Transformer-Based Few-Shot Learning for Image Classification

Abstract

Few-shot learning (FSL) remains a challenging research endeavor. Traditional few-shot learning methods mainly consider the distance relationship between the query set and the support set, while the context information between different support sets are not fully exploited. This paper proposes a Transformer-based few-shot learning method (TML). By taking advantage of the self-attention mechanism of Transformer, TML effectively exploits the correlation between support sets so as to learn highly discriminative global features. Furthermore, in order to cope with the overfitting problem introduced by the increase of model complexity, we introduce a classification loss into the total loss function as a regularization term. To overcome the limit of traditional cross-entropy loss, a label refinement method is used to refine the label assignment for classification. The experimental results show that TML improves the ability of learning hard samples and achieves higher classification accuracy than existing state-of-the-art few-shot learning methods.