Robust classifier using distance-based representation with square weights

Abstract

This paper presents a new multi-class classification method that is different from sparse representation classifier (SRC) method. SRC is a classical method which has been widely used for face recognition and digit identification. However, SRC method only looks for the sparsest solution using $$l_1$$ l 1 norm minimization with high computation complexity. The sparsest representation cannot show the space distribution feature of samples. Moreover, the sparsest representation does not mean obtaining the highest recognition rate for data classification. This paper proposes a distance-based representation method for classification. The distance between samples is used to measure the similarity. It is crucial that square weights $$x_i^2$$ x i 2 are used as the weight of distance instead of $$x_i$$ x i . Furthermore, a closed form solution is obtained so that the computation complexity is lower than that of SRC. The extensive experiments show that the proposed method achieves very competitive classification results.