Robust latent discriminative adaptive graph preserving learning for image feature extraction

Many feature extraction methods based on subspace learning have been proposed and applied with good performance. Most existing methods fail to achieve a balance between characterizing the data and the sparsity of the feature weights. At the same time, the assumptions on one specific type of noise may degrade the performance of feature extraction when the data contains complex noise. For these, this paper proposes a robust latent discriminative adaptive graph preserving learning model for feature extraction (RLDAGP). The F-norm is used to preserve the global structure of the data instead of the widely used nuclear norm. Moreover, we prove that the proposed method has a low-dimensional grouping effect, which means that highly correlated samples will be grouped together. Further, a correntropy-inducing metric (CIM) is introduced to the noise matrix to suppress complex noise. Besides, an adaptive graph regularizer is integrated into the model to enhance its robustness while preserving the local structure and enhancing the intra-class compactness. In particular, a transformed l2,1-norm regularization, which is smoothly interpolated in l2,1-norm and F-norm, is introduced to the projection matrix to adaptively extract discriminative features from the data. In order to solve the proposed nonconvex model, we design an algorithm based on nonconvex-ADMM framework and prove the convergence of the proposed algorithm theoretically. Experiments demonstrate the superiority of the proposed method over the existing state-of-the-art methods.(c) 2023 Elsevier B.V. All rights reserved.