Abstract: Few-shot learning aims to learn to classify new categories based on just one or a few examples. To address this special problem, data augmentation is often used as a direct and effective approach, and augmented data should be diverse and discriminable. This paper proposes a new two-stage data augmentation method based on transferring the intra-class variations of base classes. The learning process is divided into representation learning stage and few-shot learning stage. In the representation learning stage, representations of instance-specific features of the base class data are obtained through self-supervised tasks, and class-specific features are obtained through supervised tasks. Intra-class variations of the base class data are calculated through these two features and distributions of intra-class variations for each class of the base class are modeled . In the few-shot learning stage, the model samples task related intra-class variation information from the intra-class variation distributions of the base class and adds it to the few-shot features to enhance the few-shot data. The experimental results show that in 1-shot, the classification performance of our method on the miniImageNet, tieredImageNet and CUB datasets is improved by 4% to 7% compared to the baseline model, and by 3% to 7% in 5-shot. Compared with other existing data augmentation methods, it also demonstrates competitive performance. This indicates that the generated enhanced data can enhance the diversity of few-shot data while maintaining discriminability, verifying the feasibility and effectiveness of our method.

摘要： 小样本学习致力于通过极少数量的训练样本，甚至一个样本，来实现对新类数据的分类。面对这种挑战，数据增强成为小样本学习中一种直接而有效的解决方法，但是确保增强数据的多样性和可辨别性是数据增强的关键。为此提出一种基于迁移基类类内变化的两阶段数据增强方法，分为特征学习和小样本学习阶段。在特征学习阶段，模型通过自监督任务学习基类数据的个体特征的表达，有监督任务则学习类辨别特征，模型通过这两种特征获得基类数据的类内变化并建模基类每类的类内变化分布。在小样本学习阶段，模型从基类的类内变化分布中采样与任务相关的类内变化信息并添加到小样本特征中，以实现增强小样本数据的目的。实验结果表明，1-shot情况下，本文方法在miniImageNet、tieredImageNet和CUB数据集上的分类性能相较于基线模型提升了4%到7%，5-shot情况下提升了3%到7%，相较于其他数据增强方法，也展现了具有竞争力的性能，这表明生成的增强数据可以在保持可辨别性的同时增强小样本数据的多样性，并验证了该方法的可行性和有效性。