Elin Stenbäcken and Albert Heurlin de Oliveira will present their master's thesis.

Title: Fingerprint Image Restoration Using the U-Net Deep Learning Model

Authors: Elin Stenbäcken and Albert Heurlin de Oliveira

Supervisors: Anders Heyden 

Examinator: Kalle Åström

Abstract: 

This thesis investigated the use of convolutional neural networks (CNNs) for
image restoration of fingerprint images, focusing on the reconstruction from raw
sensor images to denoised fingerprint images, suitable for matching. Fingerprint
images from sensors often introduce noise and artifacts that inhibit the matching
process. The goal was to develop a CNN-based approach capable of removing
artifacts such as moiré patters that degrade fingerprint quality, thereby restoring
the fingerprint image. A U-Net architecture was used as a baseline model and
extended with several architectural modifications such as Convolutional Block
Attention Modules (CBAM), gated skip connections and dilated convolutions.
In addition, the effects of different batch sizes and learning rates for the ADAM
optimizer were evaluated. The models were trained with synthetic fingerprint
data, including a dataset augmented with image transformations to increase
the training dataset size. The results show that CNN-based restoration can
surpass traditional ISP pipelines, especially when training and test domains are
well aligned. Learning rate selection held significant importance, with 10−4
consistently yielding the lowest False Reject Rates (FRR). Data augmentation
improved robustness and frequently reduced FRR relative to identical models
trained on non-augmented data. However, generalization to different datasets
remained limited, highlighting the need for a training dataset representative of
diverse data. The most promising model architecture for restoration was U-Net,
which is the simplest of all the proposed models. Since the training dataset was
very limited in size, the risk of overfitting was high, and it seemed that the
more complex models tended to overfit relatively quickly. The simple U-Net, in
contrast, seemed to generalize the best, and had more consistent performance
on the testing data than every other model. Overall, it seems that CNN-based
restoration methods offers a promising approach to reconstruct fingerprints from
raw sensor images, and that their role and performance could be increased with
further development.