Fusion of Multifeature Low-Rank Representation for Synthetic Aperture Radar Target Configuration Recognition.
Xinzheng Zhang,Yijian Wang,Dong Li,Zhiying Tan,Shujun Liu
DOI: https://doi.org/10.1109/lgrs.2018.2842068
IF: 5.343
2018-01-01
IEEE Geoscience and Remote Sensing Letters
Abstract:In this letter, we propose a synthetic aperture radar (SAR) target configuration recognition algorithm based on the fusion of multifeature low-rank representations (LRRs). First, Gabor, principal component analysis, and wavelet features are extracted for the SAR training set and test set, respectively. Second, with the LRR model, each feature of the test samples is represented by those of the training set, leading to the corresponding coefficient matrix. Then, the preliminary prediction labels of all features of the test sample are obtained according to the LRR coefficients. Third, in order to further improve the confidence of recognition and reduce the instability of the algorithm, a two-stage decision fusion strategy is adopted to obtain the final prediction labels. The first stage utilizes a vote fusion for the recognition results of multiaspect neighborhood test samples for each feature pattern, which exploits the strong correlation of these neighborhood samples. Furthermore, the second stage fuses the three results obtained in the first stage through Bayesian inference. Bayesian inference is widely used in decision fusion, which can improve the confidence of results by about 3%. Experiments on the moving and stationary target acquisition and recognition data set demonstrate the effectiveness and superiority of the proposed algorithm.