Research On Algorithms Of SAR Target Discrimination

Synthetic aperture radar (SAR) is the most popular research topic in radar area at present. And the SAR automatic target recognition (ATR) system which is considered to play a big role in the future battle field is attracting more and more attention from the country related. This article is investigating the technology of target discrimination in preference which is one of the three main phrases included in the framework of the typical SAR ATR model proposed by Lincoln Laboratory. Target discrimination technology can be further divided into three steps, feature extraction, feature selection, and discriminator designing included.As in the aspect of feature extraction, the article introduced fourteen features for discrimination extracted by Lincoln Laboratory and analyzed them in detail theoretically and experimentally. And then in the phase of discriminator designing, one-class classifier was introduced, in which Support Vector Data Description (SVDD) was analyzed particularly. SVDD algorithm bases on the common idea of Support Vector Learning, however, which is different from traditional support vector classifier is that it constructs a tight hyper sphere boundary around target data and makes the algorithm showed good property on the one-class classify problem. Compared with classic Gaussian classifier, neither prior knowledge about distribution nor the prior probability of input feature vector need to estimated, and besides, few sample can still get superior performance.Last part of this article introduced the discriminating algorithm which is based on covariance matrix feature and manifold SVDD. Covariance matrix feature used for discrimination was firstly extracted. However, since the space of the new feature is not an Euclidean space, the classical one-class classifier originally defined in the Euclidean space can not be used directly. Therefore, we further improved classical SVDD algorithm which associated with characteristic of covariance matrix. The model of region covariance matrix with discriminator will provide a new possibility for SAR target discrimination.