Pseudo-Label Supervision in Unsupervised Industrial Anomaly Detection

Xin Chen, Yan Zhang, Shengchuan Zhang, Xiawu Zheng, Yunhang Shen, Ke Li, Sicheng Zhao, Liujuan Cao, Rongrong Ji

Keywords: Industrial anomaly Detection, Anomaly Detection, Unsupervised Industrial Anomaly Detection

Abstract: The method based on reconstruction and discrimination has made significant progress in unsupervised industrial anomaly detection (IAD) by using generative models to accurately reconstruct normal regions while exhibiting reconstruction failures in anomalous areas. However, current methodologies present two primary limitations. First, reliance on synthetic anomalies and reconstruction loss metrics introduces inadequate supervisory guidance for targeted model optimization. Second, uniform optimization strategies applied indiscriminately across all image regions neglect spatial discrepancies in model confidence levels. We propose Pseudo-Label Supervision in Unsupervised Industrial Anomaly Detection (PLSAD), a novel framework integrating unsupervised learning with pseudo-label supervision. Our methodology focuses on the differences between the original images and the synthetic anomaly images, thereby decoupling reconstruction processes from discriminative feature learning. This dual-stream architecture not only enhances feature representation robustness but also mitigates error propagation through explicit separation of learning objectives. Furthermore, we introduce Adaptive Intersection-over-Union Weighting (AIW), which dynamically evaluates the model's local performance through pseudo-label and synthetic ground truth alignment, and automatically emphasizes challenging regions. Comprehensive experiments on three IAD benchmarks (MVTec-AD, MVTec-LOCO, VisA) confirm PLSAD's competitive performance in both detection accuracy and anomaly localization.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5844