Few-Shot Classification Study for Prototype Fusion and Completion

Deep learning models face significant challenges in image classification due to the limited availability of training samples. To address this issue, few-shot learning, which enables model training with a small number of samples, has emerged. When applied to classification tasks, it is referred to as few-shot classification. Prototype networks are an effective approach for few-shot classification, however, the prototypes obtained by averaging sparse samples often lack sufficient representativeness, leading to reduced classification accuracy. To tackle these challenges, this paper proposes a novel prototype fusion completion network (ProfcNet). The proposed method enhances prototype representation and improves class representativeness using only the data of the prototype itself. Specifically, the network utilizes a fusion module to generate fusion feature maps with more class-specific characteristics, and a completion module to refine and complete the prototype using these fusion feature maps, yielding the final prototype. Additionally, an inference testing method tailored for the prototype fusion completion network is introduced. Extensive experiments on four publicly available few-shot classification datasets demonstrate that the proposed method significantly improves classification accuracy compared to prototype networks, achieving accuracy gains of nearly 30% and 20% in 1-shot and 5-shot classification tasks, respectively. Furthermore, ablation studies and visualization experiments further validate the effectiveness of the individual modules within the network. © 2013 IEEE.