Robust image recognition by L1-norm twin-projection support vector machine

Abstract

In this paper, we propose a L1-norm twin-projection support vector machine (TPSVM-L1) for robust representation and recognition of images. The robustness of our TPSVM-L1 method is mainly driven by the L1-norm based distance metric that is proven to be robust to noise and outliers in data. For discriminant twin-projection learning, our TPSVM-L1 aims to compact L1-norm regularized intra-class scatter and separate L1-norm regularized inter-class scatter in addition to enabling the learning system to be robust against noise or outliers. As a result, an optimal pair of robust and descriptive linear projective subspaces or the most discriminating linear vector projections for constructing two hyper-planes can be trained. The twin-vector projection subspace is effectively achieved by an iterative approach. Note that the linear twin-projections can be used to extract features from images, and the hyper-planes can decide the categories of test data by embedding them onto the hyper-planes. Simulations on UCI and real image datasets verified the validity of our TPSVM-L1, compared with the other related twin SVM classification algorithms.