基于仿真SAR图像深度迁移学习的自动目标识别  

Study of deep transfer learning for SAR ATR based on simulated SAR images

作　　者：王泽隆 徐向辉[1] 张雷 WANG Zelong;XU Xianghui;ZHANG Lei(Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China;University of Chinese Academy of Sciences, Beijing 100049, China;Unit 95899 of PLA, Beijing 100076, China)

机构地区：[1]中国科学院电子学研究所,北京100190 [2]中国科学院大学,北京100049 [3]中国人民解放军95899部队,北京100076

出　　处：《中国科学院大学学报（中英文）》2020年第4期516-524,共9页Journal of University of Chinese Academy of Sciences

基　　金：国家重点研发计划(2017YFB0503001)资助。

摘　　要：使用深度卷积神经网络实现SAR图像的自动目标识别在训练过程中需要大量的标注数据。为解决由SAR实测数据获取成本高、标注数据量不足带来的问题,提出一种在由CReLU激活函数和批归一化改进的卷积神经网络上,使用仿真SAR图像提升最终目标识别性能的方法,把从大量仿真SAR图像学习到的有效知识迁移到实测SAR图像数据上。在训练中,先用仿真SAR图像预训练卷积神经网络,结合迁移学习的方法,有效地解决由SAR图像数据不足带来的过拟合问题。在MSTAR数据集上验证方法的有效性,识别准确率提高到99.78%,并在少量SAR图像样本数据上也取得不错的识别效果。Using deep convolutional neural networks to realize automatic target recognition of SAR requires a large amount of labeled data.In order to solve the problem caused by the scarcity of SAR real images,we propose a method for improving the target recognition performance of SAR by using simulated SAR images on convolutional neural networks improved by CReLU activation function and batch normalization.The method transfers the effective knowledge learned from a large number of simulated SAR images onto the real SAR images.In the training,the pre-trained convolutional neural networks can be obtained by training by using the simulated SAR images firstly,and the deep transfer learning method is used to effectively solve the problem caused by the insufficiency of SAR image data.The validation experiment is carried out on the MSTAR dataset.The highest recognition accuracy reaches 99.78%,and good recognition results are obtained based on a small amount of SAR image data.

关 键 词：合成孔径雷达 仿真SAR图像 迁移学习 自动目标识别(ATR) 

分 类 号：TN957[电子电信—信号与信息处理]