Enhancement-Driven Pretraining for Robust Fingerprint Representation
Ekta Gavas
1 a
, Kaustubh Olpadkar
2 b
and Anoop Namboodiri
1 c
Centre for Visual Information Technology, International Institute of Information Technology, Hyderabad, India
Stony Brook University, U.S.A.
Fingerprint Representation Learning, Fingerprint Verification, Self-Supervised Learning, Deep Learning.
Fingerprint recognition stands as a pivotal component of biometric technology, with diverse applications from
identity verification to advanced search tools. In this paper, we propose a unique method for deriving robust
fingerprint representations by leveraging enhancement-based pre-training. Building on the achievements of U-
Net-based fingerprint enhancement, our method employs a specialized encoder to derive representations from
fingerprint images in a self-supervised manner. We further refine these representations, aiming to enhance the
verification capabilities. Our experimental results, tested on publicly available fingerprint datasets, reveal a
marked improvement in verification performance against established self-supervised training techniques. Our
findings not only highlight the effectiveness of our method but also pave the way for potential advancements.
Crucially, our research indicates that it is feasible to extract meaningful fingerprint representations from de-
graded images without relying on enhanced samples.
Fingerprint recognition remains a cornerstone in bio-
metric identification, valued for its uniqueness, per-
manence, and user-friendliness (Maltoni et al., 2022;
Wayman et al., 2005; Allen et al., 2005). As demand
in law enforcement, personal identification, and se-
cure authentication continues to rise, the need to en-
hance precision and efficiency in fingerprint recogni-
tion systems becomes increasingly vital (Allen et al.,
Despite advancements in the field, challenges per-
sist, including handling partial or distorted finger-
prints, managing high interclass similarity, and ad-
dressing the expansive dimensionality of the feature
space (Maltoni et al., 2022; Hong et al., 1998; Cap-
pelli et al., 2007). Many state-of-the-art works in fin-
gerprint matching rely on minutia-based approaches
(Ratha et al., 1996; Chang et al., 1997; Maltoni et al.,
2022; Cappelli et al., 2010b; Cappelli et al., 2010a;
Jain et al., 2001; Jain et al., 1997). This involves
extracting minutiae and matching templates to deter-
mine similarity, but traditional minutia-based meth-
ods face limitations like noise sensitivity and diffi-
culty with partial prints (Maltoni et al., 2009; Hong
et al., 1998; Maltoni et al., 2009; Zaeri, 2011).
In contrast, Convolutional Neural Networks
(CNNs) present a contemporary solution, effec-
tively overcoming limitations and improving accu-
racy. CNNs handle partial prints, tolerate distortions,
and adapt to diverse finger conditions, showcasing
scalability and efficient comparison even with grow-
ing databases (Nguyen et al., 2018; Deshpande et al.,
2020; Darlow and Rosman, 2017; Tang et al., 2017;
Engelsma et al., 2019).
The surge in self-supervised learning techniques
in machine learning has extended to fingerprint bio-
metrics (Jaiswal et al., 2020; Liu et al., 2021; Jing
and Tian, 2020). Offering solutions to challenges
in data acquisition, self-supervised learning bypasses
time-consuming labeling processes. In this paper,
we explore the potential of deep CNNs for supe-
rior matching performance, proposing a pretraining
technique based on U-Net for fingerprint enhance-
ment. The U-Net model, known for biomedical im-
age segmentation (Ronneberger et al., 2015), effec-
tively enhances fingerprints by extracting contextual
information, aiming to derive compact, discrimina-
tive fingerprint embeddings. Our study pursues two
objectives: proposing a pretraining technique with
Gavas, E., Olpadkar, K. and Namboodiri, A.
Enhancement-Driven Pretraining for Robust Finger print Representation Learning.
DOI: 10.5220/0012474900003660
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 2: VISAPP, pages
ISBN: 978-989-758-679-8; ISSN: 2184-4321
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
U-Net and assessing the efficacy of these represen-
tations through verification performance against ex-
isting self-supervised methods. We experiment with
training and inference techniques to optimize the use
of representations for fingerprint verification tasks.
This paper aims to deepen our understanding of fin-
gerprint recognition, inspiring future progress in this
1.1 Contributions
Here are the main contributions of this work:
1. We suggest a pre-training technique with fin-
gerprint enhancement task on our encoder and
demonstrate the usefulness of this approach in
representation learning in self-supervised setting.
2. We describe a method to fine-tune the learned em-
beddings for fingerprint verification task.
3. We evaluate our approach with various evaluation
metrics demonstrating its effectiveness in finger-
print verification task and also provide a compar-
ison with previous state-of-the-art self-supervised
learning methods.
1.2 Related Work
The need for improved fingerprint recognition tools
has spurred the development of effective fingerprint
representation methods. Various approaches, draw-
ing on domain knowledge, have enhanced the accu-
racy and speed of fingerprint identification (Engelsma
et al., 2019; Tang et al., 2017). This paper explores
a pretraining technique, focusing on an enhancement
task to optimize model learning for representation.
