Learning Class-Aware Prototypical Representations for Few-Shot Learning

Few-shot learning in visual classification aims to emulate human capacity to recognize unseen classes using only a few samples. Prototype-based methods have shown promising classification outcomes, but most of them focus only on exploiting the intra-class distribution information while neglecting the importance of inter-class information. In this paper, we propose a class-aware approach, Boosted Prototypical Networks (BPN), which comprehensively evaluates the intra- and cross-class performance of samples by using their silhouette coefficients, and weights them accordingly to construct more representative prototypes. Furthermore, we design a discriminant loss function to guide the model to learn hard features adaptively, resulting in an embedding space that is more discriminative for hard classes. Ultimately, our comprehensive experimental results on benchmark datasets demonstrate that the proposed BPN is more competitive than existing algorithms.