基于通道注意力机制的小样本SAR飞机图像分类方法  

Few⁃shot SAR aircraft image classification method based on channel attention mechanism

作　　者：赵一铭 王佩瑾 刁文辉 孙显[1,2,3,4] 邓波 Zhao Yiming;Wang Peijin;Diao Wenhui;Sun Xian;Deng Bo(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing,100049,China;University of Chinese Academy of Sciences,Beijing,100049,China;Institute of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing,100049,China;Key Laboratory of Network Information System Technology(NIST),Institute of Electronics,Chinese Academy of Sciences,Beijing,100190,China)

机构地区：[1]中国科学院空天信息创新研究院,北京100049 [2]中国科学院大学,北京100049 [3]中国科学院大学电子电气与通信工程学院,北京100049 [4]中国科学院网络信息体系技术重点实验室,北京100190

出　　处：《南京大学学报（自然科学版）》2024年第3期464-476,共13页Journal of Nanjing University（Natural Science）

基　　金：科技创新2030“新一代人工智能”重大项目(2022ZD0118402);中国人工智能学会-华为MindSpore学术奖励基金。

摘　　要：合成孔径雷达(Synthetic Aperture Radar,SAR)以其全天候、全天时、高分辨率、大幅宽的特点,成为对地观测的重要手段,图像分类是SAR图像解译的一个重要方向.和光学图像相比,SAR图像的成像机理较复杂,存在较多噪声干扰,导致图像清晰度较差、样本标注的难度大,无法保证深度学习算法对样本量的需求,因此,对小样本SAR图像进行图像分类成为当前SAR图像解译领域的重点研究问题之一.基于这一问题展开了基于元学习的SAR图像分类模型的研究,以实现小样本条件下SAR图像的高精度识别.构建基于注意力机制的原型网分类方法,设计了通道注意力模块来自动获取图像特征的重要程度,促进提取对图像分类更有判别力的特征;同时,对模型设计预训练网络,以充分利用已有数据的信息,学习更好的先验信息,提高分类的准确率.在自建的高分辨率SAR图像数据集上对该小样本分类模型进行了实验.消融实验表明,注意力模块和预训练模块对模型的性能均有一定的提升效果.通过对比实验,证明和当前常用的小样本学习方法相比,构建的分类方法能在SAR图像分类中获得较高的准确率,在第一组实验的5-way 1-shot实验中得到的分类精度提高了5.9%,在5-way 5-shot实验中提高了1.92%.Synthetic Aperture Radar(SAR)has become an important device in earth observation because of its all⁃weather and all⁃time service,high resolution and wide width,and image classification is an important direction of SAR image interpretation.Compared with the optical image,the imaging mechanism of the SAR image is more complex.There are more noise interference,resulting in poor image clarity and difficulty in sample labeling,which can not guarantee the sample size requirements of the depth learning algorithm.In this context,how to classify few⁃shot SAR images has become one of the key research issues in the field of SAR image interpretation.To solve this problem,this paper carries out the research of SAR image classification model based on meta⁃learning,hoping to achieve high⁃precision recognition of SAR images under the condition of few⁃shot.A prototypical net classification method based on attention mechanism is constructed,and the importance of automatic acquisition of image features by channel attention module is designed to promote the extraction of features that are more discriminative to image classification.At the same time,a pretraining network is designed for the model to make full use of the information of existing data and learn better priori information,so as to improve the accuracy of classification.Experiments are carried out on the few⁃shot classification model on the self⁃built high⁃resolution SAR image dataset.The ablation experiment shows that both the attention module and the pretraining module improve the performance of the model to a certain extent.Experimental results show that compared with the commonly used few⁃shot learning methods,the classification method constructed in this paper achieves higher accuracy in SAR image classification,the classification accuracy of the 5⁃way 1⁃shot experiment in the first group is improved by 5.9%,and the classification accuracy of the 5⁃way 5⁃shot experiment is improved by 1.92%.

关 键 词：SAR图像分类 元学习 小样本学习 通道注意力模块 预训练 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]