Cluster-HGNN: Deep Local Features Clustering for Few-Shot Image Classification With Hybrid Graph Neural Networks

Graph neural networks (GNNs) have shown great promise in few-shot learning, where they typically represent the entire feature of a sample as a node. However, this approach can overlook finer details within the sample, as GNNs usually measure the distance between nodes to determine overall differences, making them prone to background noise interference. To alleviate this limitation, we propose a novel framework based on a hybrid graph neural network for few-shot image classification, incorporating deep local features clustering (Cluster-HGNN). This framework comprises two types of GNNs: a global feature GNN and a local feature GNN. The former applies traditional label propagation to classify query nodes, while the latter treats feature embeddings as multiple deep local features, aggregating these by analyzing the categories of each query node’s K nearest neighbours to infer the sample’s category. The proposed design incorporates local features to rectify the final classification outcome, even in scenarios where the global features may inaccurately assess the current sample. Furthermore, the framework integrates multi-scale techniques and clustering to expand the feature space and reduce the influence of background noise on classification performance. As a result, Cluster-HGNN achieves state-of-the-art results on standard few-shot image classification benchmarks.