DiffusionAD

DiffusionAD: Norm-guided One-step Denoising Diffusion for Anomaly Detection

Hui Zhang, Zheng Wang, Zuxuan Wu, Yu-Gang Jiang

Paper link

Introduction

This repo contains source code for DiffusionAD implemented with PyTorch.

DiffusionAD is a novel framework for anomaly detection and localization, which consists of a reconstruction sub-network and a segmentation sub-network. The reconstruction sub-network is implemented via a diffusion model and is tasked with recovering anomalous images to anomaly-free ones. The segmentation sub-network predicts pixel-level anomaly scores using the input image and its anomaly-free recovery. Remarkably, we adopt a one-step denoising paradigm, which is considerably faster than iterative denoising approaches. Furthermore, the proposed norm-guided paradigm enhances the fidelity of the anomaly-free reconstruction.​

Get Started

Environment

pip install -r requirements.txt

Data

MvTec-AD

Download the dataset from here.

VisA

Download the dataset from here.

Foreground

In the anomaly synthetic strategy, we employ different foreground extraction methods for datasets belonging to different classes. For the object dataset, we use DIS to extract foregrounds. For the textural dataset, we design the foreground as a random part of the entire image.​ ​​The Describable Textures dataset(DTD) is one of the anomaly source image sets and can be downloaded from here.

Images of these foregrounds can be downloaded from Google Drive.

Finally, make sure that these datasets follow the data tree.

MVTec-AD
|-- carpet
|-----|----- thresh
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ ...
|-----|----- train
|-----|--------|------ good
|-- cable
|-----|----- DISthresh
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ ...
|-----|----- train
|-----|--------|------ good

VisA
|-- candle
|-----|----- DISthresh
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ bad
|-----|----- train
|-----|--------|------ good
|-- capsules
|-- ...

Train

Please specify the dataset path(MVTec-AD,VisA), anomaly_source_path(DTD), and output folder in args.json and run:

python train.py

Evaluation

To perform inference with checkpoints, please run:

python eval.py

Citation

@article{zhang2023diffusionad,
  title={DiffusionAD: Norm-guided One-step Denoising Diffusion for Anomaly Detection},
  author={Zhang, Hui and Wang, Zheng and Wu, Zuxuan and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2303.08730},
  year={2023}
}

License

All code within the repo is under MIT license