Center-aware Residual Anomaly Synthesis for Multi-class Industrial Anomaly Detection

TL;DR

CRAS tackles multi-class industrial anomaly detection by unifying category distributions through center-aware residual learning, reducing inter-class interference, and mitigating intra-class overlap with distance-guided anomaly synthesis. The approach combines four modules—MCMC, HPI, DAFS, and CRD—to extract robust normal features, align centers globally and locally, synthesize discriminative anomalies, and perform center-aware discrimination. It achieves state-of-the-art image- and pixel-level AUROC and AP across MVTec AD, VisA, and MPDD, with strong real-world applicability demonstrated on ITDD and robust inference speed. The results indicate practical impact for scalable, accurate, and efficient industrial inspection systems.

Abstract

Anomaly detection plays a vital role in the inspection of industrial images. Most existing methods require separate models for each category, resulting in multiplied deployment costs. This highlights the challenge of developing a unified model for multi-class anomaly detection. However, the significant increase in inter-class interference leads to severe missed detections. Furthermore, the intra-class overlap between normal and abnormal samples, particularly in synthesis-based methods, cannot be ignored and may lead to over-detection. To tackle these issues, we propose a novel Center-aware Residual Anomaly Synthesis (CRAS) method for multi-class anomaly detection. CRAS leverages center-aware residual learning to couple samples from different categories into a unified center, mitigating the effects of inter-class interference. To further reduce intra-class overlap, CRAS introduces distance-guided anomaly synthesis that adaptively adjusts noise variance based on normal data distribution. Experimental results on diverse datasets and real-world industrial applications demonstrate the superior detection accuracy and competitive inference speed of CRAS. The source code and the newly constructed dataset are publicly available at https://github.com/cqylunlun/CRAS.

Figures (9)

• Figure 1: Conceptual illustration of Industrial Anomaly Detection (IAD) settings and our motivation. (a) The single-class setting trains separate models for each category independently. (b) The multi-class setting trains a unified model for all known categories. (c) Our method aims to address inter-class interference and intra-class overlap in the multi-class setting through center-aware residual learning and distance-guided anomaly synthesis.
• Figure 2: Schematic of the proposed CRAS. (a) Multi-class Contextual Memory Center (MCMC) extracts features and initializes the multi-class centers. (b) Hierarchical Pattern Integration (HPI) aligns normal features with the contextual centers. (c) Distance-guided Anomaly Feature Synthesis (DAFS) synthesizes anomaly features based on Gaussian noise. (d) Center-aware Residual Discrimination (CRD) enhances discriminative ability through residual learning. The training stage is depicted with solid and dashed arrows, while the inference stage is indicated by solid arrows.
• Figure 3: Overview of the DAFS module. (a) Pipeline of anomaly synthesis based on recomposed center and Gaussian noise. (b) Anomaly feature distribution synthesized by the distance-guided mechanism. (c) Residual feature distribution after center-aware residual learning.
• Figure 4: t-SNE visualization of adapted patch-level feature distributions. (a) Raw features show significant overlap between normal and abnormal samples. (b) Center-aware residual features generated by CRD form more compact normal clusters and better separation from anomalies.
• Figure 5: Performance comparison of various methods on MPDD under the single-class setting, as measured by I-AUROC% and P-AUROC%.