Contrastive prototype learning with semantic patchmix for few-shot image classification

Few-shot image classification aims to learn unseen classes with only a few training samples for each class. However, most existing models still suffer from weak feature representation due to data scarcity. To this end, a novel contrastive learning framework is proposed for few-shot image classification that utilizes patch-wise and class-wise features. Concretely, a semantic patchmix scheme is designed to effectively capture patch-wise features with more discriminative representation. Specifically, a new information noise contrastive estimation loss with modulating factor is proposed to adjust the weights of samples, which is adaptive and trades off different samples. For class-wise features, contrastive prototype learning on two correlated views is leveraged to enhance the generalization of representations. Experiments demonstrate that our method achieves competitive performances on five popular datasets for few-shot image classification. In particular, our method brings a 1. 74% improvement in accuracy over state-of-the-art methods on 5-way 1-shot.

Authors

• Mengping Dong
• Fei Li
• Zhenbo Li China Agricultural University
• Xue Liu McGill University

Keywords

• Few-shot learning
• Contrastive learning
• Semantic patchmix
• Image classification