Research On Aircraft Target HRRP Recognition Based On Dictionary Learning And Convolutional Neural Network

The High Resolution Range Profile(HRRP)of the radar target can represent the radial distribution of the target scattering point along the radar line of sight,and can provide information about the target size,scattering point distribution,etc.,which facilitates target classification.Compared with Synthetic Aperture Radar(SAR)images and Inverse Synthetic Aperture Radar(ISAR)images,high resolution range profile(HRRP)have certain advantages in the way of acquisition and processing.Due to the presence of noise interference in the actual situation,and in the case of non-cooperative targets,the available training data sets are often limited,incomplete,and even some of the target azimuth data is missing,which requires the identification method to have more good generalization performance.In view of the above problems,this paper will study the noise-robust dictionary learning method and the Convolutional Neural Network(CNN)method with azimuth robustness to achieve HRRP target recognition.Due to the noise interference in the actual situation,how to achieve robust identification under low SNR is an important issue in HRRP target recognition.Studies have shown that the Sparse Representation(SR)method is extremely robust to noise and has achieved remarkable results in face classification.In this paper,the sparse representation and solving algorithm of signal are studied,including Matching Pursuit(MP)and Orthogonal Matching Pursuit(OMP),and the K-Singular Value Decomposition(K-SVD)algorithm is studied.Because the dictionary of sparse representation method is artificially designed,the dictionary obtained by adaptive learning by K-SVD method is more in line with the characteristics of the observed signal,which can better characterize the essential structure of the observed signal.Since the K-SVD method only seeks the best characterization of the sample data,although the reconstruction error and the sparsity of the coefficients are explored,it is not explored whether the dictionary has discriminative power.Therefore,the dictionary of K-SVD learning is not optimized for the classification task,resulting in more training of one dictionary for each type of target.Based on this,this paper studies the discriminative dictionary learning,including the Discriminating K-SVD(DK-SVD)and the Label-Consistent K-SVD(LC-KSVD)method.This type of method is to add a category analysis to the training target to explore how to obtain a discernible dictionary during the training process.Among them,the DK-SVD method can learn a single dictionary with both representation and discriminative power,which can solve the problem that the K-SVD method does not have discriminative power.Each atom of the dictionary obtained by the final training of the LC-KSVD method can be associated with a specific class,which is advantageous for improving the classification performance of the dictionary learning method.Because HRRP has the sensitivity of translation,angle,etc.,this has caused problems for target recognition.This paper studies the One-dimensional Convolutional Neural Network(1D-CNN)method for HRRP recognition for this problem.The 1D-CNN method can use the convolution kernel to perform feature extraction,and filter the spatial position information through the pooling operation,thus having a certain spatial transformation invariance.This feature of 1D-CNN can overcome the translational sensitivity of HRRP,and 1D-CNN has the ability to extract deep features,which makes the method robust to the angular sensitivity of HRRP.Experiments show that the 1D-CNN method is robust to target angle changes,and can extract the hierarchical features of HRRP data,which can improve the accuracy of HRRP recognition.