Introduction

DDAD, pr24: anomaly detection with conditioned denoising diffusion models

3rd row of fig 1: Input, GT, reconstruction (like the target image in training process), heatmap

see more examples in Figure 1.

Method

1. Training a denoising U-Net: image level, conditioned diffusion
   1. reconstruct / denoise X to X0, conditioned on Y. wish that
      1. input X is nominal.
      2. target Y is noimal and has different pose.
      3. X0 is a defect-free and similar to Y.
   2. Domain Adaptation
      1. distill the pretrained feature extraction model ϕ (classification on ImageNet) to the domain-adapted network ϕ^, satisfying that
         1. distill: ϕ^(x0) and ϕ(x0) should be similar; ϕ^(y) and ϕ(y) should be similar;
         2. adaption: ϕ^(x0) and ϕ^(y) should be similar;
      2. note that ϕ and ϕ^ have the same structure;
      3. ϕ^: making the model robust to nominal variation of the object and spurious anomalies in the background present in the reconstruction.
         1. why? see In Appendix, Section 10.3.
2. Comparison
   1. feature-wise + pixel-wise, see fig 3, 10
      1. feature-wise for poked parts or dents 戳伤的部分或凹痕, whereby visible colour variations are not present.
      2. pixel-wise for defects with different colors.

Related work

1. AnoDDPM [46] substituting Gaussian noise with simplex noise results in a slower inference time. Generally, the time complexity of sampling simplex noise, which is O(n2), is typically higher than that of Gaussian noise, which is O(1).
   1. this paper: Gaussian noise
2. Score-based perturbation resilience [40] 性能不如本文

Experiments

比PatchCore快一点。

Future work

1. dynamically selecting the denoising starting points

   1. the initial denoising point T ′ = 250, in training?

      which presents a greater challenge to reconstruct large missing components (such as some samples in the transistor category).

   2. However, starting from earlier time steps, T>T’ , introduces ambiguities in the reconstruction and leads to increased inference time. Some failure modes of the model are presented in the Appendix, Section 13.1, Fig 11 (这是GT不准吧？看看正常数据啥样)

   3. ⇒ should be dynamic for different category & anomaly types?

References

1. DDAD, pr24: anomaly detection with conditioned denoising diffusion models
   1. has code