嵌入注意力机制的深度可分离卷积SAR目标识别  

Depthwise Separable Convolutional SAR Target Recognition Embedded with Attention Mechanism

作　　者：卢小华[1] 李爱军[2] LU Xiaohua;LI Aijun(School of Computer Information Engineering,Shanxi Technoloy and Business College,Taiyuan 030006,China;School of Information,Shanxi University of Finance and Economics,Taiyuan 030006,China)

机构地区：[1]山西工商学院计算机信息工程学院,山西太原030006 [2]山西财经大学信息学院,山西太原030006

出　　处：《无线电工程》2024年第5期1083-1090,共8页Radio Engineering

基　　金：山西省教育厅教学改革创新项目(J2021865)。

摘　　要：深度可分离卷积(Depthwise Separable Convolution,DSC)的应用使得深度学习的网络模型轻量化。在此基础上,提出了嵌入注意力机制的DSC合成孔径雷达(Synthetic Aperture Radar,SAR)目标识别方法。通过将DSC与注意力机制结合,提高网络对目标重要特征的学习能力;将多个DSC进行叠加和并联,设计多尺度网络模块,增强不同深度网络的特征提取能力;通过残差连接缓解深层网络的梯度弥散和梯度爆炸问题。使用公开数据集实验表明,所提方法在网络模型参数量较小的情况下,获得99.0%的平均识别率,具有较强的识别优势。The application of Deep Separable Convolution(DSC)makes the deep learning network model lightweight.On this basis,a DSC Syntheic Aperture Radar(SAR)target recognition method embedded with attention mechanism is proposed.By combining the depth separable convolution with the attention mechanism,the ability of network learning the important target features is improved;At the same time,multiple depth separable convolutions are superposed and paralleled,and multi-scale network modules are designed to enhance the feature extraction capability of different depth networks;Finally,residual connection is used to alleviate gradient dispersion and gradient explosion of deep network.Experiments show that the proposed method achieves an average recognition rate of 99.0%under the condition of small network model parameters,and has strong recognition advantages.

关 键 词：合成孔径雷达 目标识别 深度可分离卷积 注意力机制 

分 类 号：TN957[电子电信—信号与信息处理]