Few-Shot Learning Based on Attention Relation Compare Network

Abstract

In the field of computer vision, great progress has been made in the classification of few-shot images recently. However, challenges in improving classification accuracy and training stability remain. To address this issue, a novel meta learning method aiming at few-shot was proposed in this paper, which is a relationship comparison network based on attention mechanism. The network is an end-to-end network, consisting of three folds: feature coding, feature combination, and relationship coding. Firstly, the attention mechanism is introduced into the feature coding module and spectral normalization is used to normalize the weight parameters of the convolutional layer of the module, resulting in better extraction of key feature information from few-shot images and enhanced training stability of the attention mechanism comparison network. Secondly, in order to compare the feature information of the sample set and the query set, the feature information of the extracted sample set and the query set is combined into a new feature information in the feature combination module. Besides, the relationship coding module can learn a depth measurement method to calculate the relationship score of combined feature information and realize the classification of few-shot images. The performance of Attention Relation Compare Network Meta-Learning was evaluated on few-shot data set, namely the Omniglot and Mini ImageNet. It was indicated that our method excelled other few-shot learning method in accuracy and efficiency and improve the training stability to some extent . Experiments show that our method produces very competitive results when compared to the state-of-the-art method, and our model also exhibits the training stability.