A collection industrial anomaly detection dataset for Printcd Cicuils Board (PCB).

• Data description
• Data download
• Data preparation
• Citation
• License

The PCB-Bank dataset is composed of Printed Circuit Board (PCB) data collected from publicly available industrial anomaly detection datasets. There are 6,749 images with 5,352 normal and 1,397 anomalous samples. It contains 7 subsets corresponding to 7 different PCB as shown in the above figure. 4 subsets are different types of printed circuit boards (PCB) from VisA. 1 subset is top-view of PCB from RealIAD. We use AnyLabeling tool to crop each PCB from the original image in VISION PCB2 subset and construct 2 subsets. More details are shown in the following table.

You can download it by this BaiduYun[Code:mx57]. Note that you should send an email to [email protected] to get the PCB5 and PCB6 (from RealIAD) subsets. The id2class map functions for multi-class masks can be found in ./id2class.py. Here the masks for normal images are not stored to save space. The data tree of the downloaded data is as follows.

PCB-Bank
|-- pcb1
|-----|--- Data
|-----|-----|----- Images
|-----|-----|--------|------ Anomaly 
|-----|-----|--------|------ Normal 
|-----|-----|----- Masks
|-----|-----|--------|------ Anomaly 
|-----|--- image_anno.csv
|-- pcb2
|-----|----- ...

1-class setup: It means model trained with normal samples and test on both normal and anomaly samples. We follows the VisA split setting on PCB1,PCB2,PCB3,PCB4. 200 normal samples from PCB5,PCB6,PCB7 are used for training, while others are used for testing. We give a sample command line for 1-class setup preparation as follows.

python prepare_data.py --split-type 1cls --data-folder <PCB-Bank_path> --save-folder <1cls_output_path> --split-file <1cls.csv_path>

The data tree of the reorganized 1-class setup is as follows.

PCB-Bank-1cls
|-- 1cls
|-----|--- pcb1
|-----|-----|----- ground_truth
|-----|-----|----- test
|-----|-----|-------|------- good 
|-----|-----|-------|------- bad 
|-----|-----|----- train
|-----|-----|-------|------- good
|-----|--- pcb2
|-----|--- ...

Specifically, the reorganized data for 1-class setup follows the data tree of MVTec-AD. For each object, the data has three folders:

• 'train', which includes the normal training images
• 'test', which includes the normal and anomalous test images
• 'ground_truth', which includes the pixel-level annotations of anomalous images

Please cite the following paper if this dataset helps your project:

@article{hang2024glad,
  title={GLAD: Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection},
  author={Yao, Hang and Liu, Ming and Wang, Haolin and Yin, Zhicun and Yan, Zifei and Hong, Xiaopeng and Zuo, Wangmeng},
  journal={arXiv preprint},
  year={2024}
}

The data is released under the CC BY 4.0 license.