只目标样本量受限的问题,提出了基于VGG16迁移学习的识别方法,该方法在已有模型的基础上进行参数的微调,使其适应目标数据集,从而解决在训练样本缺失情况下,识别过程中存在的过拟合和局部最优解等问题。利用Terra SAR数据库进行对比实验,结果表明该方法优于传统识别方法。The mainstream method of synthetic aperture radar(SAR)target recognition is neural network,which depends on the training of large amount of data.However,due to the small sample size of ships,the features extracted by the traditional recognition methods are highly variable and the classification effect is poor.Aiming at the problem of limited sample size of SAR ship targets,a target recognition method based on VGG16 transfer learning is proposed.This method slightly adjusts the parameters based on the existing models to be adapted to the target data set,and thus solves the problems of over fitting and local optimal solution in the recognition process when the training samples are missing.The comparative experiments using TerraSAR database show that the method is superior to traditional recognition methods.

关 键 词：小样本 SAR船只识别 迁移学习 VGG16 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]