Abstract
In recent years, the integration of deep learning techniques has significantly advanced subspace clustering. However, existing deep subspace clustering methods predominantly rely on self-expressive layer structures, generally lacking solid theoretical foundations and interpretability. Furthermore, these methods typically require loading the entire dataset into memory simultaneously for learning, a strategy that not only consumes substantial computational resources but also severely restricts the application of deeper network architectures. To address these issues, we propose a novel deep subspace clustering framework called UIDSSC (Unfolded ISTA for Deep Sparse Subspace Clustering). By designing a new subspace clustering objective function based on mini-batch deep features and unfolding its optimization steps into a network structure, we not only apply the unfolded form of the Iterative Shrinkage-Thresholding Algorithm (ISTA), which has a rigorous mathematical foundation, to deep sparse subspace clustering for the first time, but also successfully combine subspace clustering with self-supervised learning, thereby leveraging the powerful representation capabilities of deep networks. The resulting framework achieves both mathematical interpretability through the unfolded ISTA architecture and powerful representation capabilities through deep self-supervised features, while supporting efficient mini-batch training on large-scale datasets. Comprehensive experiments on four benchmark datasets (CIFAR-10, ImageNet 10, ImageNet-Dogs, and STL-10) demonstrate that UIDSSC significantly outperforms existing methods, particularly achieving 83.3% accuracy and 79.7% normalized mutual information on the CIFAR-10 dataset, and reaching 96.6% accuracy and 92.1% normalized mutual information on the ImageNet-10 dataset, both attaining state-of-the-art performance. Our code is publicly available at https://github.com/rook1ess/UIDSSC.
•First unfolding ISTA for deep sparse subspace clustering.•Mini-batch objective enables deep networks for subspace clustering.•Self-supervised learning enhances feature dictionary construction.•Adaptive sparsity through learnable soft-thresholding operations.