Multi-granularity awareness via cross fusion for few-shot learning

Many existing few-shot learning methods based on a single-granularity structure have achieved promising results, but they insufficiently exploit the latent information within granularity structures. In addition, these methods still struggle with unclear class distinctions and scattered intra-class feature spaces, posing significant challenges for few-shot learning. To address these issues, we propose a multi-granularity awareness-guided cross-fusion (MACF) method for few-shot image classification. MACF enhances inter-class separation while promoting intra-class cohesion. Specifically, we introduce multi-granularity awareness to construct both coarse- and fine-grained representations, strengthening class affiliations and improving feature discriminability across classes. Moreover, we utilize local descriptors as features to develop a cross-fusion strategy that generates distinguishable pseudo-features, enriching intra-class feature diversity and mitigating the impaired generalization ability for unseen classes. Furthermore, we establish a reliability assessment metric for synthetic features to ensure the generation of representative features while reducing the impact of outlier perturbations. Extensive experiments on six few-shot benchmark datasets demonstrate the effectiveness of the proposed MACF method. The basic data and code can be accessed via the link: https://github.com/woodszp/MACF.