Feature Augmentation Reconstruction Network for Few-Shot Image Classification

In recent years, deep learning has achieved significant success in various computer vision tasks, notably in image classification. Nonetheless, labeled data is frequently insufficient in real-world applications, presenting significant impediments to the efficacy of deep learning models. This issue is especially pronounced in few-shot image classification tasks. To alleviate the above problem, we propose to augment features in few-shot image classification to increase sample diversity. This is achieved through learning and fitting the features' distribution, generating pseudo-data based on the distribution, and ultimately augmenting the features. Specifically, the adherence of the sample to the Gaussian distribution is learned using both local and global information, and then feature data is generated based on the distribution for feature augmentation. We evaluated the proposed method on four benchmark datasets, and the experimental results show that our method achieves state-of-the-art performance compared to prevailing few-shot image classification methods in most experimental settings.