Retargeted broad learning systems for image classification

Abstract

The Broad Learning System (BLS) is recognized for its adept balance between efficiency and accuracy, displaying notable performance in image classification tasks owing to its streamlined network architecture and effective learning methodology. However, it faces significant challenges due to two prominent deficiencies that notably impede its learning efficacy. Firstly, the rigid binary labeling strategy inherent in BLS-based models imposes constraints on the model's adaptability. Additionally, the resultant broad features often exhibit redundancy, posing a risk of incorporating extraneous features. To address these issues, this article proposes three refined BLS-based models. Initially, a retargeting methodology is integrated into the standard BLS framework to alleviate constraints on regression targets, introducing the ${\ell }_2$-based retargeted BLS (L2ReBLS) model. Subsequently, to mitigate the adverse effects of redundant features, the ${\ell }_{2,1}$ regularizer is adopted as a replacement for the Frobenius norm in feature selection, resulting in the L21ReBLS model. Furthermore, the projection matrix of BLS is concurrently constrained with ${\ell }_2$ and ${\ell }_{2,1}$ regularization method simultaneously. Efficient iterative optimization methodologies via the alternating direction method of multipliers are devised for the purpose of solving the proposed approaches. Ultimately, comprehensive experiments conducted on diverse image databases are to highlight the superior performance of our proposed approaches in comparison to other state-of-the-art classification algorithms.