A multi-view stable feature extraction network for SAR target recognition

Since the viewing angle sensitivity of Synthetic Aperture Radar (SAR) image, targets’ shape changes dramatically in different viewing angle, the accuracy of SAR target recognition base on deep learning in single view is often affected by this difference between various viewing angle of SAR. So the multi-view approaches have been proposed to solve this limitation by learning the targets’ relationships between images from different viewings. In this paper, a multi-view stable feature extraction network (MV-SFN) is proposed which constructs stable feature that doesn’t change with viewing angle by using multi-view images. It uses Convolutional Neural Networks (CNN) to gain image feature, then take Gate Recurrent Unit (GRU) to analyze varying regularity of series viewing features. At last a novel features fusion module was designed for fusing different viewing features according to their similarity, and finally going to classify. MSTAR database is used to validate MV-SFN method.