Anomaly Detection with Conditioned Denoising Diffusion Models

Presenter: Arian Mousakhan, PhD student at the Computer Vision Lab, University of Freiburg

Abstract: Traditional reconstruction-based visual anomaly detection methods struggle to achieve competitive performance. This is primarily because most existing approaches are unable to precisely reconstruct anomalous inputs resulting in restored pattern that diverge from the original image. Moreover, methods often fail to conduct a robust comparison between the reconstructed and input images. In this paper, we propose Denoising Diffusion Anomaly Detection (DDAD) whereby a generic diffusion model is first trained only on nominal data. During inference, the reverse process is conditioned on the unperturbed input image by correcting the predicted noise at each denoising step. This novel mechanism accurately reconstructs anomalous regions while preserving the in-distribution patterns of the image.  Our validate demonstrates the efficacy of DDAD on various datasets, achieving state-of-the-art results

In compliance with GDPR consent requirements, presentations given in a Visual Intelligence context may be recorded with the consent of the speaker. All recordings are edited to remove all faces, names and voices of other participants. Questions and comments by the audience will hence be removed and will not appear in the recording.  With the freely given consent from the speaker, recorded presentation may be posted on the Visual Intelligence YouTube channel.