QGF-Net: Fine-Grained Image Classification via Self-Supervised Vision Transformer and Quantum Measurement Attention

<title>Abstract</title> <p>Fine-grained image classification (FGIC) aims to distinguish subordinate categories within the same basic-level class, and remains highly challenging due to small inter-class differences, large intra-class variations, and the fact that discriminative cues are often concentrated in local regions [2, 3]. Recently, self-supervised Vision Transformer (ViT) models have achieved remarkable success in visual representation learning and provide powerful global semantic features for downstream classification tasks [10, 11]. However, existing methods still suffer from two limitations. First, self-supervised ViT tends to focus more on global semantics while being insufficiently sensitive to subtle local discriminative regions that are crucial for fine-grained recognition [7, 8]. Second, conventional attention mechanisms mainly rely on classical similarity computation, which may be insufficient for modeling complex high-order nonlinear dependencies among local regions; meanwhile, recent quantum-enhanced vision studies suggest that parameterized quantum circuits can provide an alternative mechanism for modeling structured dependencies [19, 20]. To address these issues, this paper proposes QGF-Net, a hybrid quantumclassical fine-grained image classification network that integrates self-supervised visual representations, illumination-consistent local feature stabilization, local discriminative region proposal, and quantum measurement attention into a unified architecture. Specifically, a DINOv2-pretrained ViT is first employed to extract global semantic features and patch token representations from input images [11]. Then, an Illumination-Consistent Local Coarse Fusion (ICLCF) module is introduced to improve the robustness and consistency of local token features under illumination changes. Based on the stabilized local representations, a Local Discriminative Region Proposal (LDRP) module is designed to identify the most informative local regions, thereby narrowing the search space for fine-grained recognition. A Quantum Measurement Attention (QMA) module is further constructed to model the relationships among key local regions through a lightweight parameterized quantum circuit, where qubit measurement results are used to generate adaptive attention weights for enhancing local discriminative representations [18–20]. Finally, the global semantic features and quantum-enhanced local features are fused and fed into a Hybrid Quantum-Classical Classifier (HQCC) for final prediction. Experimental results on three representative fine-grained benchmarks, namely CUB-200-2011, Stanford Cars, and FGVC-Aircraft, demonstrate that QGF-Net consistently outperforms strong CNN, Transformer, and locally enhanced baselines in terms of Top-1 Accuracy, Macro-F1, and robustness [13, 15, 16]. The full model achieves 92.0%, 94.2%, and 89.5% Top-1 accuracy on CUB, Cars, and Aircraft, respectively, with an average Top-1 accuracy of 91.9%. Robustness experiments under brightness, shadow, and occlusion perturbations further show that QGF-Net achieves the smallest average accuracy drop among compared methods. Ablation studies verify the effectiveness of ICLCF, LDRP, QMA, and HQCC, and additional sensitivity analysis demonstrates that the proposed quantum module provides a favorable balance between representation gain and training stability. Overall, QGF-Net provides a promising paradigm for fine-grained image classification by combining self-supervised ViT with hybrid quantum-classical modeling.</p>