Data representation via refined discriminant analysis and common class structure

The main goal of discriminant embedding is to extract features that form a compact and informative representation of the original feature set. In this paper, we propose an improved hybrid method aimed at extracting linear features for supervised multiclass classification. We implement a unifying criterion that is able to preserve the benefits of robust sparse linear discrimination along with inter-class sparsity. The expected transformation involves two forms of discrimination, namely: common class or group sparsity in addition to robust discriminant analysis with feature ranking. For the purpose of solving the proposed criterion, an iterative alternating minimization framework is used to evaluate the linear transformation and orthogonal matrix. The presented scheme is generic enough that it can be used to consolidate and tune several linear embedding methods. In the light of experiments conducted on various image datasets with different types, the suggested scheme was able to outperform other methods in most cases. (C) 2021 Published by Elsevier B.V.