DouN-GNN:Double nodes graph neural network for few-shot learning

Abstract

In recent years, graph neural networks (GNNs) for few-shot learning have garnered significant attention due to their powerful learning capabilities. However, previous methods typically construct nodes using single-modal samples, often overlooking additional information (e.g., high-frequency details information) that can be provided by other modalities, which may limit model performance. To fully leverage multi-dimensional information from various sample modalities, we propose a novel double-node graph neural network (DouN-GNN). In our approach, each node comprises two sub-nodes, with each sub-node representing a different modality of the sample image. To address the issue of information redundancy between modalities when constructing sub-nodes, we introduce an orthogonal transformation method to orthogonalize the sub-node features. Additionally, we develop a graph update module for double-nodes, which alternately updates the nodes and edges of the graph to facilitate the aggregation of multi-dimensional information from multi-modal images. As the number of graph update layers increases, the edge features become more reliable, further enhancing performance. Extensive experiments on the miniImageNet, TieredImageNet, and CUB-200-2011 datasets demonstrate that our method outperforms existing state-of-the-art approaches.