基于多头注意力和对抗网络的晶圆图缺陷生成方法

Wafer Map Defect Generation Model Based on Multi-Head Attention and Generative Adversarial Networks

摘要: 为了解决深度学习分类模型对少数样本的晶圆图缺陷模式的分类准确率低的问题。该研究提出了一种融合多头注意力机制和对抗网络模型的生成算法Multi-SAGAN。多头注意力机制拥有多个特征子空间，可以生成更丰富的图像细节和全局特征，利用Multi-SAGAN生成的晶圆图来扩充数量较少的晶圆图缺陷模式，能够提高分类器的分类准确率。为了比较DCGAN、SAGAN、数据增强模型和Multi-SAGAN的生成性能，分别把生成的图像和原始数据集组合成新的数据集，放入同一个分类模型中比较分类准确率。最终实验结果表明由Multi-SAGAN生成的图像组成的数据集准确率比原始数据集准确率高18.9%，比数据增强和DCGAN的扩充数据集准确率分别高7.4%和6.4%。比SAGAN的扩充数据集准确率高2.2%。

Abstract: To address the issue of low classification accuracy of deep learning models on wafer map defect patterns with scarce quantities, this study proposes a novel generative algorithm named Multi-SAGAN, integrating multi-head-attention and SAGAN. Generating multiple feature subspaces through multi-attention enables the creation of richer image details and global features. Ultimately, utilizing this generative algorithm to augment the limited quantity of wafer map defect patterns aims to enhance the classifier’s classification accuracy. The images generated by DCGAN, SAGAN, data augmentation, and Multi-SAGAN are respectively mixed with the original dataset to create new datasets, which are then input into the same classification model to compare classification accuracy. The final experimental results show that the dataset generated by Multi-SAGAN achieves an accuracy 18.9% higher than the original dataset, surpassing the accuracy of the datasets generated by data augmentation and DCGAN by 7.4% and 6.4% respectively. Additionally, it outperforms the accuracy of the dataset generated by SAGAN by 2.2%.