Deep supervised t-SNE for SAR target recognition

In this paper, a novel feature extraction method based on t-distributed stochastic neighbor embedding (t-SNE) is presented for target recognition in synthetic aperture radar (SAR) images. It aims to search and preserve the local structure characteristic of SAR images. Recently, t-SNE has been widely studied in finding the underlying structure of data as an efficient technique. However, less attention has been paid to apply it for target recognition, since t-SNE cannot provide a parametric mapping to deal with the out-of-sample problem. To solve this problem and capture the complex characteristics of SAR images as well as possible, t-SNE is extended by a deep feed-forward network with a complex nonlinear mapping function. The network is pre-trained using a stack of Restricted Boltzmann Machines (RBMs). To boost the performance, the aspect angles of SAR images are used for supervising the t-SNE, which makes the same class more concentrated. Experimental results on moving and stationary target automatic recognition (MSTAR) dataset reveal the effective performance of the proposed method.