Hierarchical few-shot learning based on coarse- and fine-grained relation network

摘要

Few-shot learning plays an important role in the field of machine learning. Many existing methods based on relation network achieve satisfactory results. However, these methods assume that classes are independent of each other and ignore their relationship. In this paper, we propose a hierarchical few-shot learning model based on coarse- and fine-grained relation network (HCRN), which constructs a hierarchical structure by mining the relationship among different classes. Firstly, we extract deep and shallow features from different layers at a convolutional neural network. The shallow feature information contains more common features among similar classes, while the deep feature information is more specific. The complementary of these different types of data features can effectively construct coarse- and fine-grained structures by clustering. Secondly, we design coarse- and fine-grained relation networks to classify according to the guidance of the hierarchical structure. The hierarchical class structure learned from data is important auxiliary information for classification. Experimental results show that HCRN can outperform several state-of-the-art models on the Omniglot and miniImageNet datasets. Especially, HCRN obtains 6.47 improvement over the next best under the 5-way 1-shot setting on the miniImageNet dataset.