Robust sparse discriminative least squares regression for image classification

Abstract

Discriminative Least Squares Regression (DLSR) is a method used for multi-class classification tasks that expands the distance between different classes through an ε-dragging technique. However, it also amplifies the differences in intra-class regression targets. Moreover, the samples contain a significant amount of noise, which negatively affect the classification performance. To mitigate these problems, we propose Robust Sparse Discriminative Least Squares Regression (RSDLSR) approach to enhance the model's discriminative power. Firstly, we maintain the original data structure by matrix decomposition in the label space. Secondly, the noise is fitted using sparse constrained noise matrix to enhance the model's denoising ability. Furthermore, we select important features from label space using a linear discriminant analysis criterion to minimize the influence of redundant features. Finally, $l_{2,1}$ norm constraint is imposed on the relaxation matrix to improve the sparsity and robustness of the model. Comparative evaluations demonstrate that our proposed method exhibits significant advantages over various existing methods across different classification tasks.