Incremental Robust Non-negative Matrix Factorization for SAR Image Recognition

Abstract

In the face of incremental learning of synthetic aperture radar (SAR) image recognition, recognition model is easy to affect by the increase of training samples. In this paper, incremental robust non-negative matrix factorization(IRNMF) scheme is proposed to relieve the problem. The noise and outliers transform the projection direction from the original space to feature space as the increase of training samples. Thus conventional non-negative matrix factorization(NMF) is easily prone to noise and outliers since its loss function includes the squared errors. The loss function of IRNMF is defined by using $\mathrm{L}_{2,1}$ -norm so that the proposed algorithm can avoid the influence of the squared errors. In the process of incremental learning, some experiments on Moving and Stationary Target Acquisition and Recognition(MSTAR) dataset prove that our scheme consistently outperforms NMF and RNMF.