Adaptive deep clustering integrating DINOv2 embeddings, graph attention, and bio-inspired optimization

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised image clustering lacks robust methods for integrating diverse learning paradigms.
• Existing transformer- or GNN-based systems often treat components in isolation, limiting adaptive refinement.
• Need for interpretable internal metrics in fully unsupervised scenarios.

Purpose of the Study:

• To present a unified and adaptively integrated framework for unsupervised image clustering.
• To establish a synergistic interaction between self-supervised representation learning, graph-based embedding refinement, and bio-inspired optimization.
• To introduce interpretable composite internal indices for reliable unsupervised evaluation.

Main Methods:

• Utilizes pretrained DINOv2 Vision Transformers for high-level feature extraction.
• Employs a multi-head Graph Attention Network (GAT) for refining relational structures.
• Integrates a bat-inspired metaheuristic for joint cluster number estimation and adaptive hyperparameter tuning.

Main Results:

• Achieved state-of-the-art performance on CIFAR-10 (NMI=0.938, ARI=0.932) and other benchmarks.
• Demonstrated effectiveness and generalization capability across diverse datasets.
• Introduced two composite internal indices ([Formula: see text] and [Formula: see text]) correlating strongly with external metrics.

Conclusions:

• The proposed DINOv2-GAT-BAT pipeline offers a novel adaptive deep clustering approach.
• The integrated framework dynamically refines semantic embeddings and structural information.
• The study advances unsupervised evaluation with interpretable composite metrics and achieves superior clustering results.