ProtoRefine: Enhancing Prototypes with Similar Structure in Few-Shot Learning

Few-shot learning presents a substantial challenge in developing robust models due to the inherent scarcity of samples within each category. To overcome this challenge, metric-based methods have been introduced, classifying images based on the relationships among samples within a given embedding space. While these methods are effective, the limited samples often result in an incomplete representation of the category’s feature space, leading to sub-optimal prototypes for classification. Recognizing this shortcoming, we identified that categories in new tasks often exhibit structural similarities with those in the relative base domain. Driven by this observation, we introduce ProtoRefine. Our approach employs the structural information of categories within the base domain that bear relevance to the new tasks, generating additional sample embeddings. This strategy refines the prototype representation, thus providing a more accurate prototype for category classification. We performed extensive experiments on popular few-shot learning benchmarks, with the results highlighting the effectiveness of ProtoRefine, especially within the 5-way 1-shot settings. Matching the competitive results of state-of-the-art methods, our work underlines the significant advantage of enhancing prototypes with structurally similar information from the base domain in the context of few-shot learning.