Breaking the Bias: Quantifying the Attention of Industrial Anomaly Detection

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

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

Abstract: Industrial anomaly detection (IAD) predominantly utilizes unsupervised learning due to the scarcity and unpredictability of defect samples. A major challenge in unsupervised IAD methods is the inherent bias in normal samples, which causes models to focus on variable regions while overlooking potential defects in invariant areas. In this paper, we propose Recalibrating Attention of Industrial Anomaly Detection (RAAD), which decomposes and recalibrates the input data to highlight anomalies better. Additionally, Hierarchical Quantization Scoring (HQS) is introduced to refine the detection process by assigning quantization scores at multiple levels. These strategies work together to mitigate the bias toward normal samples and improve the accuracy of anomaly detection. We validate the effectiveness of RAAD on three IAD datasets: MVTec-AD, MVTec-LOCO, and VisA. The experimental results demonstrate that RAAD exhibits competitiveness in both detection and localization tasks, providing a robust solution for industrial anomaly detection. The source code will be released to promote further research and application.

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

Submission Number: 5846