基于混合核深度适配网络的HRRP目标识别  

HRRP Target Recognition Based on Mixed Kernel Deep Adaptation Networks

作　　者：王国帅 汪文英[1] 魏耀[1] 郑玄玄 雷志勇 WANG Guoshuai;WANG Wenying;WEI Yao;ZHENG Xuanxuan;LEI Zhiyong(Nanjing Research Institute of Electronics Technology, Nanjing 210039, China)

机构地区：[1]南京电子技术研究所,江苏南京210039

出　　处：《雷达科学与技术》2020年第6期618-624,共7页Radar Science and Technology

摘　　要：为了改善基于卷积神经网络(CNN)的雷达目标识别模型的泛化能力,本文将深度适配网络(DAN)方法引入到高分辨一维距离像(HRRP)目标识别中。为了进一步提高DAN方法的性能,创新性地提出了混合核函数MMD代替传统DAN中的多核MMD,设计了基于混合核函数的MMD损失函数。本文使用服从瑞利分布的海杂波来干扰目标域数据。网络模型中使用一维CNN提取特征,使用混合核函数DAN来减少源域和目标域间特征分布的差异。实验表明,相对常规迁移学习方法和DAN方法,在海杂波影响下该方法可将目标域数据识别率提高15%左右,显著提高了模型的泛化性能和鲁棒性。In order to improve the generalization performance of radar target recognition model based on convolutional neural network(CNN),this paper introduces the deep adaptation networks(DAN)method into the high resolution range profile(HRRP)target recognition.In order to further improve the performance of DAN method,the mixed kernel MMD is proposed to replace multi-kernel MMD(MK-MMD)in DAN,and the MMD loss function based on the mixed kernel function is designed.In this paper,the sea clutter obeying Rayleigh distribution is used to interfere with target domain data.In the network model,one-dimensional CNN is used to extract features.The mixed kernel function DAN is used to reduce the difference of feature distribution between source domain and target domain.The experimental results show that compared with the conventional transfer learning method and DAN method,this method can improve the recognition rate of target domain data by about 15%under the influence of sea clutter.It greatly improves the generalization and robustness of the model.

关 键 词：深度迁移学习 目标识别 高分辨一维距离像 混合核深度适配网络 

分 类 号：TN957.51[电子电信—信号与信息处理] TP183[电子电信—信息与通信工程]