Proceedings of the Annual Meeting of the Cognitive Science Society

Few-shot fine-grained image classification (FSFGIC) is the task of developing computational models to recognize highly similar sub-categories from minimal training data. This task poses a core dilemma: over-emphasizing local details makes models vulnerable to noise and overfitting, while relying on global features often lacks the sensitivity required for fine-grained discrimination. Inspired by the human visual cognition strategy of dynamically balancing global and local information, we propose an Adaptive Global–Local Balance Network (AGLB-Net), a framework that computationally implements adaptive global– local integration for FSFGIC. AGLB-Net introduces two key modules: a Hierarchical Discriminative Feature Refinement (HDFR) module that progressively integrates representations from global semantics to fine-grained details, and an Adaptive Regional Re-Attention Module (ARRM) that automatically localizes and emphasizes discriminative regions without additional supervision. Extensive experiments on widely-used benchmarks demonstrate that AGLB-Net consistently achieves state-of-the-art performance across various few-shot settings, validating the effectiveness of human visual cognition-inspired adaptive global–local balancing in FSFGIC.