Feature Enhanced Graph Neural Network for Few-Shot Image Classification

There has been a rising trend of solving few-shot image classification problems utilizing graph neural networks (GNN) in recent years. However, most GNN-based approaches fail to fully exploit relationships among samples, resulting in poor image classification. To address this problem, we propose a feature-enhanced GNN model appropriate for few-shot image classification tasks. The suggested method first exploits an efficient convolutional block to generate accurate feature maps, enhancing the expressivity of the feature extraction module. Then, our technique employs flexible combinatorial distance metric functions to compute the exact relation score between samples and determine image relevance by minimizing the matching cost. Moreover, a multilayer perceptron based on the residual structure with an attention mechanism is developed to produce focused feature representations, allowing the model to obtain selective and relevant information among the samples. The proposed model is evaluated on a supervised few-shot image classification task utilizing four benchmark datasets, with the corresponding results demonstrating that our model achieves a higher accuracy performance than traditional few-shot image classification methods.