Removing Adverse Background Shortcut with Text for Few-Shot Classification

Few-shot image classification is to categorize novel classes with limited training instances. A key hurdle in few-shot image classification arises from the disjoint nature of training and testing categories, which restricts the seamless transfer of knowledge acquired during base-class training to the classification of novel classes. Background information often serves as a convenient shortcut for typical classification. However, the inability to effectively leverage base-class knowledge for novel-class classification renders such shortcuts detrimental to the task of few-shot classification. In this paper, we present a simple yet effective approach to tackle this problem. By incorporating linguistic information, we adjust the image features, to eliminate background cue. Furthermore, we investigate the selection of texts that effectively capture the abstract features of the background in few-shot learning, aiming to provide a more robust soft-label for background. Our comprehensive experimental results validate the efficacy of our proposed method.