Learning salient self-representation for image recognition via orthogonal transformation

Abstract

Self-representation is a learning paradigm that exploits the intrinsic information from the given observation by representing the observation itself with a linear combination. Recent works have not considered learning the self-representation from disparate spaces, which cannot fully exploit the discriminative property for classification. To resolve this issue, this paper proposes an approximated self-representation, termed as salient self-representation (S2R), which learns an approximated self-representation between the given data itself and its projection in the L4 space. We will show that we can project the data to the L4 space via a linear orthogonal transformation. Here, the salient information will be preserved when we pursue the sparsity from both the L0 and L4 spaces. A classifier is proposed to apply the learned salient self-representation to pattern classification. Furthermore, we proved that the S2R can well incorporate the salient information with supervised information for pattern classification. Several numerical experiments including comparisons and visualizations with the state-of-the-art methods are provided to verify the effectiveness of S2R for pattern classification.