SimpleNet: A Simple Network for Image Anomaly Detection and Localization

Zhikang Liu, Yiming Zhou, Yuansheng Xu, Zilei Wang*

Paper link

Introduction

This repo contains source code for SimpleNet implemented with pytorch.

SimpleNet is a simple defect detection and localization network that built with a feature encoder, feature generator and defect discriminator. It is designed conceptionally simple without complex network deisng, training schemes or external data source.

Get Started

Environment

Python3.8

Packages:

• torch==1.12.1
• torchvision==0.13.1
• numpy==1.22.4
• opencv-python==4.5.1

(Above environment setups are not the minimum requiremetns, other versions might work too.)

Data

Edit run.sh to edit dataset class and dataset path.

Custom dataset without ground-truth masks

SimpleNet now supports training and evaluation on datasets that only provide images. Use the following directory layout:

data/custom/train/good/*.jpg
data/custom/test/good/*.jpg
data/custom/test/bad/*.jpg

Notes:

• The ground_truth directory is optional; segmentation masks are not required.
• Good images in the test split will automatically receive a zero mask during evaluation, so alignment errors will not occur even without GT files.
• Because pixel-level metrics depend on masks, they are disabled when no masks are present or all provided masks are empty. Instance-level AUROC continues to be reported normally.

Update run.sh to point to your custom data folder and classname. Replace the default MVTec setup by setting datapath to the parent folder that contains your class subfolders (for example, data when using data/custom) and listing each class folder inside the datasets array in run.sh (for instance, datasets=('custom') for a single class). The script builds the appropriate flags automatically and runs SimpleNet with the custom layout.

MvTecAD

Download the dataset from here.

The dataset folders/files follow its original structure.

Run

Demo train

Set datapath and datasets in run.sh before running so they match your custom dataset layout. The script now defaults to the custom directory structure without ground-truth masks (with datapath=data and datasets=('custom') for data/custom/...) and uses resize/imagesize=128 to match a 128x128 custom dataset; switch the path, class list, or image sizes if you want to run on MVTecAD instead.

Inference only (using a trained checkpoint)

Use the infer subcommand to load a saved checkpoint and evaluate the test split without running training. Point --checkpoint_dir to the folder that contains models.ckpt (created during training), and pass the same net/dataset settings you used for training so the model and dataloader are rebuilt correctly. For the custom dataset setup described above, you can keep the configuration in inference.sh (update datapath and checkpoint_dir if needed) and simply run:

The script rebuilds the dataloader with the correct options and reports AUROC/PRO metrics along with FPS for the restored model. Optionally save segmentation images by adding --save_segmentation_images inside the script command.

Citation

@inproceedings{liu2023simplenet,
  title={SimpleNet: A Simple Network for Image Anomaly Detection and Localization},
  author={Liu, Zhikang and Zhou, Yiming and Xu, Yuansheng and Wang, Zilei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20402--20411},
  year={2023}
}

Acknowledgement

Thanks for great inspiration from PatchCore

License

All code within the repo is under MIT license