Hypergraph-Augmented Reconstruction and Adaptive Metric Aggregation Network for Few-Shot Fine-Grained Image Classification

Abstract

Few-shot fine-grained image classification aims to recognize unseen fine-grained categories from very few labeled examples. The task is difficult because visually similar classes may differ only in subtle details, while images from the same class can vary considerably. Useful evidence is therefore often localized. Reconstruction-based metric methods reduce local misalignment by reconstructing query features from class-wise support features, but a small support set may still provide an incomplete reconstruction basis. In addition, spatial reconstruction errors are usually combined with fixed or nearly uniform weights, and intermediate features receive no direct reconstruction-oriented supervision. We develop HARA, a Hypergraph-Augmented Reconstruction and Adaptive Metric Aggregation Network, to address these limitations. HARA forms learnable soft hyperedges over local support features and uses the resulting aggregated features to enrich the class-wise reconstruction basis. It then combines class-conditional reconstruction with query self-reconstruction to obtain position-wise joint errors, which are aggregated using query-dependent weights. A training-only auxiliary branch applies reconstruction-based supervision to intermediate backbone features. For 5-way evaluation, HARA obtains 84.09%/93.48% on CUB-200-2011, 76.91%/88.69% on Stanford Dogs and 88.53%/95.87% on Stanford Cars under the 1-shot/5-shot settings, respectively. Ablation and diagnostic analyses further characterize the contributions and behavior of the proposed framework. Overall, HARA provides an effective extension to reconstruction-based few-shot fine-grained classification.