摘要
为了解决深度学习方法网络模型在小样本遥感图像目标识别场景下过拟合、性能急剧下降等问题,设计了基于度量学习的小样本目标识别方法RS-DN4。在度量模块中选择前k个具有代表性的特征进行目标相似度的计算;引入元学习中阶段式训练理念,执行上万次任务级迭代训练;基于互联网开源遥感影像数据集和自有遥感影像数据集,构建了一套包含21类不同遥感目标切片数据的多尺度、多分辨率的遥感影像数据集RSD-FSC,并据此进行RS-DN4方法的实验验证。结果表明针对新类小样本目标,当训练样本分别只有1,5和10个时,平均识别准确率可分别达到59.13%,82.55%和87.80%,相对于其他方法,RS-DN4在遥感图像小样本目标识别中具有识别准确率高、泛化能力强等优势。RS-DN4方法实现难度适中,在小样本遥感目标检测识别领域的工程应用场景中具有推广应用价值。
In order to solve the problems of overfitting and drastic performance degradation of network models with deep learning methods in few-shot remote sensing image target recognition scenarios,the few-shot target recognition method RS-DN4 was designed based on metric learning.The top k representative features in the metric module were selected to calculate the similarity of the targets,and the concept of meta-learning method for episodic training was introduced to conduct thousands of task-level iterative training.In addition,based on the internet open source remote sensing image dataset and own remote sensing image dataset,a multi-scale and multi-resolution remote sensing image dataset RSD-FSC,which contained sliced data of 21 different classes of remote sensing targets,was constructed.On this basis,the RS-DN4 method was verified by experiments.The results show that:for the new few-shot targets,the average recognition accuracy can reach 59.13%,82.55%and 87.80%respectively when there are only one,five and ten training samples.Compared with other methods,RS-DN4 has the advantages of high recognition accuracy and strong generalization ability in few-shot target recognition of remote sensing images.The RS-DN4 method is moderately difficult to implement and has the value to be promoted in engineering application scenarios in the field of few-shot remote sensing target detection and recognition.
作者
张萌月
陈金勇
王港
王敏
帅通
孙康
ZHANG Mengyue;CHEN Jinyong;WANG Gang;WANG Min;SHUAI Tong;SUN Kang(The 54th Research Institution of CETC,Shijiazhuang,Hebei 050081,China;CETC Key Laboratory of Aerospace Information Applications,Shijiazhuang,Hebei 050081,China)
出处
《河北工业科技》
CAS
2021年第2期116-122,共7页
Hebei Journal of Industrial Science and Technology
基金
中国电子科技集团公司航天信息应用技术重点实验室开放基金(SXX19629X060)。
关键词
模式识别
小样本
遥感目标识别
度量学习
阶段式训练
pattern recognition
few-shot
remote sensing target recognition
metric learning
episodic training
