Abstract
This paper presents a unified and adaptively integrated framework for unsupervised image clustering that establishes a novel synergistic interaction between self-supervised representation learning, graph-based embedding refinement, and bio-inspired optimization. Rather than employing DINOv2, GAT, and the Bat Algorithm as isolated components, the proposed DINOv2-GAT-BAT pipeline introduces a closed-loop adaptive mechanism in which semantic embeddings, attention-guided structural information, and cluster-shaping optimization dynamically influence one another. The framework first extracts high-level visual features using pretrained DINOv2 Vision Transformers, then refines relational structures through a multi-head Graph Attention Network (GAT), and finally employs a bat-inspired metaheuristic that jointly estimates the optimal number of clusters and adaptively tunes structural and hyperparameter configurations. This tightly coupled interaction results in a new form of adaptive deep clustering not present in existing transformer- or GNN-based systems. To improve interpretability, two composite internal indices-[Formula: see text] and [Formula: see text]-are introduced, jointly capturing separability, entropy, compactness, stability, and outlier sensitivity. These indices exhibit strong correlations with external evaluation metrics, enabling reliable and meaningful assessment in fully unsupervised scenarios. Extensive experiments on CIFAR-10, Oxford-IIIT Pet, and STL-10 demonstrate the effectiveness and generalization capability of the proposed framework. On CIFAR-10, it achieves NMI = 0.938, ARI = 0.932, and a Composite Score = 0.894, surpassing several state-of-the-art baselines. Overall, this work (1) introduces a novel adaptive integration mechanism linking transformers, graph attention, and metaheuristic optimization, (2) proposes interpretable composite metrics for unsupervised evaluation, and (3) achieves state-of-the-art clustering performance across diverse benchmarks.