智能系统学报2025，Vol.20Issue(2)：376-388,13.DOI:10.11992/tis.202402022
基于ODE扩散模型的多类异常检测和定位
ODE diffusion model for multiclass anomaly detection and localization
摘要
Abstract
Multiclass anomaly detection and localization methods aim to train a single model capable of identifying an-omalous regions that deviate from normal across multiple categories.Diffusion-based methods have recently attracted attention due to their excellent performance in this task.However,existing methods concentrate on enhancing the de-noising network of the diffusion model by adding more constraints to ensure high consistency in multistep generation and achieve superior reconstruction performance.Additional sampling steps also lead to higher computational costs.We propose a novel multiclass anomaly detection and localization method called TimeNet to address these issues.It is based on the diffusion model of ordinary differential equations and achieves high-quality reconstruction with only one-step generation.We introduce a time-perceptive network to address the consistency and identity shortcut problems that may arise from small sampling steps,which further improves the reconstruction quality.Experiments on the most popular benchmark MVTec-AD dataset demonstrate that our TimeNet competes with the current state-of-the-art methods in terms of accuracy while requiring less computational effort and achieving faster speeds.The high accuracy and real-time performance of TimeNet satisfy the requirements for industrial anomaly detection and localization.关键词
缺陷检测/异常检测/异常定位/扩散模型/去噪网络/常微分方程/无监督学习/时间步感知Key words
defect detection/anomaly detection/anomaly localization/diffusion model/denoising network/ordinary dif-ferential equations/unsupervised learning/timestep-perceptive分类
信息技术与安全科学引用本文复制引用
蒋世杰,夏秀山,翟伟,曹洋..基于ODE扩散模型的多类异常检测和定位[J].智能系统学报,2025,20(2):376-388,13.基金项目
安徽省重点研究与开发计划项目(2022107020030). （2022107020030）
