Summary
International Symposium on Nonlinear Theory and Its Applications
2022
Session Number:B2L-B
Session:
Number:B2L-B-01
Application of Denoising Image Restoration to Anomaly Detection
Yu Kashihara , Takashi Matsubara,
pp.260-263
Publication Date:12/12/2022
Online ISSN:2188-5079
Summary:
Generative models learn complicated distributions and generate new samples that follow the learned distributions. Approximating the input image with the learned model is called reconstruction. However, existing generative models often lose the original features in the reconstructions, such as the original orientation and the flaws with production. The anomaly detection often fails if the reconstruction loses the original features. We propose the anomaly detection model based on the diffusion model to avoid this problem. In this study, the model is evaluated on industrial anomaly detection dataset and demonstrates excellent anomaly detection performance and restoration of anomalous regions.