AMSL Synthetic Fingerprint Models

This license agreement covers six AMSL fingerprint generative models:

1. AMSL StyleGAN2-ada. The model is trained from scratch with the built-in augmentation based on 408 samples from the Neurotechnology dataset collected with the Cross Match Verifer 300 sensor (https://www.neurotechnology.com/download.html). All samples are padded to 512x512 pixels prior to training. The training is performed with standard hyperparameters. The 10800 kImages model snapshot picked due to the best visual results. Due to the low number of unique identities in the training dataset, the synthetic samples may lack diversity and it is not assured that no identity leakage happens in regard to the training data.
2. AMSL pix2pix aug39k DL BN. The model which is trained based on 39k real fingerprints making use of the directed line encoding. Training is done with batch normalization for 120 epochs (60 epochs with the constant learning rate and 60 epochs with learning rate decay)
3. AMSL pix2pix aug39k DL IN. The model trained based on 39k real fingerprints making use of the directed line encoding. Training is done with instance normalization for 15 epochs.
4. AMSL pix2pix aug39k PM IN. The model trained based on 39k real fingerprints making use of the pointing minutiae encoding. Training is done with instance normalization for 15 epochs.
5. AMSL pix2pix synth50k DL IN. The model trained based on 50k synthetic fingerprints generated by the AMSL StyleGAN2-ada model making use of the directed line encoding. Training is done with instance normalization for 15 epochs.
6. AMSL pix2pix synth50k PM IN. The model trained based on 50k synthetic fingerprints generated by the AMSL StyleGAN2-ada model making use of the pointing minutiae encoding. Training is done with instance normalization for 15 epochs.

The 39k training dataset is created by applying data augmentation (horizontal flip and rotations to +/-5, +/-10, +/-15 and +/-20 degree) to 2168 samples taken from the three following datasets:

- The Neurotechnology dataset collected with a Cross Match Verifer 300, 408 samples, https://www.neurotechnology.com/download.html
- The FVC2002 DB1 A+B dataset collected with a TouchView II scanner by Identix, 880 samples, http://bias.csr.unibo.it/fvc2002/databases.asp
- The FVC2004 DB1 A+B dataset collected with a Cross Match Verifer 300, 880 samples, http://bias.csr.unibo.it/fvc2004/databases.asp

By downloading any of the models, you agree the following license conditions:

The models will not be used to harm anyone!

The models are distributed in the hope that they will be useful, but WITHOUT ANY WARRANTY. No author or distributor accepts responsibility to anyone for the consequences of using it or for whether it serves any particular purpose or works at all.

The models are licensed under the Creative Commons Attribution-ShareAlike 3.0 License (CC BY-SA 3.0). You are free to share, to copy, distribute and transmit the data and to remix or adapt the data under the following conditions:

If you use any of the models in your research leading to a publication, you must cite the following paper:

@inproceedings{Makrushin23b,
author = {Andrey Makrushin and Venkata Srinath Mannam and Jana Dittmann},
title = {Data-Driven Fingerprint Reconstruction from Minutiae Based on Real and Synthetic Training Data},
booktitle = {Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 4: VISAPP},
pages = {229-237},
publisher = {{SCITEPRESS}},
year = {2023},
isbn = {978-989-758-634-7},
issn = {2184-4321}
}

If you alter, transform or build upon the data, you may distribute the resulting work only under the same, similar or a compatible license.

You are not allowed to use or distribute any of the models for commercial purposes as, for example, licensing or selling the data, or using the data with a purpose to procure a commercial gain without written permission.

All rights not expressly granted to you are reserved by the authors.