SAR target recognition via linear t-stochastic neighbor embedding and sparse representation

In this paper, a novel technique based on linear t-stochastic neighbor embedding (t-SNE) and sparse representation is presented for target recognition in SAR images. Though t-SNE has a powerful performance on dimension reduction and visualization task, the lack of a parametric mapping makes t-SNE less suitable for using in target recognition. Here we extent t-SNE toward an efficient linear projection to solve the out-of-sample problem. What's more, the projection preserves the local structure characteristic of SAR images in a manifold. Specifically, each class belongs to each sub-manifold. Then, the test sample is represented by sparse linear combination of the training samples in the feature space. The best class estimation of the test sample corresponds to the least distances between the test sample and training samples with a certain sub-manifold. Experimental results on moving and stationary target automatic recognition (MSTAR) database evaluate the effectiveness of the proposed method.