The task of few-shot learning (FSL) is to classify never-before-seen samples by merely depending on few examples. This work, called the Relative Relation Network ( \(R^{2}Net\) ), is presented to introduce the relative metric concept into the design of FSL architecture in order to improve discriminability of the model for classification based on limited support samples. Concretely, we first perform a intra-class local enhancement strategy on the same kind of query images before the feature embedding network during training stage. Then the concatenated local feature descriptors of query images and support samples are fed into the proposed relative relation module to learn relative relation scores. This design can effectively operate under the challenging and restrictive conditions of the FSL tasks for enabling the enhancement of sample features within the same class and obtaining more discriminative scores under the comparison of few support samples from different classes. Furthermore, an optimal permutation mechanism of support samples is proposed to obtain the strong representational support samples in the testing stage. It is worth to note that \(R^{2}Net\) is trained end-to-end from scratch, which can be easily embedded into any episodic training manners. Extensive experiments on FSL tasks demonstrate that the superiority of the proposed methods over the state-of-the-art benchmarks.
