Evaluating Modern Visual Anomaly Detection Approaches in Semiconductor Manufacturing: A Comparative Study

TL;DR

This paper tackles the problem of visual anomaly detection in semiconductor SEM images, where anomalous labeled data are scarce. It introduces MIIC as a large-scale SEM-specific VAD benchmark and evaluates both reconstruction-based and modern feature-embedding methods across image-level and pixel-level metrics. The results show that feature-based methods like CFA perform exceptionally well on MIIC, while surprisingly strong reconstruction-based approaches (e.g., inpainting) can approach or match ROC performance, highlighting domain effects and the value of large normal datasets. The findings emphasize the importance of domain-aware benchmarking and suggest future work on domain-tuned features and self-supervised fine-tuning to further improve SEM defect localization and explainability.

Abstract

Semiconductor manufacturing is a complex, multistage process. Automated visual inspection of Scanning Electron Microscope (SEM) images is indispensable for minimizing equipment downtime and containing costs. Most previous research considers supervised approaches, assuming a sufficient number of anomalously labeled samples. On the contrary, Visual Anomaly Detection (VAD), an emerging research domain, focuses on unsupervised learning, avoiding the costly defect collection phase while providing explanations of the predictions. We introduce a benchmark for VAD in the semiconductor domain by leveraging the MIIC dataset. Our results demonstrate the efficacy of modern VAD approaches in this field.

Figures (2)

• Figure 1: Representative images from the MIIC dataset. In the first row are reported normal images, in the second row anomalous images with the anomalies highlighted in red, in the last row the segmentation masks produced by CFA.
• Figure 2: Comparison of F1 scores by VAD method on image level(left) and on pixel level(right).