Abstract
This study establishes a comprehensive benchmark for unsupervised Visual Anomaly Detection (VAD) in semiconductor manufacturing using the large-scale MIIC dataset. By evaluating nine state-of-the-art methods, we demonstrate that while modern feature-based approaches excel at defect localization, reconstruction-based methods remain highly effective for image-level detection in specialized Scanning Electron Microscope (SEM) imagery.
Introduction
In semiconductor manufacturing, even microscopic defects can lead to substantial financial losses and equipment downtime. Traditional defect detection systems often rely on supervised learning, which requires thousands of manually labeled anomalous samples—a resource-intensive requirement in real-world production where defects are rare. Visual Anomaly Detection (VAD) addresses this by learning exclusively from "normal" (defect-free) samples.
However, most VAD research has historically focused on natural images. Semiconductor images captured via Scanning Electron Microscope (SEM) present unique challenges: they are grayscale, high-resolution, and characterized by intricate, repetitive textures. This paper bridges the gap by evaluating how state-of-the-art VAD algorithms perform on these specialized industrial images.
Methodology
The study benchmarks nine VAD methods categorized into two primary families:
- Reconstruction-based methods: Models like GANomaly and IAD+Inpainting attempt to rebuild an image. Anomalies are identified by the reconstruction error when the model encounters features unseen during training.
- Feature Embedding-based methods: These utilize pre-trained neural networks to extract features from image patches. This category includes Teacher-Student distillation (STFPM, RD4AD), Memory Bank approaches (PatchCore, CFA), and Normalizing Flow (FastFlow).
The authors highlight CFA (Coupled-hypersphere-based Feature Adaptation), which uses an auxiliary network to better cluster normal features specifically for the target domain.
The MIIC Dataset
The primary benchmark used is the Microscopic Images of Integrated Circuits (MIIC) dataset. It contains 25,276 grayscale SEM images (25,160 normal, 116 anomalous) at 512 × 512 resolution. MIIC is the first large-scale, publicly available VAD dataset for the semiconductor domain, offering significantly more data and heterogeneity than previous SEM benchmarks like NanoTwice.
Key Results and Insights
The evaluation reveals a notable "reconstruction surprise": the older IAD+Inpainting method outperformed modern feature-based methods in image-level detection, achieving a 99.27% ROC and 91.23% F1-score. This performance is attributed to the abundance of normal data in MIIC, which allows reconstruction models to learn the "normal" manifold with high precision.
While reconstruction is effective for detection, feature-based methods like STFPM and CFA are superior for localization—precisely highlighting the specific pixels that constitute a defect. STFPM achieved the highest localization F1-score (77.30%), followed by CFA (75.66%). This localization capability is critical for industrial applications requiring explainable AI results.
Conclusion and Future Work
The study concludes that while feature-based methods provide superior localization, reconstruction-based approaches remain highly competitive for image-level detection in data-rich environments. A key limitation remains the reliance on ImageNet pre-training; future research should explore self-supervised pre-training directly on SEM data to eliminate the domain gap and further improve detection accuracy.
Summary updated