摘要
传统的合成孔径雷达(SAR)图像目标识别算法需要独立设计特征提取算法和分类器,限制了SAR快速拓展到实际应用中。基于卷积神经网络(CNNs)构建一种可直接从输入图像到输出类别的一体化SAR图像目标识别框架,并引入AdaDelta梯度下降优化算法来进行网络优化学习。同时,由于SAR图像获取困难、数量有限,无法保证CNNs网络的大数据量训练样本需求,因此设计了一种基于多样本扩充CNNs的SAR图像目标识别算法。实验证实设计的算法在MSTAR数据集上10类军事目标平均识别率可达97.28%,且对目标平移、旋转、相干斑噪声和目标遮挡具有较强的鲁棒性。
Conventional synthetic aperture radar(SAR)image target recognition algorithms need to design feature extraction algorithms and classifiers independently, which limits the rapid development of SAR to practical applications.In this paper,an integrated SAR image target recognition algorithm model based on the convolutional neural networks(CNNs)that can directly transfer the input image to the output category has been successfully constructed,and AdaDelta gradient descent optimization algorithm is introduced for network optimization learning.Due to the difficulty of acquiring SAR images and the limited sizes of training samples,the requirements for large data CNNs networks is unable to guarantee.Therefore,a SAR image target recognition algorithm based on multi-sample expansion CNNs is designed.Experiments show that the proposed algorithm is effective in MSTAR dataset.The average recognition rate of the ten military targets is up to 97.28%,and it has strong robustness to the target translation,rotation,speckle noise and target occlusion.
作者
张笑
刘文波
Zhang Xiao;Liu Wenbo(College of Automation Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 211100,China)
出处
《电子测量技术》
2018年第14期92-96,共5页
Electronic Measurement Technology
基金
国家自然科学基金(61471191)
航空科学基金(20152052026)
南京航空航天大学研究生创新基地(实验室)开放基金(kfjj20170313)
中央高校基本科研业务费专项资金
江苏省研究生科研与实践创新计划(KYCX17_0249)项目资助
关键词
合成孔径雷达
卷积神经网络
目标识别
多样本扩充
synthetic aperture radar
convolution neural networks
target recognition
multi-sample expansion