小样本条件下异源图像迁移学习的红外目标检测与识别  

Infrared target detection and recognition based on transfer learning with small samples

作　　者：龚任杰 郑智辉 丛龙剑 徐振涛 韦海萍[1] 唐波[1] 李全运[1] GONG Renjie;ZHENG Zhihui;CONG Longjian;XU Zhentao;WEI Haiping;TANG Bo;LI Quanyun(Beijing Aerospace Automatic Control Institute,Beijing 100854,China)

机构地区：[1]北京航天自动控制研究所,北京100854

出　　处：《西北工业大学学报》2021年第S01期84-88,共5页Journal of Northwestern Polytechnical University

摘　　要：针对小样本条件下红外目标检测任务中缺少足够的训练样本,网络泛化能力不理想的问题,提出了一种基于迁移学习的红外目标检测算法。论证了使用可见光图像通过迁移学习技术训练红外图像目标检测算法的可行性。设计了一种基于注意力机制的域自适应方法的深度神经网络,通过可见光数据迁移学习实现小样本条件下的红外目标检测与识别。通过VisDrone2019数据集、Street Scece红外数据集进行验证。结果表明:所提算法实现小样本条件下高精度的红外目标检测与识别。In infrared target detection and recognition,aiming at the problem of the unsatisfactory of network generalization capability due to the limited number of infrared samples,this thesis proposes an infrared target detection algorithm based on transfer learning.Firstly,this thesis demonstrates the feasi-bility of the infrared target detection using transfer learning with visible images.In order to achieve the goals of infrared target detection with small samples,the algorithm designs a deep neural network based on domain adaptation.The verification of the VisDrone2019 UAV data sets and Street Scene infrared images data sets shows that it implements high-precision infrared target detection algorithms to complete tasks with small samples.

关 键 词：红外目标检测 异源图像 迁移学习 注意力机制 泛化能力 域自适应 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]