基于属性散射中心卷积核调制的SAR目标识别深层网络  

Deep Network for SAR Target Recognition Based on Attribute Scattering Center Convolutional Kernel Modulation

作　　者：李毅[1] 杜兰[1] 周可儿 杜宇昂 LI Yi;DU Lan;ZHOU Ke’er;DU Yuang(National Key Laboratory of Radar Signal Processing,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学雷达信号处理全国重点实验室,西安710071

出　　处：《雷达学报（中英文）》2024年第2期443-456,共14页Journal of Radars

基　　金：国家自然科学基金(U21B2039)。

摘　　要：卷积神经网络(CNN)的特征提取能力与其参数量有关,一般来说,参数量越多,CNN的特征提取能力越强。但要学好这些参数需要大量的训练数据,而在实际应用中,可用于模型训练的合成孔径雷达(SAR)图像往往是有限的。减少CNN的参数量可以降低对训练样本的需求,但同时也会降低CNN的特征表达能力,影响其目标识别性能。针对此问题,该文提出一种基于属性散射中心(ASC)卷积核调制的SAR目标识别深层网络。由于SAR图像具有电磁散射特性,为了提取更符合SAR目标特性的散射结构和边缘特征,所提网络使用预先设定的具有不同指向和长度的ASC核对少量CNN卷积核进行调制以生成更多卷积核,从而在降低网络参数量的同时保证其特征提取能力。此外,该网络在浅层使用ASC调制卷积核来提取更符合SAR图像特性的散射结构和边缘特征,而在高层使用CNN卷积核来提取SAR图像的语义特征。由于同时使用ASC调制卷积核和CNN卷积核,该网络能够兼顾SAR目标的电磁散射特性和CNN的特征提取优势。使用实测SAR图像进行的实验证明了所提网络可以在降低对训练样本需求的同时保证优秀的SAR目标识别性能。The feature extraction capability of Convolutional Neural Networks(CNNs)is related to the number of their parameters.Generally,using a large number of parameters leads to improved feature extraction capability of CNNs.However,a considerable amount of training data is required to effectively learn these parameters.In practical applications,Synthetic Aperture Radar(SAR)images available for model training are often limited.Reducing the number of parameters in a CNN can decrease the demand for training samples,but the feature expression ability of the CNN is simultaneously diminished,which affects its target recognition performance.To solve this problem,this paper proposes a deep network for SAR target recognition based on Attribute Scattering Center(ASC)convolutional kernel modulation.Given the electromagnetic scattering characteristics of SAR images,the proposed network extracts scattering structures and edge features that are more consistent with the characteristics of SAR targets by modulating a small number of CNN convolutional kernels using predefined ASC kernels with different orientations and lengths.This approach generates additional convolutional kernels,which can reduce the network parameters while ensuring feature extraction capability.In addition,the designed network uses ASC-modulated convolutional kernels at shallow layers to extract scattering structures and edge features that are more consistent with the characteristics of SAR images while utilizing CNN convolutional kernels at deeper layers to extract semantic features of SAR images.The proposed network focuses on the electromagnetic scattering characteristics of SAR targets and shows the feature extraction advantages of CNNs due to the simultaneous use of ASC-modulated and CNN convolutional kernels.Experiments based on the studied SAR images demonstrate that the proposed network can ensure excellent SAR target recognition performance while reducing the demand for training samples.

关 键 词：合成孔径雷达(SAR) 目标识别 卷积神经网络(CNN) 属性散射中心(ASC) 卷积核调制 

分 类 号：TN957.51[电子电信—信号与信息处理]