基于特征注意力融合元残差网络的小样本SAR目标识别

Few-Shot SAR Target Recognition Method Based on Feature Fusion Attention and Meta-ResNet

深度学习方法在合成孔径雷达(Synthetic Aperture Radar,SAR)目标识别方面表现良好,但是这些深度模型往往需要大量训练数据来对模型参数进行优化,否则经常会遇到严重的过拟合问题,导致目标识别准确率低,模型泛化能力差.针对上述问题,基于元学习框架,提出了针对小样本条件下SAR目标识别的特征注意力融合元残差网络.在该方法中,基于残差网络设计了全新的学习器,通过利用残差结构,有效传递SAR图像的对比度信息,提高目标识别准确率,而且多尺度特征注意力融合模块通过注意力机制,融合不同残差层特征,为目标识别提供更丰富的特征信息.改进的元学习器不仅能够为学习器学习到易于优化的初始化参数,而且能够为学习器的每一个参数学习一个不同但是合适的学习率.与其他三种小样本目标识别方法在MSTAR数据集上进行对比实验,结果表明本文方法提高了小样本条件下SAR目标识别方法的识别准确率和识别模型的泛化性能.对所提方法进行了鲁棒性验证实验,网络结构消融实验,并展示了元学习器为学习器参数学习到的不同但是合适的学习率.

In recent years,traditional synthetic aperture radar(SAR)target recognition methods based on deep learning have achieved promising results.However,these deep models need lots of training samples for parameter optimization,otherwise they would possibly encounter serious overfitting problem,resulting in low recognition accuracy and poor generalization ability of these deep models.Aiming to above problems,this paper proposed a few-shot SAR target recognition method named multi feature attention fusion Meta-ResNet.In this method,we designed a learner based on residual network,which can effectively transmit contrast information in SAR images and thus improve recognition accuracy.Furthermore,the multi feature attention fusion module can provide weighted multi scale features by fusing features of different ResNet layers.The improved meta-learner can not only learn good initialization parameters for learner,but also learn a different but appropriate learning rate for each learner parameter.The comparative experiment between the proposed method and other three few-shot recognition methods demonstrated the effectiveness and progressiveness of the proposed method.We also conducted experiment to verify the robustness and study the influence of the network structure on recognition accuracy.We showed the different but appropriate learning rates of learner parameters learned by the meta-learner.