MPI-PCB Dataset

The Multi-Perspective and Illumination PCB (MPI-PCB) dataset contains 1742 images with dimensions of 4096 × 2816-pixel showing an unmodified PCB from a gas pump. The images were captured using a Canon EOS 1100D camera with 18–55 mm lenses. The set also contains 55 images showing the board with modifications manually added by the authors, which are meant to be representative of situations encountered in actual frauds. These samples also contains labeled semantic segmentation masks that represents the localization of the frauds, being able to execute the comparision between different techniques.

Images were captured from a generally overhead view, but without strict demands on position or illumination, as expected in a real-world situation. To reduce variations that may occur in the image registration step and focus on the anomaly detection problem, the dataset contains the images after the registration procedure described in the original paper: Image-Based Detection of Modifications in Assembled PCBs with Deep Convolutional Autoencoders (https://www.mdpi.com/1424-8220/23/3/1353)

The dataset is available on several formats following what was used on the paper:

The images in the original size with alignment.

The images already split in grids as presented in the paper.

The .tfrecord used in the training using Google Colab TPUs*.

In the images format the dataset were organized following the MVTec-AD structure, so it is easy to use the same source codes for MVTec-AD.

Please cite:

@article{candido2023image,
title={Image-Based Detection of Modifications in Assembled PCBs with Deep Convolutional Autoencoders},
author={Candido de Oliveira, Diulhio and Nassu, Bogdan Tomoyuki and Wehrmeister, Marco Aurelio},
journal={Sensors},
volume={23},
number={3},
pages={1353},
year={2023},
publisher={MDPI}
}