UFSMatAD: A Unified Framework for Few-Shot Material Anomaly Detection Across Nanofiber SEM and Wafer Imaging

Automated detection of nanoscale defects in materials imagery is challenging due to scarce labels, high morphological variability, and strict latency requirements in inline inspection. We present UFSMatAD, a unified, parameter-efficient framework for few-shot, multi-class anomaly detection in SEM and wafer AOI images. UFSMatAD replaces decoder feed-forward networks with Adapter Blocks configured with two bottleneck sizes and uses deterministic routing at inference to ensure stable optimization and predictable latency. A lightweight reconstruction head produces pixel-level maps and image-level scores. With the backbone frozen and only the decoder and head trainable, UFSMatAD matches transformer and diffusion baselines while substantially reducing trainable parameters and computational cost, and it remains robust under SEM-to-AOI domain shift. These results indicate that deterministic adapter mixtures provide a practical, scalable path to generalizable and resource-efficient industrial inspection.