ReNDCF: Relation network with dual-channel feature fusion for few-shot learning

Yi Xu, Wenke Chu, Mingyue Lu

Abstract: RelationNet is a highly effective metric-based few-shot learning method. However, RelationNet uses only shallow convolutional networks in the feature extraction stage, yielding less representative sample features. In addition, RelationNet simply uses summing or averaging to compute prototypes in the prototype learning stage, making it difficult to obtain very representative prototypes. To address these two aspects, we propose a Relation Network with Dual-Channel Feature Fusion for Few-shot Learning(ReNDCF). Firstly, in the feature extraction stage, a Dual-Channel feature extractor is proposed and a multi-head self-attention mechanism is introduced to enhance the feature extraction ability for fine-grained samples by exploiting the correlation between sample categories. Secondly, in the prototype learning module, an improved adaptive prototype learner is proposed, which further enables the final prototype to represent a certain class of samples more accurately. Finally, our ReNDCF achieves better classification performance than RelationNet and other state-of-the-art classification methods on three widely used fine-grained benchmark datasets CUB-200-2011, Stanford-Cars, and Stanford-Dogs.