An Attention Mechanism-based Relation Network for Few-Shot Image Classification

Abstract

The key to solving the few-shot image classification problem is learning image category information from a handful of image samples. Few-shot image classification can easily cause overfitting problems that impact classification performance due to insufficient labeled data. In this paper, an attention mechanism-based relation network model for few-shot image classification is proposed. Inspired by classic methods of relation networks for few-shot learning, the attention mechanism in a convolutional block attention mechanism (CBAM) for metric learning is utilized in the feature extraction network. Then, an update strategy selecting a validation set during the training is adopted to reduce the possibility of overfitting. Through comparison experiments on different datasets, the results demonstrate that our model has better accuracy than traditional methods.