| 陈晓雷,温润玉,杨富龙,李正成,沈星阳.基于空频特征融合的双流晶圆缺陷分类网络[J].电子测量与仪器学报,2024,38(8):56-67 |
| 基于空频特征融合的双流晶圆缺陷分类网络 |
| Dual-stream wafer defect classification network based onspatial and frequency domains feature fusion |
| |
| DOI： |
| 中文关键词: 晶圆缺陷分类 双流网络 小波变换 注意力机制 卷积神经网络 |
| 英文关键词:wafer defect classification dual-stream network wavelet transform attention mechanism convolutional neural network |
| 基金项目:甘肃省科技计划资助（24JRRA179）、甘肃省科技重大专项（23ZDGE001）资助项目 |
|
|
| Author | Institution |
| Chen Xiaolei | College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730000, China |
| Wen Runyu | College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730000, China |
| Yang Fulong | College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730000, China |
| Li Zhengcheng | College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730000, China |
| Shen Xingyang | College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730000, China |
|
| 摘要点击次数: 26 |
| 全文下载次数: 107 |
| 中文摘要: |
| 晶圆缺陷模式分类在晶圆制造过程中扮演着至关重要的角色，准确识别晶圆缺陷能够确定缺陷产生的根本原因，进而定位生产流程中的问题。然而，现有深度学习晶圆缺陷分类方法仅从空间域或者频率域出发进行网络设计，未能实现空频信息的相互补充与融合，限制了晶圆缺陷分类准确性的进一步提高。针对这一问题，提出了一种基于空间域和频率域特征融合的双流晶圆缺陷分类网络—SFWD-Net。该网络利用提出的多尺度特征提取卷积模块和多视角注意力模块构成空间流分支提取晶圆图的空间信息，利用离散小波变换构成频率流分支提取晶圆图的频率信息，空频信息融合后再进行缺陷分类。在大规模半导体晶圆图数据集WM-811K的实验证明，SFWD-Net由于同时从空间域和频率域出发进行网络设计，其分类准确度达到99.299 2%，优于其他5种先进方法，能够显著提高晶圆缺陷分类的精度。 |
| 英文摘要: |
| The classification of wafer defect patterns plays a crucial role in the wafer manufacturing process. Accurate identification of wafer defects enables the determination of the root causes of defects, thereby pinpointing issues in the production process.However, existing deep learning-based wafer defect classification methods are designed solely from the spatial or frequency domain, failing to achieve mutual supplementation and integration of spatial and frequency information. This limitation constrains the improvement of wafer defect classification accuracy. To address this issue, a dual-stream wafer defect classification network based on the fusion of spatial and frequency domain features, named SFWD-Net, is proposed.The network utilizes the proposed multi-scale feature extraction convolution module and multi-view attention module to form the spatial stream branch, which extracts spatial information from wafer images. The frequency stream branch, utilizing discrete wavelet transform, extracts frequency information from wafer images. After integrating spatial and frequency information, defect classification is performed. Experiments on the large-scale semiconductor wafer image dataset WM-811K demonstrate that SFWD-Net, by simultaneously designing the network from both spatial and frequency domains, achieves a classification accuracy of 99.299 2%, outperforming five other state-of-the-art methods and significantly improving the accuracy of wafer defect classification. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
