Guided mask-enhanced diffusion framework for multi-class anomaly detection

Document Type

Article

Publication Date

4-1-2026

Abstract

Industrial anomaly detection plays a critical role in ensuring product quality and production safety, particularly in automated manufacturing environments where early identification of defects is essential. Traditional reconstruction-based methods have struggled to achieve competitive performance in anomaly detection. In this paper, we propose MG-DDAD (Mask-Guided Denoising Diffusion for Anomaly Detection), a novel diffusion-based framework that extends latent diffusion with mask-guided supervision for industrial anomaly detection. By incorporating mask guidance into the denoising process, our method reconstructs normal regions with high fidelity while preventing the restoration of anomalous areas, thereby improving both reconstruction quality and anomaly localization. Anomalies are identified through pixel- and feature-level discrepancies between the original and reconstructed images. Extensive experiments on the MVTec-AD dataset demonstrate that MG-DDAD achieves superior performance over state-of-the-art approaches, attaining 100% AUROC and at 99.5% AUROC at the image level and 98.6% AUROC at the pixel level on the bottle category, while consistently delivering the best results across most other categories. These results highlight the effectiveness of mask-guided diffusion in enhancing robustness, accuracy, and interpretability for next-generation industrial visual inspection.

Publication Title

Journal of Visual Communication and Image Representation

ISSN

1047-3203

DOI

10.1016/j.jvcir.2026.104741

Funders

Youth Science Fund Project

Volume

117

Publisher

Academic Press Inc.

Share

COinS