当采用高分辨雷达对空间微动目标进行观测时,往往能同时获得其窄带、宽带回波.为充分利用其中蕴含的丰富电磁散射、形状、结构及运动信息,该文提出基于稀疏自编码器(SAE)的空间微动目标特征级融合识别方法.在训练阶段,首先采用卷积神经网络(CNN)分别提取训练集中微动目标回波的1维高分辨距离像(HRRP)、时频图(JTF)及距离-瞬时多普勒像(RID)层级特征.随后,将提取的3个深层特征进行1维拼接形成联合特征向量,并采用SAE自动学习联合特征向量的隐层特征.进而剔除SAE解码部分并在编码器后接入Softmax分类器构成识别网络.最后,利用SAE网络参数对识别网络进行初始化,并利用上述联合特征向量对其进行微调得到训练好的识别网络.在测试阶段,将CNN所提测试集的联合特征向量直接输入训练好的识别网络以得到融合识别结果.不同条件下的电磁仿真数据识别结果证明了所提方法的有效性及稳健性.
During the observation of micro-motion targets in space,high resolution radar collects the narrowband and wideband echoes simultaneously.This paper proposes a fusion method based on a Sparse Auto-Encoder(SAE)for recognizing space micro-motion targets to exploit their rich electromagnetic scattering,shape,structure,and motion information.In the training phase,the proposed method extracts the hierarchical features from High Resolution Range Profiles(HRRP),Joint Time-Frequency(JTF)images,and Range-Instantaneous-Doppler(RID)images using Convolution Neural Networks(CNN).The joint feature vector is then created by concatenating the relevant deep features,and SAE learns autonomously its hidden features unsupervised.After that,the decoder is removed and the Softmax classifier is introduced after the encoder to create the recognition network.Finally,parameters of the optimized SAE network are used for the initialization of the recognition network,which is then fine-tuned by the joint feature vectors of training samples.In the test phase,the trained recognition network is supplied directly with the joint feature vectors of the test samples recovered by CNN to produce the fusion recognition results.Experimental results of simulated EM data under different conditions show the efficacy and robustness of the proposed method.