Few-shot Classification Network Based on Feature Differentiation

Aiming at the existing problems in the few-shot learning methods which treat samples in an isolated perspective and ignore the difference information between samples, we propose a few-shot learning classification network based on feature differentiation. A new feature adaptive fusion module and a feature conversion module form our network, where the former is proposed to fuse global information and detailed features, and the latter one marks the semantic features which have high recognition, so as to narrow the semantics within the same category and widen the semantic gap between different categories. CUB dataset and mini-ImageNet dataset were used in the experiment, and the accuracy of 5way-lshot and 5way-5shot tasks respectively achieved 57.63%, 76.54% and 54.39%, 73.19%. Experimental results show that our method can further learn how to distinguish different category concepts through differentiated features, thus the proposed few-shot learning model has higher accuracy and robustness.