Few-shot image classification aims to learn novel classes with limited labeled samples for each class. Recent research mainly focuses on reconstructing a query image from a support set. However, most methods overlook the nearest semantic base parts of support samples, leading to higher intra-class semantic variation. To address this issue, we propose a novel prototype resynthesis network (PRSN) for few-shot image classification that includes global-level and local-level branches. Firstly, the prototype is compounded from semantically similar base parts to enhance the representation. Then, the query set is used to reconstruct the prototypes, further reducing intra-class variations. Additionally, we design a cross-image semantic alignment to enforce global-level and local-level semantic consistency between different query images of the same class. Our empirical results demonstrate that PRSN achieves remarkable performance across a range of widely recognized benchmarks. For instance, our method outperforms the second-best by 0.69% under 5-way 1-shot settings with ResNet-12 backbone on the miniImageNet dataset.
Read full abstract