MCANet: A Multi-Scale Class-Related Attention Network for Few-Shot Learning

Among various few-shot learning methods, metric learning approaches based on local feature representation of images have achieved remarkable results. However, most of these methods coarsely compare support sets and query sets at a single scale, failing to consider the contextual information of image data and the diversity of multi-scale information. To address these issues, this paper proposes a novel Multi-scale Class-related Attention Network for few-shot learning, referred to as MCANet. In MCANet, we first construct a multi-scale feature extraction module that adopts a bottom-up architecture to capture image features from multiple scales and levels, thereby combining high-level semantic representations with low-level visual information. The images are then represented as sets of local features at different scales. Next, we utilize data co-occurrence information to select class-related features for each category, highlighting semantically dominant objects at various scales while reducing the influence of irrelevant local features. Finally, we integrate contextual information to fuse similarity measures across different scales for comparing query and support sets. On the miniImageNet dataset, our MCANet achieves accuracies of 71.08% (5-way 1-shot) and 83.34% (5-way 5-shot), and demonstrates competitive performance on other popular benchmarks. Experimental results indicate that the proposed method is effective for few-shot learning.