Learning to Distinguish: A General Method to Improve Compare-Based one-shot Learning Frameworks for Similar Classes

Abstract

How to recognize unseen classes given only few examples remains a challenge in image classification. One-shot learning, different from the standard paradigm that deals with unseen examples of learned classes, is a new area and developing rapidly. However, we notice that existing frameworks abandon the inter-class information to gain the generalization ability in new classes. But this information is important in distinguishing similar classes, like fine-grained images. So we propose a method focused on hard examples to regain it. Firstly, we select out hard examples via comparison scores from the basic network, which belong to the similar classes that the basic network cannot distinguish well. And then we design a targeted loss function that enlarges inter-class gap to fortify the network to distinguish similar classes. The method can be applied to existing and forthcoming compare-based One-shot learning frameworks generally. And the experiments prove that our algorithm improves state-of-the-art networks on miniImageNet and several fine-grained datasets.