Registration based Few-Shot Anomaly Detection

This is an official implementation of “Registration based Few-Shot Anomaly Detection” (RegAD) with PyTorch, accepted by ECCV 2022 (Oral).

Paper Link

@inproceedings{huang2022regad,
  title={Registration based Few-Shot Anomaly Detection}
  author={Huang, Chaoqin and Guan, Haoyan and Jiang, Aofan and Zhang, Ya and Spratlin, Michael and Wang, Yanfeng},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

Abstract: This paper considers few-shot anomaly detection (FSAD), a practical yet under-studied setting for anomaly detection (AD), where only a limited number of normal images are provided for each category at training. So far, existing FSAD studies follow the one-model-per-category learning paradigm used for standard AD, and the inter-category commonality has not been explored. Inspired by how humans detect anomalies, i.e., comparing an image in question to normal images, we here leverage registration, an image alignment task that is inherently generalizable across categories, as the proxy task, to train a category-agnostic anomaly detection model. During testing, the anomalies are identified by comparing the registered features of the test image and its corresponding support (normal) images. As far as we know, this is the first FSAD method that trains a single generalizable model and requires no re-training or parameter fine-tuning for new categories.

Keywords: Anomaly Detection, Few-Shot Learning, Registration

Get Started

Environment

python >= 3.7.11
pytorch >= 1.11.0
torchvision >= 0.12.0
numpy >= 1.19.5
scipy >= 1.7.3
skimage >= 0.19.2
matplotlib >= 3.5.2
kornia >= 0.6.5
tqdm

Files Preparation

Download the MVTec dataset here.
Download the support dataset for few-shot anomaly detection on Google Drive or Baidu Disk (i9rx) and unzip the dataset. For those who have problem downloading the support set, please optional download categories of capsule and grid on Baidu Disk (pll9) and Baidu Disk (ns0n).
```
tar -xvf support_set.tar
```
We hope the followers could use these support datasets to make a fair comparison between different methods.
Download the pre-train models on Google Drive or Baidu Disk (4qyo) and unzip the checkpoint files.
```
tar -xvf save_checkpoints.tar
```

After the preparation work, the whole project should have the following structure:

./RegAD
├── README.md
├── train.py                                  # training code
├── test.py                                   # testing code
├── MVTec                                     # MVTec dataset files
│   ├── bottle
│   ├── cable
│   ├── ...                  
│   └── zippper
├── support_set                               # MVTec support dataset files
│   ├── 2
│   ├── 4                 
│   └── 8
├── models                                    # models and backbones
│   ├── stn.py  
│   └── siamese.py
├── losses                                    # losses
│   └── norm_loss.py  
├── datasets                                  # dataset                      
│   └── mvtec.py
├── save_checkpoints                          # model checkpoint files                  
└── utils                                     # utils
    ├── utils.py
    └── funcs.py

Quick Start

python test.py --obj $target-object --shot $few-shot-number --stn_mode rotation_scale

For example, if run on the category bottle with k=2:

python test.py --obj bottle --shot 2 --stn_mode rotation_scale

Training

python train.py --obj $target-object --shot $few-shot-number --data_type mvtec --data_path ./MVTec/ --epochs 50 --batch_size 32 --lr 0.0001 --momentum 0.9 --inferences 10 --stn_mode rotation_scale

For example, to train a RegAD model on the MVTec dataset on bottle with k=2, simply run:

python train.py --obj bottle --shot 2 --data_type mvtec --data_path ./MVTec/ --epochs 50 --batch_size 32 --lr 0.0001 --momentum 0.9 --inferences 10 --stn_mode rotation_scale

Then you can run the evaluation using:

python test.py --obj bottle --shot 2 --stn_mode rotation_scale

Results

Results of few-shot anomaly detection and localization with k=2:

AUC (%)	Detection		Localization
K=2	RegAD	Inplementation	RegAD	Inplementation
bottle	99.4	99.7	98.0	98.6
cable	65.1	69.8	91.7	94.2
capsule	67.5	68.6	97.3	97.6
carpet	96.5	96.7	98.9	98.9
grid	84.0	79.1	77.4	77.5
hazelnut	96.0	96.3	98.1	98.2
leather	99.4	100	98.0	99.2
metal_nut	91.4	94.2	96.9	98.0
pill	81.3	66.1	93.6	97.0
screw	52.5	53.9	94.4	94.1
tile	94.3	98.9	94.3	95.1
toothbrush	86.6	86.8	98.2	98.2
transistor	86.0	82.2	93.4	93.3
wood	99.2	99.8	93.5	96.5
zipper	86.3	90.9	95.1	98.3
average	85.7	85.5	94.6	95.6

Results of few-shot anomaly detection and localization with k=4:

AUC (%)	Detection		Localization
K=4	RegAD	Inplementation	RegAD	Inplementation
bottle	99.4	99.3	98.4	98.5
cable	76.1	82.9	92.7	95.5
capsule	72.4	77.3	97.6	98.3
carpet	97.9	97.9	98.9	98.9
grid	91.2	87	85.7	85.7
hazelnut	95.8	95.9	98.0	98.4
leather	100	99.9	99.1	99
metal_nut	94.6	94.3	97.8	96.5
pill	80.8	74.0	97.4	97.4
screw	56.6	59.3	95.0	96.0
tile	95.5	98.2	94.9	92.6
toothbrush	90.9	91.1	98.5	98.5
transistor	85.2	85.5	93.8	93.5
wood	98.6	98.9	94.7	96.3
zipper	88.5	95.8	94.0	98.6
average	88.2	89.2	95.8	96.2

Results of few-shot anomaly detection and localization with k=8:

AUC (%)	Detection		Localization
K=8	RegAD	Inplementation	RegAD	Inplementation
bottle	99.8	99.8	97.5	98.5
cable	80.6	81.5	94.9	95.8
capsule	76.3	78.4	98.2	98.4
carpet	98.5	98.6	98.9	98.9
grid	91.5	91.5	88.7	88.7
hazelnut	96.5	97.3	98.5	98.5
leather	100	100	98.9	99.3
metal_nut	98.3	98.6	96.9	98.3
pill	80.6	77.8	97.8	97.7
screw	63.4	65.8	97.1	97.3
tile	97.4	99.6	95.2	96.1
toothbrush	98.5	96.6	98.7	99.0
transistor	93.4	90.3	96.8	95.9
wood	99.4	99.5	94.6	96.5
zipper	94.0	93.4	97.4	97.4
average	91.2	91.2	96.8	97.1

Visualization

Acknowledgement

We borrow some codes from SimSiam, STN and PaDiM

Contact

If you have any problem with this code, please feel free to contact huangchaoqin@sjtu.edu.cn.

Haoyan-Guan/RegAD