Three-Dimension Attention Mechanism and Self-Supervised Pretext Task for Augmenting Few-Shot Learning

Abstract

The main challenge of few-shot learning lies in the limited labeled sample of data. In addition, since image-level labels are usually not accurate in describing the features of images, it leads to difficulty for the model to have good generalization ability and robustness. This problem has not been well solved yet, and existing metric-based methods still have room for improvement. To address this issue, we propose a few-shot learning method based on a three-dimension attention mechanism and self-supervised learning. The attention module is used to extract more representative features by focusing on more semantically informative features through spatial and channel attention. Self-supervised learning mainly adopts a proxy task of rotation transformation, which increases semantic information without requiring additional manual labeling, and uses this information for training in combination with supervised learning loss function to improve model robustness. We have conducted extensive experiments on four popular few-shot datasets and achieved state-of-the-art performance in both 5-shot and 1-shot scenarios. Experiment results show that our work provides a novel and remarkable approach to few-shot learning.

Keywords

• Few-shot
• self-supervised pretext task learning
• deep learning
• image classification
• attention mechanism