Scholarlink

<p>Deep learning has achieved significant progress in computer vision; however, its dependence on large-scale, labeled datasets limits its application in scenarios with scarce annotations. Few-shot learning (FSL) addresses this issue by enabling generalization from only a few samples, but many existing methods rely on simple feature aggregation, which overlooks fine-grained semantic cues. To overcome these limitations, we propose APNet that combines a strengthened backbone with prototype-based meta-learning. In the classification training stage, a ResNet-12_SCAA backbone equipped with a Spatial-Channel Adaptive Attention (SCAA) module captures cross-channel dependencies and long-range spatial correlations, producing richer feature representations. In the meta-learning stage, we introduce Local Importance-based Pooling (LIP) with local–global fusion to dynamically recalibrate support and query features, preserving discriminative local details while incorporating global context. To further enhance generalization, a prototype alignment loss enforces intra-class compactness and inter-class separability in the embedding space. Extensive experiments on MiniImageNet, TieredImageNet, CUB-200-2011, and CIFAR-FS show that APNet achieves consistent improvements over state-of-the-art methods, particularly under the challenging 1-shot setting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

APNet: attention-enhanced representation and prototype alignment for few-shot learning

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Jianlie Lou,
  • Ailan Wu,
  • Jianjun Zhang,
  • Hongfeng Han

摘要

Deep learning has achieved significant progress in computer vision; however, its dependence on large-scale, labeled datasets limits its application in scenarios with scarce annotations. Few-shot learning (FSL) addresses this issue by enabling generalization from only a few samples, but many existing methods rely on simple feature aggregation, which overlooks fine-grained semantic cues. To overcome these limitations, we propose APNet that combines a strengthened backbone with prototype-based meta-learning. In the classification training stage, a ResNet-12_SCAA backbone equipped with a Spatial-Channel Adaptive Attention (SCAA) module captures cross-channel dependencies and long-range spatial correlations, producing richer feature representations. In the meta-learning stage, we introduce Local Importance-based Pooling (LIP) with local–global fusion to dynamically recalibrate support and query features, preserving discriminative local details while incorporating global context. To further enhance generalization, a prototype alignment loss enforces intra-class compactness and inter-class separability in the embedding space. Extensive experiments on MiniImageNet, TieredImageNet, CUB-200-2011, and CIFAR-FS show that APNet achieves consistent improvements over state-of-the-art methods, particularly under the challenging 1-shot setting.