Bidirectional Patch-Aware Attention Network for Few-Shot Learning

Few-shot learning (FSL) aims to train a model using a minimal number of samples and subsequently apply this model to recognize unseen classes. Recently, metric-based methods mainly focus on exploring the relationship between the support set and the query set through attention mechanism in solving FSL problems. However, these methods typically employ unidirectional computation when calculating the attention relationship between support and query. This unidirectional approach not only limits the depth and breadth of knowledge acquisition but may also lead to mismatched patches between support and query, thereby affecting the overall performance of the model. In this article, we propose a bidirectional patch-aware attention network for few-shot learning (BPAN) to address this issue. First, we extract subimages via grid cropping and feed them into the learned feature extractor to obtain patch features. Moreover, self-attention is used to assign different weights to patch features and reconstruct them. Then, PFCAM is proposed to mutually explore the patch feature relationship between the support set and the support set, further reconstruct the patch features, and aggregate multiple patch features of each image into one feature through a learnable parameter matrix for the purpose of prediction. Finally, the template for each class is constructed to extend the results of PFCAM to the few-shot classification scenario. Experiments on three benchmark datasets show that BPAN achieves superior performance.