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

Yan Zhang, Xudong Zhou, Nian Wang, Jun Tang, Tao Xuan

Abstract: Highlights•We propose a novel double-node graph neural network with an improved graph update strategy, which optimizes sample topological relationships by alternately updating nodes and edges to aggregate more reliable multidimensional information.•We introduce gray image features as sub-nodes and design an orthogonal transformation module to minimize redundancy between different modality sub-nodes, allowing each sub-node to focus on expressing discriminative information specific to its modality.•The DouN-GNN demonstrates strong performance across three benchmark datasets: miniImageNet, TieredImageNet, and CUB-200-2011. Compared to existing state-of-the-art few-shot learning classification networks, it demonstrates a significant improvement in performance.