Feature Extraction And Recognition Methods For Radar High Range Resolution Profiles

Radar high-resolution range profile (HRRP) denotes the coherent summation of projection vectors of the complex echoes from target scatterers along the radar line-of-sight (LOS). HRRP is easy to obtain, store and process, and contains informative target structure signatures, e.g. target size, scatterer distribution, etc. Therefore, radar HRRP target recognition has received extensive attention in the radar automatic target recognition (RATR) community. This dissertation focuses on the theories and techniques of HRRP-based RATR in feature extraction and recognition methods. Our researches are supported by Advanced Defense Research Programs of China and National Science Foundation of China.The main works of this dissertation is summarized as follows.1. We introduce the basic concepts of HRRP and the fundamental theories of RATR at first, and then review several related works in recent years. Finally, a brief description of the main works in our dissertation is given.2. To deal with the limited data description capability of traditional subspace methods, a dictionary learning based radar HRRP target recognition method is established. The proposed method can adaptively select the sparsity level based on the estimated test noise level. Compared with the traditional HRRP recognition methods, the proposed algorithm has higher recognition accuracy and is more robust to the test noise environment. Furthermore, this method achieves satisfactory generalization performance even with a few HRRP training samples from partial target-aspect angles (i.e., the training dataset is incomplete), thereby it can be used for HRRP dataset extension. The experiments based on measured HRRP data validate the proposed method.3. The sparse representation of signal via dictionary learning is widely used in signal processing field. Since there is redundancy in the new space defined by overcomplete dictionary atoms, the problem of finding sparse representation may bring the uncertainty and ambiguity in the presence of unknown amplitude perturbations, which is unfavorable to radar HRRP target recognition task. To deal with this issue, our works include:(1) A novel algorithm called Dropout-based Stable Dictionary Learning (Drop-SDL) is proposed, which constructs a robust loss function via marginalizing dropout to learn a stable adaptive dictionary. This algorithm considers the structure similarity among the adjacent HRRPs without scatterers’motion through range cells (MTRC), and enforces the constraints that the sparse representations of adjacent HRRPs should have the same supports. Moreover, Drop-SDL utilizes the structured sparse regularization learned in the training phase to automatically select the optimal sub-dictionary basis vectors, which is used for the classification of the test sample. (2) The stability conditions in the presence of amplitude perturbations with a fixed dictionary are established first. Then a similar but more flexible stability conditions for adaptive dictionary is utilized to propose a stable dictionary learning (SDL) method. The proposed method relies on the constraints that the sparse representations of adjacent HRRPs should have the same support and lower variance. The structured sparsity regularization is then employed to automatically select the optimal dictionary basis vectors for stable sparse coding. Experimental results on measured radar HRRP dataset validate the effectiveness of the proposed method.4. Traditional dictionary learning algorithms only depict the linear relationship between the observation and sparse representation. In practice, however, there is complex nonlinear relationship between them. Hence, linear dictionary learning algorithms cannot be guaranteed a good performance. To deal with this issue, the main works here are aimed at nonlinear dictionary learning algorithm. (1) A dictionary learning algorithm with simultaneous nonnegative sparsity constraints is presented. An adaptive clustering algorithm can be derived from the proposed dictionary learning algorithm by constraining the sparsity degree of signal to be 1, which does not need to artificially specify the centers and is independent of initialization. Next, we extend it to a nonlinear clustering algorithm via using kernel tricks. (2) A kernel dictionary learning algorithm with simultaneous nonnegative sparsity constraints is established. This algorithm is used to extract the basis vectors in the feature space and remove redundant samples and outliers in the observation space. After obtaining the basis vectors, K-PCA is utilized to realize feature extraction. Experimental results on UCI dataset and measured radar HRRP dataset validate the effectiveness of the proposed method.5. Feature extraction is the key technique for RATR based on HRRP. Traditional feature extraction algorithms usually use shallow models. When applying such models, the inherent structure of the target is always ignored, this is disadvantageous for learning effective features. To address this issue, deep architecture for radar HRRP target recognition, which adopts multi-layered nonlinear networks, is studied. The main works include:(1) Because of the stable physical properties of the average profile in each HRRP frame without MTRC, Stacked Robust Autoencoders (SRAEs) are further developed, which are stacked by a series of RAEs. SRAEs can not only reconstruct the original HRRP samples, but also constrain the HRRPs in one frame close to their average profile. Then the top-level output of the networks is used as the input to the classifier. Experimental results on measured radar HRRP dataset validate the effectiveness of the proposed method. (2) To learn a more stable structure and a higher-order correlation of targets from unlabeled data, Stacked Corrective Autoencoders (SCAEs) is further proposed with the utilization of the HRRP’s characteristics. The proposed model, which is an extension of Stacked Denoising Autoencoders (SDAEs), is stacked by a series of learned Corrective Autoencoders (CAE) and employs the average profile of each HRRP frame as the correction term. The covariance matrix of each HRRP frame is considered for establishing an effective loss function under the Mahalanobis distance criterion. The proposed model can be used as the tool to reduce the dimensionality of HRRP data which facilitates the classification, visualization, and storage tasks. With the proper optimization procedure and selection of hyper-parameters, we demonstrate our model can be beneficial, even with a moderately incomplete training set.