PCB-Bank

PCB-Bank is a unified printed circuit board anomaly detection benchmark constructed from three public industrial datasets:

• PCB1–PCB4: VisA
• PCB5: Real-IAD
• PCB6–PCB7: VISION

All subsets are reorganized into a common one-class anomaly detection format compatible with common MVTec AD / VisA-style pipelines.

Dataset Structure

PCB-Bank/
├── pcb1/
├── pcb2/
├── pcb3/
├── pcb4/
├── pcb5/
├── pcb6/
└── pcb7/

Each category follows:

pcbX/
├── train/
│   └── good/
├── test/
│   ├── good/
│   └── bad/
└── ground_truth/
    └── bad/

• train/good: normal training images
• test/good: normal test images
• test/bad: anomalous test images
• ground_truth/bad: binary pixel-level anomaly masks corresponding to test/bad

Data Sources

Category	Source
PCB1	VisA
PCB2	VisA
PCB3	VisA
PCB4	VisA
PCB5	Real-IAD
PCB6	VISION
PCB7	VISION

For PCB5, the original Real-IAD train/test split is retained and all available camera views are included.

For PCB6 and PCB7, the original COCO annotations from VISION are converted into binary pixel-level anomaly masks using the provided segmentation polygons.

Usage

The dataset can be used directly with anomaly detection pipelines expecting a MVTec AD / VisA-style directory structure.

Example:

dataset_root = "/path/to/PCB-Bank"
category = "pcb5"

Notes

This repository only reorganizes and preprocesses the original datasets into a common anomaly detection format. Please refer to the original datasets for licensing terms, dataset details, and attribution requirements.

Original Datasets

Please cite the corresponding original datasets when using PCB-Bank:

• VisA
• Real-IAD
• VISION

If results are reported across the complete PCB-Bank benchmark, we recommend acknowledging all three source datasets.