Prior knowledge-based denoising diffusion for semi-supervised anomaly detection

In industrial quality inspection, defect samples are rare and diverse, often following a long-tail distribution. This scarcity limits the reliability of supervised detection methods when facing previously unseen defects. Semi-supervised approaches, such as the Denoising Diffusion Anomaly Detection (DDAD), address this challenge by training solely on normal samples within a stable diffusion framework, enabling detection of unknown anomalies. However, DDAD remains limited in reconstruction accuracy and its ability to integrate prior knowledge. We introduce IPAD, a semi-supervised diffusion-based framework that augments training on normal samples with a small set of known defect types, used as priors to guide learning and improve sensitivity to anomalous regions. A dynamic noise adjustment mechanism further enhances reconstruction robustness, while a region-adaptive dual loss function applies differentiated constraints to defective and non-defective regions through learnable weighting. Experiments on the MVTec dataset show that IPAD achieves an average PRO score of 93.5%, surpassing DDAD by 4.7%. These results demonstrate that incorporating limited prior knowledge effectively overcomes the bottlenecks of unsupervised methods, offering a robust and generalizable framework for unknown defect detection in industrial inspection.