DFPN: a dynamic fusion prototypical network for few-shot learning

Abstract

Prototypical networks have been widely adopted for few-shot image classification. However, due to data scarcity, these methods often suffer from bias and struggle to capture discriminative features effectively. To address the problem, we propose a novel dynamic fusion prototypical network (DFPN) that learns more representative prototypes from limited training samples. In particular, we present a dynamic prototypical network that leverages dynamic routing within a meta-learning framework, effectively mapping sample representations to prototype representations. To further enhance prototype estimate, we design a distribution-based fusion strategy that mitigates biased distributions by integrating mean-based prototypes with adaptively generated dynamic prototypes. Moreover, we employ the Yeo-Johnson transformation to make the feature distribution more Gaussian-like, thereby improving representation quality. Extensive experiments on five benchmark datasets demonstrate the effectiveness of our method. Notably, our DFPN achieves state-of-the-art performance on the miniImageNet dataset, reaching 74.34% accuracy in the 5-way 1-shot setting and 86.56% in the 5-way 5-shot setting. These results demonstrate DFPN can learn more expressive prototypes, significantly advancing few-shot image classification performance.