摘要
深度神经网络技术在为合成孔径雷达(SAR)自动目标识别领域带来了较高的识别精度的同时,也在持续进行样本训练的过程中产生了灾难性遗忘问题。目前,学界使用增量学习的方法来缓解深度神经网络持续学习过程中的灾难性遗忘问题。增量学习的关键问题在于提取并保留用于区分新类和旧类的特征,该问题也成为增量学习性能提升的主要瓶颈。主流的增量学习方法一般通过筛选并保留一定数量的旧样本,来保留关键的旧类特征。为了进一步提升增量学习方法的性能,增强增量学习的实用性,文中提出了一种新的增量学习样本保留方法,该方法保留的旧样本具有更强的旧类特征代表性;利用了k均值方法选择代表性旧样本,再利用蒸馏损失训练新模型;通过在MSTAR数据集上的实验可知,该方法能够进一步提升神经网络对SAR图像的增量学习能力。
Although deep neural network technology brings high recognition accuracy to the field of synthetic aperture radar image-based automatic target recognition,it has also caused catastrophic forgetting problems during the continuous sample training process.Currently,academia uses incremental learning methods to alleviate the catastrophic forgetting problem in the continuous learning process of deep neural networks.The key problem of incremental learning is to extract and retain the features used to distinguish new classes from old classes,which has also become the main bottleneck for improving incremental learning performance.Mainstream incremental learning methods generally retain key old class features by selecting and retaining a certain number of old samples.In order to further improve the performance of incremental learning methods and enhance the practicality of incremental learning,this paper proposes a new incremental learning sample retention method,which retains old samples with stronger representativeness of old class feature.We utilize the K-means method to select representative old samples,and then use the distillation loss to train the new model.Available from experiments on the MSTAR dataset.Our method can further improve the incremental learning ability of neural networks for SAR images.
作者
胡超
郝明
汪文英
HU Chao;HAO Ming;WANG Wenying(Nanjing Research Institute of Electronics Technology,Nanjing Jiangsu 210039,China)
出处
《现代雷达》
北大核心
2025年第4期46-51,共6页
Modern Radar
关键词
合成孔径雷达
SAR自动目标识别
深度神经网络
增量学习
灾难性遗忘
synthetic aperture radar
SAR image-based automatic target recognition
deep neural network
incremental learning
catastrophic forgetting
作者简介
胡超,男,1997年生,硕士生,研究方向为雷达目标识别;郝明,男,1971年生,硕士,研究员级高级工程师,研究方向为雷达目标识别;汪文英,男,1981年生,博士,研究员级高级工程师,研究方向为雷达目标识别。
