Controllable reconstruction for diffusion-based anomaly detection

Diffusion-based deep learning has shown promising performance in industrial visual anomaly detection, but diffusion reconstruction may over-reconstruct anomalous regions and provide insufficiently discriminative cues, particularly in complex and highly repetitive structures. To address this issue, this paper develops a controllable diffusion reconstruction framework that regulates the denoising process and exploits complementary anomaly cues. A dynamic denoising guidance strategy progressively relaxes structural constraints along the denoising trajectory, while a dual-path anomaly modeling strategy combines reconstruction discrepancy and structural consistency cues for anomaly detection. A Braided Catheter Sheath Anomaly Detection dataset was constructed from an industrial production environment to evaluate anomaly detection in highly repetitive and fine-grained braided structures. Experiments on public benchmarks and the proposed dataset demonstrate competitive performance on public benchmarks and the best performance across all five reported evaluation metrics on the proposed dataset. Specifically, the proposed framework achieves an image-level area under the receiver operating characteristic curve of 99.42%, a pixel-level area under the receiver operating characteristic curve of 96.47%, and an area under the per-region overlap curve of 97.94%.