Scholarlink

<p>Few-shot learning (FSL) is a critical challenge in computer vision, aiming to classify novel classes using only a limited number of labeled examples. Prototypical networks are known for their simple architecture and robust performance. However, most existing prototypical methods rely on averaging a small number of labeled samples and fail to fully leverage intra-class and inter-class distribution information. This limitation generates redundant features by assuming equal importance for all samples. As a result, it increases the risk of overfitting and reduces generalization capability. To address these challenges, we propose Graph-Based Co-Attention Discriminative Prototype Learning Networks (GCDPL-Net), a novel meta-learning framework designed to enhance discriminative power and generate more representative and distinct prototypes. This approach improves classification accuracy for unseen queries. GCDPL-Net introduces three core modules, i.e., the Co-Attention Discriminative (CoAD) module, the Instance Graph Prototypical Network (IGPN) module, and the Attentive Weighted Prototypical Graph Network (AWPGN) module. CoAD integrates intra-class and inter-class self-attention to construct an informative and discriminative representation space. IGPN precisely evaluates and emphasizes the significance of individual samples within each class. AWPGN captures and refines inter-class relationships. Additionally, GCDPL-Net incorporates a memory attention layer to enhance scalability and robustness against overfitting and over-smoothing while effectively managing long-range dependencies. Experimental results on five standard FSL benchmark datasets demonstrate that GCDPL-Net achieves state-of-the-art or competitive performance compared to recent approaches.</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

GCDPL-Net: graph-based co-attention discriminative prototype learning networks for few-shot classification

  • 期刊论文

  • 发表日期：

    2025年8月9日

查看全文
  • Mohammed Al-Habib,
  • Zuping Zhang,
  • Abdulrahman Noman,
  • Raeed Al-sabri,
  • Majjed Al-Qatf

摘要

Few-shot learning (FSL) is a critical challenge in computer vision, aiming to classify novel classes using only a limited number of labeled examples. Prototypical networks are known for their simple architecture and robust performance. However, most existing prototypical methods rely on averaging a small number of labeled samples and fail to fully leverage intra-class and inter-class distribution information. This limitation generates redundant features by assuming equal importance for all samples. As a result, it increases the risk of overfitting and reduces generalization capability. To address these challenges, we propose Graph-Based Co-Attention Discriminative Prototype Learning Networks (GCDPL-Net), a novel meta-learning framework designed to enhance discriminative power and generate more representative and distinct prototypes. This approach improves classification accuracy for unseen queries. GCDPL-Net introduces three core modules, i.e., the Co-Attention Discriminative (CoAD) module, the Instance Graph Prototypical Network (IGPN) module, and the Attentive Weighted Prototypical Graph Network (AWPGN) module. CoAD integrates intra-class and inter-class self-attention to construct an informative and discriminative representation space. IGPN precisely evaluates and emphasizes the significance of individual samples within each class. AWPGN captures and refines inter-class relationships. Additionally, GCDPL-Net incorporates a memory attention layer to enhance scalability and robustness against overfitting and over-smoothing while effectively managing long-range dependencies. Experimental results on five standard FSL benchmark datasets demonstrate that GCDPL-Net achieves state-of-the-art or competitive performance compared to recent approaches.