How Many Normal Images Are Needed? Training-Free Few-Shot Anomaly Detection on Scanning Electron Microscope Images of Integrated Circuits

Abstract: Anomaly detection on scanning electron microscope (SEM) images of integrated circuits is an essential element in the semiconductor manufacturing quality control process to ensure that no defective chip is delivered. Because SEM data is specific to each product and imaging tool, inspection models are usually built or tuned for each production line. Developing these models involves a large quantity of images and takes considerable effort to collect the data and refine the model until it works reliably. Also, most of the time, collecting data results in a disproportionate amount of normal images compared to anomalous images. Vision foundation models, pretrained on very large image collections, can describe images in great detail and have become strong tools for finding anomalies in photographs. Even so, the best anomaly detection models on the public MIIC dataset have been trained on 23,888 normal images, which is still a large number. In this research, we studied training-free few-shot anomaly detection on MIIC using frozen vision foundation backbones to assess whether their features transfer to SEM images. We extracted features from k normal reference images to form a memory bank, and assigned an anomaly score to each test feature based on its distance to the nearest stored reference feature in the memory bank. We tested six backbones with four values of k and three random seeds. DINOv2- B with 64 references yielded 94.53% image AUROC, which was better than full-data PatchCore (93.33%), while DINOv3 and CLIP were less transferable. A layer analysis revealed that averaging the last four transformer layers played a critical role in localization, raising the AUPRO of DINOv3-S to 95.2% compared with 76.2% for the last layer alone. When the 16 most similar normal references were retrieved for every test image, AUPRO reached 99.23%, nearly equivalent to full-data CFA (99.04%), and the nearest-neighbour search was 129 times quicker than searching through a 2,000-image pool. However, image-level AUROC (94.78%) lagged behind full-data CFA (99.01%). A decision layer calibrated solely on normal images rejected only 0.39% of normal images and detected every anomaly covering at least 0.25% of the image.

Keywords: Few-shot anomaly detection, foundation models, defect inspection, semiconductor manufacturing, electron microscopy.