Abstract:The aim of few-shot learning (FSL) is to achieve efficient learning via a small number of labeled samples. However, it remains challenging to deal with intra-class differences and inter-class similarity. In particular, FSL yields unsatisfactory performance in handling complex data distributions and fine-grained classification tasks. To address the problems of large intra-class differences and high inter-class similarity in FSL, this study proposes a novel model, the AQH-GCN model. This model proposes three key technologies to enhance the FSL performance, and optimization methods to deal with the problems of intra-class differences and inter-class similarity. (1) An adaptive aggregation mechanism with dynamic weight adjustment is designed to effectively improve the intra-class feature consistency and suppress intra-class interference. (2) The node pooling strategy integrating contrastive learning enhances the model’s ability to retain discriminative nodes. (3) A multi-level feature fusion framework is constructed to achieve the dynamic optimization of multi-level feature expression. Comparative experiments on multiple standard datasets validate the superiority of the proposed method. In particular, in fine-grained classification tasks, the performance is significantly improved. For instance, compared with the mainstream baseline model CAN on miniImageNet dataset, the accuracy of this method increases by 4.38 percentage points in the 5-way 1-shot task and 4.52 percentage points in the 5-way 5-shot task.