Introduction

1. $I'=f_\theta(I)$, then anomaly map: $A=C_\theta(I, I')$

   1. Testing: 异常图像⇒正常图像. how to achieve this?

2. trained on nominal and / or abnormal samples, learns to accurately reconstruct nominal data while failing to reconstruct anomalies.

   

   1. the majority of reconstruction-based methods are trained from scratch without employing robust pretrained models, which results in inferior performance ****compared to feature embedding ?

3. memory bank based methods和库里的比；reconstruction based和自己比。

Methods

VAE ⇒ blurry and anomalies weren’t adequately removed;

GAN ⇒ rely on synthetic anomalies for training.

Diffusion based methods are better than VAE and GAN for AD.

Diffusion based

102 Diffusion models

看这些文章的fig1，都是挖掘diffusion的生成能力。

1. 先看DiffAD, iccv23 的效果示意图，？？？？？；
2. 再看DDAD, pr24，因其方法更简单；
   1. not only outperforms reconstruction-based models but also representation-based models.
   2. Ranked #1 on MVtec AD (Detection AUROC metric)
   3. Ranked #1 on VisA (Detection AUROC metric)
3. DiffusionAD, tpami25
   1. Ranked #3 on VisA (Detection AUROC metric)
   2. has code

对比

MVTec AD — Image-level AUROC（整体异常判断）

MVTec AD — Pixel-level AUROC（定位异常）