Scholarlink

<p>Generalized Zero-Shot Learning (GZSL) aims to recognize objects from both seen and unseen classes using data from only seen classes. Current feature generation methods for GZSL often suffer from category bias and a lack of discriminative information. To address these issues, we propose a novel GZSL method based on Semantic Contrast and Feature Aggregation (SCFA-GZSL). This method integrates a feature generation module, an embedding module, a relation module, a feature aggregation module, and a semantic consistency contrast module. The semantic consistency contrast module employs contrastive learning to strengthen the correlation between original semantic features and synthetic visual features, enhancing intra-class compactness and inter-class separability. The feature aggregation module aligns embedding features with the centroid of real visual features, reducing the distribution gap between synthetic and real features. Experimental results on benchmark datasets demonstrate that our approach significantly outperforms state-of-the-art methods, particularly for unseen classes, highlighting its effectiveness in improving semantic consistency and alleviating category bias. Detailed code and data are available through <a href="https://github.com/SAYXNER/SCFA-GZSL/">https://github.com/SAYXNER/SCFA-GZSL/</a>.</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

Enhancing generalized zero-shot learning through semantic contrast and feature aggregation

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Xiyu Yang,
  • Fang Zhou,
  • Jing Fang,
  • Mengdao Xing

摘要

Generalized Zero-Shot Learning (GZSL) aims to recognize objects from both seen and unseen classes using data from only seen classes. Current feature generation methods for GZSL often suffer from category bias and a lack of discriminative information. To address these issues, we propose a novel GZSL method based on Semantic Contrast and Feature Aggregation (SCFA-GZSL). This method integrates a feature generation module, an embedding module, a relation module, a feature aggregation module, and a semantic consistency contrast module. The semantic consistency contrast module employs contrastive learning to strengthen the correlation between original semantic features and synthetic visual features, enhancing intra-class compactness and inter-class separability. The feature aggregation module aligns embedding features with the centroid of real visual features, reducing the distribution gap between synthetic and real features. Experimental results on benchmark datasets demonstrate that our approach significantly outperforms state-of-the-art methods, particularly for unseen classes, highlighting its effectiveness in improving semantic consistency and alleviating category bias. Detailed code and data are available through https://github.com/SAYXNER/SCFA-GZSL/.