HCDG: unified multiclass unsupervised anomaly detection with adaptive weighted combination and error-aware conditional denoising

Abstract

In industrial visual inspection, unsupervised anomaly detection has significant application value due to the elimination of anomaly labeling requirements. However, existing methods often rely on independent modeling by category, leading to high storage and maintenance costs; unified multi-category modeling is susceptible to the diversity of normal patterns, resulting in approximate identity mappings and weakening anomaly representation capabilities. To address these issues, we propose a hierarchically conditioned denoising and guidance framework (HCDG), which combines adaptive hierarchical feature fusion with error-aware conditional denoising. HCDG integrates shallow texture and deep semantic features and uses noise prediction errors to guide adaptive denoising in the feature bottleneck. A feature-guided decoder reconstructs normal features, and reconstruction and noise prediction errors are jointly used for image-level and pixel-level anomaly scoring. HCDG achieves competitive overall performance on MVTec AD, reaching 99.7% I-AUROC, 99.8% I-AP, and 99.4% I-F1-max at the image level. At the pixel level, HCDG attains 98.4% P-AUROC, 70.2% P-AP, 69.9% P-F1-max, and 95.0% P-AUPRO. These results suggest that the proposed denoising and guidance strategy not only preserves image-level discrimination but also yields more stable pixel-level localization under unified multi-class training.