Unsupervised industrial anomaly detection using paired well-lit and low-light images

Abstract Unsupervised industrial anomaly detection trains models solely on anomaly-free images to detect unseen defects. While embedding-based methods have recently achieved state-of-the-art results, their use of memory banks substantially increases memory usage and inference times, limiting their practicality in industrial settings. In this work, we propose a lightweight and efficient framework for anomaly detection and localization using paired well-lit and low-light images. Our network learns to reconstruct well-lit features from low-light features on nominal (anomaly-free) samples, detecting anomalies by identifying inconsistencies between the reconstructed and extracted features. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches across multiple industrial datasets. Specifically, our model achieves an Image-level Area Under the Receiver Operating Characteristic (I-AUROC) of 0.854 and rea Under the Per-Region Overlap (AUPRO) of 0.823 on low-light industrial anomaly detection (LL-IAD), significantly surpassing existing methods. Furthermore, it attains I-AUROC scores of 0.864 and 0.858 on the Insulator and Clutch datasets, respectively, outperforming all prior approaches in these industrial settings. Notably, even when well-lit images are unavailable, our model maintains high performance using Retinexformer-enhanced low-light images, demonstrating its adaptability to real-world low-light scenarios. Additionally, we introduce a new industrial anomaly detection dataset featuring paired well-lit and low-light images. To our knowledge, this is the first dataset for LL-IAD dataset.