Prototype-Neighbor Networks with task-specific enhanced meta-learning for few-shot classification.

As a promising technique for Few-Shot Classification (FSC), Prototypical Networks (PN) has gained increasing attention due to their simplicity and effectiveness. However, the unimodal prototypes derived from a few labeled data may lack representativeness and fail to capture complex data distributions. Inspired by KNN, a model-free classification algorithm, we propose a Neighbor Network (NN) to compensate for the limitations of PN. Specifically, NN classifies query samples based on their neighbors and optimizes the metric space to ensure that samples of the same class are grouped together. By combining PN and NN, we propose a Prototype-Neighbor Networks (PNN) to learn a better metric space where a few labeled data suffice to learn a reliable FSC model. To enhance adaptability to new classes, we improve the meta-learning mechanism by incorporating a task-specific fine-tuning phase between the meta-training and meta-testing stages. Additionally, we present a data augmentation method that combines PN and NN to generate pseudo-labeled data. Compared to self-training approaches, our method significantly reduces pseudo-label noise and confirmation bias. The proposed method has been validated on three benchmark datasets. Compared to 24 state-of-the-art FSC algorithms, PNN outperforms others on mini-imageNet, and CUB while achieving competitive results on tiered-imageNet. The experimental results on four medical image datasets further demonstrate the effectiveness of PNN on cross-domain tasks. The source code and related models are available at https://github.com/Dracula-funny/PNN.