摘要
针对深度学习中的有监督学习需要大量的标注数据,提出了一种少量训练样本下的SAR目标识别方法,解决了SAR图像人工标注成本较高、标注样本不足的问题。首先通过构建正负样本对的策略对数据集进行样本扩充,大幅增加数据量;其次,设计了一种基于度量学习和深度学习的孪生卷积神经网络(孪生CNN),用于衡量样本之间的相似概率;然后采用多任务联合学习的方法训练模型,有效缓解了相干斑噪声对SAR图像的影响,降低了噪声过多易引起的过拟合风险;最后,设计了一种基于孪生CNN的识别样本具体类别的加权投票模型。实验采用了MSTAR和OpenSARShip数据集,在小规模训练集上通过上述方法取得了较好的识别效果。
In view of the large amount of labeling data required by supervised learning in deep learning,a SAR target recognition method based on small scale training dataset is proposed in this paper,which solves the problems of high cost of manual labeling and insufficient labeling SAR samples in SAR images.Through the strategy of constructing positive and negative sample pairs,the training set is enlarged and the data volume is greatly increased.Secondly,a siamese CNN method combining metric learning and deep CNN is proposed to measure the similarity probability between samples.Then,a multi-task joint learning for model training is adopted to effectively alleviate the problem,which reduces the risk of over-fitting caused by excessive noise in SAR images.Finally,a weighted voting model based on siamese CNN is designed to infer specific categories of samples.MSTAR and OpenSARShip datasets are used in the experiment,and good recognition results are obtained through the above methods under small dataset.
作者
王博威
潘宗序
胡玉新
马闻
WANG Bowei;PAN Zongxu;HU Yuxin;MA Wen(University of Chinese Academy of Sciences,Beijing 100049,China;Institute of Electronics,Chinese Academy of Sciences,Beijing 100190,China;Key Laboratory of Technology in Geo-Spatial Information Processing and Application System,Chinese Academy of Sciences,Beijing 100190,China)
出处
《雷达科学与技术》
北大核心
2019年第6期603-609,615,共8页
Radar Science and Technology
基金
国家自然科学基金(No.61701478)
