Scholarlink

<p>In the semiconductor manufacturing process, wafer defect inspection is crucial for ensuring product quality. However, due to the high cost and scarcity of defect samples, how to effectively enhance detection performance under data-scarce scenarios has become a significant challenge. This study proposes an innovative approach for the YOLOv11 baseline detection network to address the small-sample problem in wafer defect detection. Firstly, we adopt the Multi-Domain Offset Normalization (MDON) technique, which avoids the chaotic data distribution caused by traditional data augmentation by performing batch normalization on each augmented image domain independently, allowing for learning the most critical consistency information within features. Secondly, a simulation-to-reality sample domain alignment migration learning method based on Maximum Mean Discrepancy (MMD) is introduced. By jointly inputting simulated and real samples and incorporating MMD loss, this method reduces the gap between simulation and reality, thereby enhancing the model’s generalization ability on real-world samples. Experimental results demonstrate that the effective combination of these two methods significantly improves the performance of YOLOv11 in wafer defect detection under data-scarce scenarios, offering a new avenue for industrial applications.</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

A domain optimization approach for few-shot defect inspection on patterned wafer

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Leisheng Chen,
  • Kai Meng,
  • Hangying Zhang,
  • Yangyang Zou,
  • Peihuang Lou,
  • Xing Wu

摘要

In the semiconductor manufacturing process, wafer defect inspection is crucial for ensuring product quality. However, due to the high cost and scarcity of defect samples, how to effectively enhance detection performance under data-scarce scenarios has become a significant challenge. This study proposes an innovative approach for the YOLOv11 baseline detection network to address the small-sample problem in wafer defect detection. Firstly, we adopt the Multi-Domain Offset Normalization (MDON) technique, which avoids the chaotic data distribution caused by traditional data augmentation by performing batch normalization on each augmented image domain independently, allowing for learning the most critical consistency information within features. Secondly, a simulation-to-reality sample domain alignment migration learning method based on Maximum Mean Discrepancy (MMD) is introduced. By jointly inputting simulated and real samples and incorporating MMD loss, this method reduces the gap between simulation and reality, thereby enhancing the model’s generalization ability on real-world samples. Experimental results demonstrate that the effective combination of these two methods significantly improves the performance of YOLOv11 in wafer defect detection under data-scarce scenarios, offering a new avenue for industrial applications.