Momentum memory contrastive learning for transfer-based few-shot classification

To improve the representation ability of feature extractors in few-shot classification, in this paper, we propose a momentum memory contrastive few-shot learning method based on the distance metric and transfer learning. The proposed method adopts an external memory bank and a contrastive loss function to constrain the feature representation of the samples in training. The memory bank is maintained by the dynamic momentum update of current samples. In addition, a feature representation augmentation technique is used to improve the generalization of the feature representation centroid to the samples in the testing. Furthermore, we design a spatial pyramid fusion downscaling module to improve the extraction ability of multi-scale features. Experimental results show that our method outperforms the compared methods and achieves state-of-the-art accuracy in 5-way 1-shot and 5-way 5-shot tasks on datasets including miniImageNet, CUB-200, and CIFAR-FS. The extensive study with discussions verifies the effectiveness of each proposed component in our method.