1. DiffusionAD, tpami25: Norm-guided One-step Denoising Diffusion for Anomaly Detection

Introduction

fig 1

Method

如何使用 Diffusion:

一步去噪(One-step Denoising): 核心在于重新审视去噪过程,提出通过扩散模型预测一次噪声后直接预测重构结果,而非迭代去噪。

范数引导(Norm-guided): 同时对图像注入两种不同尺度的噪声(ts=200, tb=400), 解决了单一噪声尺度难以同时处理微小异常和语义级别大异常的问题

image.png

  1. Reconstruction

    $$ \begin{equation}\mathcal{L}=\mathbb{E}{z,\;\epsilon \sim N(0,I),\;t}\left[\left\|\epsilon - \epsilon\theta\!\left(z_t,\, t\right)\right\|_2^2\right]\end{equation} $$

    1. ϵ是forward diffusion 中显式引入的高斯噪声随机变量
  2. NOD

    1. Norm-guided: 利用大噪声尺度的重构结果在语义上的正确性来引导小噪声尺度的重构,平衡细节和语义
      1. tb尺度 ⇒ 预测出一个语义正确但细节较模糊的参考图x^0b
      2. x^0b为引导,结合ts尺度的信息进行一步重构,最终生成既符合正常语义分布又具备高保真细节的图像x^0g
    2. One-step Denoising
  3. Segmentation Sub-network

  4. DiffusionAD [51]

    1. rely on external synthetic anomalies, potentially decreasing robustness to unseen anomalies.
    2. employs two sub-networks for denoising and segmentation, inspired by DRAEM [50], showcasing the success of diffusion models over VAEs in anomaly detection.
    3. While a single denoising step accelerates the process, it makes it akin to VAEs, moving directly from noise to signal, with the distinction that in this case, the starting point is a noise-to-signal ratio.