Double relaxed broad learning system for image classification

The Broad Learning System (BLS), as a prominent network paradigm, has garnered significant attention across various domains in machine learning due to its remarkable efficiency and impressive performance. Nevertheless, two deficiencies exist when its variants are applied to supervised image classification tasks. First, they generally rely on rigid binary labels, which restricts the approximation procedure and fails to find the best classification margins. Second, using a single transformation matrix in graph-regularized BLS struggles to mitigate the overfitting problem. In this article, we focus on offering more freedom to the graph regularized BLS to explore appropriate margins between samples. Specifically, we learn the adaptive labels from data, and a marginalized constraint is implemented to enhance the distinguishability of the learned labels simultaneously. Subsequently, two distinct transformation matrices are employed for the graph embedding process, with the addition of a novel matrix designed to capture the similar structure between the transformations. The intra-class compactness of samples can be assured through experimental verification. Hence, a novel method named double relaxed BLS (DRBLS) is proposed. Further, with the help of the alternating direction method of multipliers, an efficient iterative approach is developed to find the closed-form solution. Experiments on diverse image classification tasks are conducted to certify the effectiveness of our method compared to state-of-the-art algorithms.