Defect classifications are the very important steps as the in-line defect inspection of the semiconductor manufacturing procedure. The conventional defect classifications are usually through visual judgement by engineer or technical assistant. However, it's time-consuming and laborious. In our recent study, the artificial intelligence automatic defect classification (AI-ADC) performed promisingly good accuracy and purity (A/P) of the auto defect classification by deep learning method. Nevertheless, some kinds of tiny defects are not only suffered lower A/P issues, but also suffered bad A/P stability of real defect classification. In this work, we propose the novel method, called “Hierarchical structure AI-ADC”, which introduced a second binning classifier and it's based on hierarchical clustering to achieve more precise defect classification. As a result, the proposed method shown obvious improvements to the binning purity of multi-lines bridge defect from 56% to 88% as well as the stability variation has been reduced from 55% to 22%, besides it also can be applied to classify the similar defect types efficiently. Indeed this approach achieves excellent defect classification and highly stable performance.

中文翻译：

---

人工智能自动缺陷分类中分层架构对多线桥梁缺陷分类的改进


缺陷分类是半导体制造过程中在线缺陷检测中非常重要的步骤。传统的缺陷分类通常是由工程师或技术助理通过目视判断。然而，这既费时又费力。在我们最近的研究中，人工智能自动缺陷分类（AI-ADC）在深度学习方法的自动缺陷分类中表现出了良好的准确性和纯度（A/P）。然而，某些微小缺陷不仅存在较低的A/P问题，而且实际缺陷分类的A/P稳定性也较差。在这项工作中，我们提出了一种称为“分层结构AI-ADC”的新方法，该方法引入了第二个分箱分类器，并基于分层聚类来实现更精确的缺陷分类。结果表明，该方法将多线桥缺陷的分箱纯度从 56% 明显提高到 88%，稳定性变异从 55% 降低到 22%，此外它还可以应用于分类有效地检测相似的缺陷类型。事实上，这种方法实现了出色的缺陷分类和高度稳定的性能。