Prototypical Residual Networks for Anomaly Detection and Localization

Unofficial PyTorch implementation for CVPR2023 paper, Prototypical Residual Networks for Anomaly Detection and Localization.

This paper proposes a framework called Prototypical Residual Network (PRN), which learns feature residuals of varying scales and sizes between anomalous and normal patterns to accurately reconstruct the segmentation maps of anomalous regions.

Dataset Preparation

Download MVTecAD dataset from there, put it to the directory data/mvtec_anomaly_detection.

Then unzip the foreground masks:

unzip fg_mask.zip

Prototype Features Generation

Run the create_proto_feature_maps.py to generate prototype features.

python create_proto_feature_maps.py

Training and validating

Run the following code for training and validating the MVTecAD dataset.

python train.py

We summarize the validation results on MVTecAD as follows.

Category	Image/Pixel AUC	Paper
Carpet	0.999/0.988	0.977/0.990
Grid	0.818/0.830	0.994/0.984
Leather	1.000/0.994	1.000/0.997
Tile	0.992/0.957	1.000/0.996
Wood	1.000/0.955	1.000/0.978
Bottle	1.000/0.984	1.000/0.994
Cable	0.952/0.949	0.989/0.988
Capsule	0.903/0.961	0.980/0.985
Hazelnut	1.000/0.994	1.000/0.997
Metal Nut	0.962/0.984	1.000/0.997
Pill	0.880/0.970	0.993/0.995
Screw	0.834/0.961	0.959/0.975
Toothbrush	0.983/0.963	1.000/0.996
Transistor	0.931/0.973	0.997/0.984
Zipper	0.988/0.960	0.997/0.988
Mean	0.949/0.962	0.994/0.990

🔨 Todo List

The following issues need to be further improved:

Update the dataset code. Current data processing way is not fully consistent with the description in the paper.
Address the problem of unstable results during training. The loss is stable but the validation results are unstable.

Reference

@InProceedings{Zhang_2023_CVPR,
    author    = {Zhang, Hui and Wu, Zuxuan and Wang, Zheng and Chen, Zhineng and Jiang, Yu-Gang},
    title     = {Prototypical Residual Networks for Anomaly Detection and Localization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {16281-16291}
}

Acknowledgement

The dataset part of this repository is built using the BGAD library and the DRAEM repository.

xcyao00 / PRNet