Local feature graph neural network for few-shot learning

The few-shot learning method based on local feature attention can suppress the irrelevant distraction in the global information and extract discriminating features. However, empirically defining the relationship between local features cannot fully utilize the power of local feature attention. This paper proposes a local feature graph neural network model (LFGNN), which uses the GNN to automatically extract and aggregate the relationship between different local parts and obtain features with stronger expressive ability for classification. Specifically, a sparse hierarchical connectivity graph is proposed to describe the relationship between features, in which the global features of all samples in the support set and the query set are connected in pairs, and the local features of each sample are only connected to the corresponding global features. Further, a multiple node-edge aggregation strategy is developed to learn a similarity metric. By integrating the edge loss with the classification loss, our LFGNN learns a better classifier to distinguish samples of novel classes. We conducted extensive experiments under the 5-way 1-shot and 5-way 5-shot setting on two benchmark datasets: miniImageNet, tieredImageNet. Experimental results demonstrate that the proposed approach is effective for boosting performance of meta-learning few-shot classification. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.