摘要
自动目标识别(ATR)作为合成孔径雷达(SAR)图像解译的重要手段而备受关注。由于不同方位角域下的SAR目标散射特性分布差异大,导致SAR目标图像特征对方位角高度敏感,且雷达难以在单次观测中捕获目标所有方位角域下的数据,基于历史数据训练的SAR‐ATR模型在新观测方位角域数据上识别性能下降。当新观测数据以流的形式到达时,若仅依赖新观测数据对现有模型进行再训练,容易引发“灾难性遗忘”问题。因此,本文通过引入Fisher信息矩阵调节的正则项来保护对识别任务贡献大的模型参数,并利用核心集减小推理误差,构建一种方位角自适应SAR目标识别连续学习模型。实验结果表明,方位角自适应SAR‐ATR模型能够在线学习不同方位角下SAR目标数据,不断适应其特征变化,有效提高了其对未观测方位角域数据的泛化性。
Automatic target recognition(ATR),as an important means of synthetic aperture radar(SAR)image interpretation,has attracted much attention.Due to the large difference in the scattering characteristics distribution of SAR targets in different azimuth domains,the image characteristics of SAR targets are highly sensitive to the azimuth angle.Moreover,it is difficult for radar to capture all the azimuth domain data of the target in a single observation,and the recognition performance of the SAR-ATR model trained on past data is reduced on the newly observed azimuth domain data.When the new observation data is arrived in the form of a stream,it is easy to cause the problem of“Catastrophic Forgetting”if the existing model is retrained only by relying on the new observation data.Therefore,in this paper,the regularization term adjusted by Fisher information matrix is used to protect the model parameters that contribute greatly to the recognition task,and the core set is used to reduce the inference error,so as to construct an azimuth adaptive continuous learning model for SAR target recognition.The experimental results show that the azimuth adaptive SAR-ATR model can learn the SAR target data under different azimuth angles online,continuously adapt to its feature changes,and effectively improve the generalization to the unobserved azimuth domain data.
作者
陈虹廷
武凡
杜川
龙伟军
CHEN Hongting;WU Fan;DU Chuan;LONG Weijun(Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《雷达科学与技术》
北大核心
2025年第2期167-175,共9页
Radar Science and Technology
基金
国家自然科学基金(No.62071440)。
关键词
合成孔径雷达
自动目标识别
方位角域
连续学习
Fisher信息矩阵
synthetic aperture radar
automatic target recognition
azimuth domain
continuous learning
Fisher information matrix
