Agent Attention Feature Reconstruction Network for Fine-Grained Few-Shot Image Classification

To develop fine-grained few-shot image classification, it is crucial to study how to increase inter-class variation and reduce intra-class variation. A recently proposed method called the Bidirectional Feature Reconstruction Network (BiFRN) increases inter-class variation by using support sets to reconstruct query sets and reduces intra-class variation by query sets to reconstruct support sets. In our study, we found a problem with BiFRN: The calculation of self-attention weights is affected by noise or inaccurate information, which may cause the model to over-focus on specific locations or ignore features at other locations. To solve this problem, we proposed an Agent Attention Bidirectional Feature Reconstruction Network (AAFRN), which extends BiFRN by fusing an Agent Attention Model to generate a new feature map for each feature map. Compared with BiFRN, our method can enhance feature expressiveness, facilitate global contextual information capture within feature maps and enable weighted feature fusion across various positions. The experimental results on three widely used fine-grained image classification datasets demonstrate that the proposed method achieves competitive performance compared to other methods.