Scholarlink

<p>Most existing few-shot segmentation methods based on the support-query matching framework suffer from insufficient support information, where the limited number and coverage of annotated samples often produce prototypes that are incomplete or contaminated by background features, leading to incomplete activation of target regions in the query image and false activation of irrelevant areas. To address this issue, we propose a class-aware prototype fusion network (CAPFN) for few-shot segmentation, comprising a class-aware module (CAM) and a prototype fusion module (PFM). The CAM extracts class-specific information by jointly utilizing support features, masks, and preliminary query predictions, thereby guiding the model to attend precisely to target regions. To further mitigate semantic gap between the support and query domains, the PFM constructs a hybrid prototype by fusing support-derived and query-derived prototypes based on initial predictions. This fusion enhances prototype quality, reduces information loss, and improves the discriminative capacity of query activation. Extensive experiments on benchmark datasets (PASCAL-5<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({}^i\)</EquationSource> </InlineEquation> and COCO-20<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({}^i\)</EquationSource> </InlineEquation>) demonstrate the superiority of our approach. Notably, with ResNet50 as the backbone, our model achieves a 2.32% improvement in mIoU over the baseline on the COCO-20<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({}^i\)</EquationSource> </InlineEquation> dataset. The code is available at <a href="https://github.com/ShuoWang011/CAPFN.git">https://github.com/ShuoWang011/CAPFN.git</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

Few-shot segmentation network based on class-aware prototype fusion

  • 期刊论文

  • 发表日期：

    2025年9月23日

查看全文
  • Aiping Yang,
  • Shuo Wang,
  • Zijia Sang,
  • Yaran Zhou

摘要

Most existing few-shot segmentation methods based on the support-query matching framework suffer from insufficient support information, where the limited number and coverage of annotated samples often produce prototypes that are incomplete or contaminated by background features, leading to incomplete activation of target regions in the query image and false activation of irrelevant areas. To address this issue, we propose a class-aware prototype fusion network (CAPFN) for few-shot segmentation, comprising a class-aware module (CAM) and a prototype fusion module (PFM). The CAM extracts class-specific information by jointly utilizing support features, masks, and preliminary query predictions, thereby guiding the model to attend precisely to target regions. To further mitigate semantic gap between the support and query domains, the PFM constructs a hybrid prototype by fusing support-derived and query-derived prototypes based on initial predictions. This fusion enhances prototype quality, reduces information loss, and improves the discriminative capacity of query activation. Extensive experiments on benchmark datasets (PASCAL-5 \({}^i\) and COCO-20 \({}^i\) ) demonstrate the superiority of our approach. Notably, with ResNet50 as the backbone, our model achieves a 2.32% improvement in mIoU over the baseline on the COCO-20 \({}^i\) dataset. The code is available at https://github.com/ShuoWang011/CAPFN.git.