Multi-scale feature self-enhancement network for few-shot learning

摘要

The goal of few-shot learning(FSL) is to learn from a hand of labeled examples and quickly adapt to a new task. The traditional FSL models use the single-scale feature that does not have strong representative ability. Besides, some previous methods construct graph neural network to get better classifications, while they update nodes indiscriminately, which will result in intra-class information passing between inter-class nodes. In this paper, we propose a new method called Multi-scale Feature Self-enhancement Network(MFSN) for few-shot learning, which extracts multi-scale feature through a novel extractor, and then enhance the multiple features by the selective graph neural networks that can filter out the incorrect passings between nodes through a meta-learner. At last, classification is performed by measuring distances between the augmented unlabeled features and the improved prototypes computed from augmented labeled features. Comparing to the traditional method, our method improves 1-shot accuracy by 11.8% and improves 5-shot by 10.3% on MiniImagenet dataset. Experiments on MiniImagenet, Cifar-100, and Caltech-256 datasets show the effectiveness of the proposed model.