Dual Self-expression Subspace Clustering with Multi-scale Features and Structural Consistency Alignment

In recent years, deep clustering has shown strong potential for unsupervised representation learning on high-dimensional and complex data. Nevertheless, many existing approaches struggle to simultaneously capture fine-grained visual details across different scales and the intrinsic structural relationships embedded in data manifolds, which limits their clustering performance. To address these challenges, this paper proposes a Dual Self-Expression Subspace Clustering method with Multi-Scale Features and Structural Consistency Alignment (DMSSC). In DMSSC, EfficientNet is employed to extract hierarchical multi-scale features that balance local texture information and global semantic cues, while Chebyshev graph convolution effectively models structural dependencies and mitigates the over-smoothing issue commonly observed in standard GCNs. Furthermore, dual self-expression modules are designed to learn complementary content and structural affinities, which are aligned through a consistency constraint to produce more robust and discriminative subspace representations. Extensive experiments on four benchmark datasets validate the effectiveness and stability of DMSSC, achieving 87.58% ACC and 82.12% NMI on STL10. The implementation is publicly available at https://github.com/DMSSC-123/DeepSubspaceClustering .