电子测量与仪器学报

杜先君,贾龙.融合改进多头注意力与残差结构的VGGNet晶圆缺陷检测[J].电子测量与仪器学报,2025,39(8):1-12

融合改进多头注意力与残差结构的VGGNet晶圆缺陷检测

VGGNet wafer defect detection with improved multi-headattention and residual structure

DOI：

中文关键词: 晶圆图像分类卷积神经网络不平衡数据集 VGGNet 注意力机制

英文关键词:wafer image classification convolutional neural network imbalanced dataset VGGNet attention mechanism

基金项目:国家自然科学基金（62241307）、兰州市科技支撑计划（2024-3-47）项目资助

作者	单位
杜先君	1.兰州理工大学微电子现代产业学院兰州730050；2.兰州理工大学自动化与电气工程学院兰州730050
贾龙	兰州理工大学微电子现代产业学院兰州730050

Author	Institution
Du Xianjun	1.School of Microelectronics Industry-education Integration, Lanzhou University of Technology, Lanzhou 730050, China; 2.School of Automation and Electrical Engineering, Lanzhou University of Technology, Lanzhou 730050, China
Jia Long	School of Microelectronics Industry-education Integration, Lanzhou University of Technology, Lanzhou 730050, China

摘要点击次数: 1040

全文下载次数: 798

中文摘要:

精准检测晶圆图像中的缺陷对于及时识别晶圆生产过程中的异常故障具有重要意义。在晶圆测试阶段，由于深度学习方法具备卓越的特征提取能力，其在晶圆缺陷检测中得到广泛应用。然而，传统深度学习模型通常依赖于大量标注充分且高质量的数据进行训练，而在实际应用中，均衡、充足的标注数据往往难以获得。针对这一问题,提出了一种融合改进多头注意力机制与残差结构的VGGNet深度学习模型，旨在从不平衡的数据集中提取更全面的特征，从而实现对晶圆表面缺陷的精准检测。具体而言,利用改进的多头注意力机制将输入的晶圆图像特征映射到多维子空间,显著提升了模型的表达能力和泛化性能；同时，在传统VGGNet的全连接层中引入残差连接(residual structure, RS)，有效缓解了深层网络训练中的梯度消失问题。为验证融合改进多头注意力机制与残差结构的VGGNet的有效性，在数据集WM811K上进行大量实验，其分类准确率达到94.3%，相较传统VGGNet准确率提高了3%，相较现有类似模型准确率平均提高了1%。实验结果表明，在真实数据集WM811K上，所提方法不仅提高了晶圆缺陷检测的鲁棒性，而且在非平衡数据集上的检测性能明显优于现有算法。

英文摘要:

Accurate detection of defects in wafer images is of great significance for timely identification of abnormal faults in wafer production. In the wafer testing phase, the deep learning method has been widely used in wafer defect detection due to its excellent feature extraction capability. However, traditional deep learning models often rely on a large number of adequately labeled and high-quality data for training, and in practical applications, balanced and sufficient labeled data is often difficult to obtain. To address this issue, we propose a VGGNet deep learning model that integrates an improved multi-head attention mechanism with a residual structure, aiming to extract more comprehensive features from imbalanced data sets to achieve accurate detection of wafer surface defects. Specifically, we use an improved multi-head attention mechanism to map the input wafer image features to multi-dimensional subspaces, which significantly improves the expressiveness and generalization performance of the model. At the same time, the residual connection is introduced into the full connection layer of traditional VGGNet, which effectively alleviates the problem of gradient disappearance in deep network training. To validate the effectiveness of the VGGNet with the improved multi-head attention mechanism and residual structure(RS), extensive experiments were conducted on the WM811K dataset, achieving a classification accuracy of 94.3%, which is 3% higher than the traditional VGGNet and 1% higher than existing similar models on average. The experimental results show that on the real data set WM811K, the proposed method not only improves the robustness of wafer defect detection, but also significantly outperforms the existing algorithms on the non-equilibrium data set.

查看全文查看/发表评论下载PDF阅读器