电子测量与仪器学报

任杰,迟荣华,李红旭.基于多尺度特征提取和注意力机制的轻量化晶圆缺陷检测方法[J].电子测量与仪器学报,2025,39(8):13-21

基于多尺度特征提取和注意力机制的轻量化晶圆缺陷检测方法

Lightweight wafer defect detection method based on multi-scalefeature extraction and attention mechanism

DOI：

英文关键词:wafer defect multi-scale feature extraction attention mechanism deep learning

基金项目:江苏省基础研究计划重点项目（BK20243021）、江苏省产学研合作项目（BY20230745）、江苏省高等学校基础科学研究面上项目（22KJB510043）、无锡市科技创新创业资金“太湖之光”科技攻关计划（K20241049）、无锡学院引进人才科研启动专项经费（550222001, 550223012）项目资助

作者	单位
任杰	南京信息工程大学电子与信息工程学院南京210044
迟荣华	无锡学院江苏省通感融合光子器件及系统集成工程研究中心无锡214105
李红旭	无锡学院江苏省通感融合光子器件及系统集成工程研究中心无锡214105

Author	Institution
Ren Jie	School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
Chi Ronghua	Jiangsu Engineering Research Center for Sensor Fusion Photonic Devices and System Integration, Wuxi University, Wuxi 214105, China
Li Hongxu	Jiangsu Engineering Research Center for Sensor Fusion Photonic Devices and System Integration, Wuxi University, Wuxi 214105, China

摘要点击次数: 839

全文下载次数: 507

中文摘要:

在半导体制造中，晶圆图缺陷检测至关重要，能够对缺陷进行快速定位，实现对缺陷的识别，对于提升晶圆产品质量和生产效率具有意义。然而，现有方法存在局限性，如模型过于庞大，网络模型深度过深，难以充分利用多层次特征进行精确分类。为了解决这些问题，结合了Stem-Dense特征提取模块和多尺度注意力特征融合结构，提出了一种新型网络结构——MSD-DFE。MSD-DFE通过Stem-Dense的密集连接结构和多尺度注意力特征融合技术，有效提取丰富的浅层特征信息，同时显著降低模型的参数量和计算复杂度。多尺度特征提取模块融合了不同尺度下的晶圆图信息，增强了模型对不同层次缺陷特征的提取能力。此外，引入的注意力机制使得模型能够更关注晶圆图存在缺陷区域，从而提升分类精度。实验结果表明，在减少参数量和计算量的前提下，MSD-DFE在WM-811K数据集上达到了97.4%的平均准确率，优于现有主流方法，表明其在实际生产环境中具有较高的应用潜力。

英文摘要:

In semiconductor manufacturing, wafer map defect detection is crucial for the rapid localization and identification of defects, which is significant for enhancing wafer product quality and production efficiency. However, existing methods have limitations, such as overly complex models and excessively deep network structures that struggle to leverage multi-level features for accurate classification. To address these issues, this paper combines a Stem-Dense feature extraction module with a multi-scale attention feature fusion module to propose a novel network architecture—multi-scale defect detection network with enhanced feature extraction (MSD-DFE). MSD-DFE effectively captures rich shallow feature information through the dense connection structure of Stem-Dense and multi-scale attention-based feature fusion technology, while significantly reducing the number of parameters and computational complexity of the model. The multi-scale feature extraction module integrates wafer map information from various scales, enhancing the model’s ability to extract defect features. Additionally, the introduced attention mechanism allows the model to focus more on defect areas, thereby improving classification accuracy. Experimental results show that MSD-DFE achieves an average accuracy of 97.4% on the WM-811K dataset, outperforming current mainstream methods, indicating its high potential for practical application in industrial settings.

查看全文查看/发表评论下载PDF阅读器