Robust deep dictionary learning via self-expression neighbor atom enhancement

Abstract

Dictionary learning has been widely recognized as an effective method for processing high-dimensional and nonlinear data. Although graph structures have been widely adopted to enhance optimization in irregular data handling, existing approaches inadequately utilize local graph information, leading to suboptimal dictionary update. Consequently, the learned dictionary struggles to capture local details effectively, which can negatively impact the performance of classification tasks. To tackle this problem, this study proposes a method based on graph structure, dubbed Deep Neighbor Atom-enhanced Dictionary Learning (DNADL), designed to enhance the sensitivity of dictionaries to local structure. Our DNADL segments data into internally consistent neighborhood manifold module via clustering method and enhances dictionary atoms based on their self-expression relationships, where the self-expression relationships within generated manifolds reflect the intrinsic correlation of local data structures. Enhanced dictionary atoms exploit the similarities among neighboring data points, ensuring that the constructed dictionary effectively captures local structure features. Furthermore, this study introduces an adaptive graph learning module that dynamically updates graph representations, thereby facilitating the simultaneous optimization of dictionary learning and graph topology within an integrated framework. Comprehensive experimental results demonstrate that DNADL achieves state-of-the-art classification performance across multiple benchmark datasets, with accuracies of 99.23 % on Fashion-MNIST, 94.12 % on EMNIST, and 95.3 % on Oxford Flowers 102. By effectively capturing local structural features, DNADL surpasses existing methods by a significant margin of 0.98 % to 3.70 % in accuracy.