Feature separability analysis for SAR ATR using data description method

Abstract

Feature extraction and selection play an important role in radar target recognition. This paper focuses on evaluating feature separability for SAR ATR and selecting the best subset of features. In details, fifteen features extracted from T72, BTR70 and BMP2 in MSTAR standard public dataset are examined, which are divided into seven categories: standard deviation, fractal dimension, weighted-rank fill ratio, size-related features, contrast-based features, count feature, projection feature, and moment features. Since the number of samples is small, a new separability criterion based on the overlap degree of each two class regions is proposed to assess the separability of these features. Here the class region is described by support vector data description (SVDD) method for good generalization. Based on the proposed criterion, a forward feature selection method is adopted to choose the best subset of features. Because of the strong variability of the feature against aspect, the features are analyzed under different aspect sectors within 360°angle range stepped by 15°, 30 °, and 60°, respectively. Experiments using MSTAR dataset validate the criterion, and the best subset of features is determined.