Research On The Local Property Based MMW HRRP Recognition Methods

Radar automatic target recognition (RATR) technique is one of the key technologies in the application of target detection, accurate guidance and other fields. As an important signal in RATR. high-resolution range profile (HRRP) represents the projection of the target scattering centres on the radar line-of-sight. It is easier for a millimeter wave (MMW) radar to transmit a wide-band signal. This improves the range resolution and more details in the target can be obtained. Thus, the result of the recognition can be more accurate. However, HRRP is influenced by the radar parameters, the target state, the background, the weather and other factors. HRRP shows highly nonlinear and the recognition rate of the traditional linear methods is not satisfactory.Manifold learning is a kind of nonlinear dimensionality reduction technology which has been studied widely. The low dimensional feature structure can be obtained from the high dimensional nonlinear space by manifold learning. In this paper, we consider the problem of manifold learning on the basis of the ground target MMW HRRP recognition. Issues on feature selection, design of classifier, active learning and imbalanced learning have been studied. The main achievements of this work are listed as follows.We studied the problem of feature selection from the view of algorithm and proposed three feature selection algorithms, called local reconstruction error alignment (LREA), label reconstruction based Laplacian score (LRLS) and modified constraint score (MCS) respectively, for the application of non-supervised feature selection and semi-supervised feature selection (the supervised information are label and pairwise constraints, respectively). LREA can be considered as the feature selection version of locally linear embedding. The optimal local features can be obtained by minimizing the local reconstruction error. Then, the alignment technique is applied to extend the local optimal feature sequence to a global unique feature sequence. LRLS utilizes the label reconstruction technique to estimate the labels of the unlabeled samples and extends the Laplacian score (LS) to the case of semi-supervised learning. For the better description of the similarity between two samples in the non-linearity space, we substitute the geodesic distance for the Euclidean distance. In MCS, it is assumed that the pairwise constraints and the local property of samples are not completely independent. The pairwise constraints can be utilized to improve the local property. The improved local property and the pairwise constraints are used to perform semi-supervised feature selection.To overcome the target-aspect sensitivity problem in HRRP recognition, we proposed a novel classification algorithm, called geodesic weighted sparse representation (GWSR). It is assumed that the L2-normalized HRRP samples from the same target lie on the same submanifold which is embedded in a unit hypersphere. These submanifolds can be classified by using the property that HRRP in a small aspect range has a high correlation. In GWSR, modified geodesic distance is calculated to measure the similarity among all samples. Then, geodesic weighted samples are calculated. Geodesic weighted samples unfold the submanifold that lies on the hypersphere. The nonlinear structure of HRRP can be transformed into a linear one through this step. Finally, all the labeled samples are used as a dictionary and the unlabeled HRRP is sparsely reconstructed. The sparse reconstruction weight can be utilized to estimate the label information of the unlabeled HRRP.In the traditional recognition methods, the training set is obtained by random selection. For the same classifier model, different parameters will be obtained if the training sets are different. The performance of these classifiers with different parameters may also be different. Active learning endeavors in searching the optimal training subset from the given training set. An optimal classifier can be obtained by using this subset. We introduce active learning based on locally linear reconstruction "(LLRActive) in detail firstly. Then, in the framework of LLRActive,we substitute the Laplacian matrix for the locally linear reconstruction matrix to characterize the local property and obtain Laplacian transductive optimal design (LTOD).The samples which can be used to best reconstruct the whole samples are selected as the optimal training subset. Several active learning algorithms are utilized to recognize the ground target MMW HRRP and the results are compared.Imbalanced learning is a novel issue we faced when the pattern recognition methods are applied in some practical applications, which concerns more about the recognition of the minority class.When the training samples are imbalanced, the hyperplane will move to the minority. This results in a low recognition of minority. To solve the imbalanced learning problem in HRRP recognition, we proposed a novel semi-supervised feature selection algorithm, called cost-sensitive geodesic constraint scores (CSGCS). We first extend the constraint scores to the case of imbalanced training samples by using the cost-sensitive technique. Then, we extend it to the case of semi-supervised learning by using the constraint reconstruction technique. CSGCS is designed for the use of semi-supervised imbalanced learning and will improve the recognition rate of the minority class.