基于重采样扩散模型的异常检测  

Anomaly detection based on resampling diffusion model

作　　者：白宇 张若淇 周羿旭 朱强 BAI Yu;ZHANG Ruo-qi;ZHOU Yi-xu;ZHU Qiang(School of Computer Science,Beijing Information Science and Technology University,Beijing 100101,China;College of Computer Science and Technology,Zhejiang University,Hangzhou 310013,China)

机构地区：[1]北京信息科技大学计算机学院,北京100101 [2]浙江大学计算机科学与技术学院,浙江杭州310013

出　　处：《计算机工程与设计》2025年第3期788-794,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2022YFF1202400)。

摘　　要：为提升基于图像重建的异常检测方法的检测和定位准确率,提出一种基于重采样去噪扩散概率模型的异常检测方法,名为DiffAD。仅用正常样本训练扩散模型实现无缺陷重建,通过比较输入图像和重建图像像素级特征检测和定位异常区域。在保留原始扩散模型基础上,引入语义融合重采样技术和自适应掩码策略,有效消除图像的像素级重建间隙,显著提升非结构化类型的异常检测准确率。在Mvtec AD数据集上的实验结果表明,该方法的异常检测和定位准确率均优于其它方法。To enhance the detection and localization accuracy of anomaly detection methods based on image reconstruction,an anomaly detection method was proposed based on a resampling denoising diffusion probability model,named DiffAD.The diffusion model with normal samples was solely trained to achieve defect-free reconstruction.Anomalous regions were detected and located by comparing pixel-level features between input and reconstructed images.Building upon the original diffusion model,semantic fusion resampling techniques and adaptive mask strategy were introduced,effectively eliminating pixel-level reconstruction gaps in images.The accuracy of anomaly detection is significantly improved,especially for non-structured categories.Results of experiments on the Mvtec AD dataset demonstrate that the proposed method outperforms other methods in both anomaly detection and localization accuracy.

关 键 词：机器视觉 异常检测 图像重建 扩散模型 无监督学习 重采样 自适应掩码 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]