Scholarlink

<p>In semiconductor manufacturing, rapidly identifying process faults through wafer map defect recognition can significantly improve production yield. However, annotating wafer map defect types—especially complex mixed-type defects—requires skilled technicians and substantial time, leading to high annotation costs. To address this issue, we present a transductive zero-shot learning method that classifies mixed-type defects using only labeled samples of single-type defects. The presented technique eliminates the need for labeled mixed-type defects by leveraging semantic information to bridge known classes (single-type) and unknown classes(mixed-type). This directly translates to reduced time and annotation costs, as well as faster process diagnosis in production environments. We introduce three key strategies to improve classification accuracy: (1) collaborative optimization of the visual feature extractor and semantic embedder, (2) iterative updating of the semantic space, and (3) progressive pseudo-labeling for retraining. Extensive experiments demonstrate that the proposed method substantially surpasses previous transductive zero-shot learning methods, particularly on mixed-type defects.</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

Transductive zero-shot learning for mixed-type defect classification in wafer maps

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Jun Liu,
  • Jifei Lu,
  • Tian Chen,
  • Xi Wu,
  • Huaguo Liang,
  • Xiaohui Yuan,
  • Yen Pham

摘要

In semiconductor manufacturing, rapidly identifying process faults through wafer map defect recognition can significantly improve production yield. However, annotating wafer map defect types—especially complex mixed-type defects—requires skilled technicians and substantial time, leading to high annotation costs. To address this issue, we present a transductive zero-shot learning method that classifies mixed-type defects using only labeled samples of single-type defects. The presented technique eliminates the need for labeled mixed-type defects by leveraging semantic information to bridge known classes (single-type) and unknown classes(mixed-type). This directly translates to reduced time and annotation costs, as well as faster process diagnosis in production environments. We introduce three key strategies to improve classification accuracy: (1) collaborative optimization of the visual feature extractor and semantic embedder, (2) iterative updating of the semantic space, and (3) progressive pseudo-labeling for retraining. Extensive experiments demonstrate that the proposed method substantially surpasses previous transductive zero-shot learning methods, particularly on mixed-type defects.