1.2.1 Image Enhancement
Early fingerprint image enhancement methods, such
as Gabor filters and Fourier Transform, faced chal-
lenges with poor quality, noise, and pattern varia-
tions (Greenberg et al., 2002; Hong et al., 1998; Kim
et al., 2002; Yang et al., 2002; Liu et al., 2014;
Sherlock et al., 1992; Chikkerur et al., 2005; Rah-
man et al., 2008). Convolutional Neural Networks
(CNNs), adept at hierarchical learning, have proven
effective in capturing minutiae and latent features, en-
hancing recognition accuracy (Nguyen et al., 2018;
Deshpande et al., 2020; Tang et al., 2017). U-Net,
originally designed for biomedical image segmenta-
tion, has been adapted for fingerprint enhancement
(Ronneberger et al., 2015). Various modifications
to U-Net, tailored for fingerprint enhancement tasks,
have been proposed (Gavas and Namboodiri, 2023;
Qian et al., 2019; Liu and Qian, 2020).
1.2.2 Self-supervised Learning Techniques
Self-supervised learning, an alternative to traditional
supervised learning, capitalizes on unlabeled data us-
ing pretext tasks for feature representation (Jaiswal
et al., 2020; Jing and Tian, 2020). Contrastive learn-
ing, a cornerstone of self-supervised learning, dif-
ferentiates between similar and dissimilar instances
(Liu et al., 2021). Techniques like SimCLR, MoCo,
BYOL, SwAV, and Noise Contrastive Estimation
showcase the diversity of contrastive learning ap-
proaches (Chen et al., 2020a; Chen et al., 2020b;
Grill et al., 2020; Caron et al., 2021; Gutmann and
arinen, 2010). These methods provide insights
into contrastive learning’s potential applications in
fingerprint biometrics.
The methodology for our research is constructed
around a two-stage framework to probe the potential
of self-supervised learning in fingerprint representa-
tion learning. A broad overview of the process is as
Stage 1: Self-Supervised Pre-training: This is
the initial stage of our methodology, in which
we perform pre-training of our models in a self-
supervised manner. It includes the application of
both existing self-supervised learning techniques
as well as our novel enhancement-based approach
for this task. This stage intends to leverage the
power of unlabeled data to learn meaningful rep-
resentations that can serve as a starting point for
subsequent stages. Notably, for all methods, we
keep the encoder architecture the same. While
other self-supervised methods traditionally use
encoders like ResNet or Vision Transformers, in
our framework we use the encoder of our U-Net-
based model to ensure a fair comparison.
Stage 2: Probing Experiments: Upon comple-
tion of the pre-training phase, we progress to the
second stage where a few linear layers (MLP)
are added on top of the frozen pre-trained en-
coder, making the representations 512-d. We then
perform probing experiments using this newly
formed model. By keeping the encoder part
frozen, we ensure that the model adapts the exist-
ing representations for the verification task with-
out altering the learned patterns from the self-
supervised pre-training phase.
Following this framework, we navigate through
the process of adapting and implementing self-
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
Figure 1: a) Architecture with verification objective i.e with binary classifier (at training and inference) b) Architecture to
compute similarity scores (at inference). The dotted arrows indicate networks having tied weights (siamese network structure).
Figure 2: U-Net architecture for enhancement task for the pre-training stage in the self-supervised setting. For representation
learning, the decoder is discarded and the binary classifier is attached.
supervised learning techniques, exploring a U-Net-
based pre-training strategy, and conducting probing
experiments with pre-trained networks. The sections
below provide a detailed overview of the procedures
involved in each stage.
2.1 U-Net-Based Pretraining
While applying existing self-supervised methods to
fingerprint data offers a valuable starting point, we
advocate for a self-supervised learning method tai-
lored specifically for the uniqueness of fingerprint
data. Drawing on our insights from U-Net-based en-
hancement works, our approach employs the training
of a fingerprint enhancement model as a form of self-
We employ U-Net-based fingerprint enhancement
for pre-training, hypothesizing that the U-Net en-
coder, trained on fingerprint enhancement, holds valu-
able fingerprint representations. Enhancing a fin-
gerprint image becomes an effective self-supervised
task, encouraging the model to learn useful, finger-
print representations. The pre-trained encoder al-
ready encapsulates crucial information about the fin-
gerprint, providing a foundation for further represen-
tation learning. The quality of these initial repre-
sentations hinges on the efficacy of the U-Net-based
enhancement model, emphasizing the significance of
the model’s design and training.
For enhancement-based pre-training, we use the
basic U-Net architecture (Figure 2) to optimize the
fingerprint enhancement task. This simple image-to-
image network takes a degraded fingerprint image as
input, degraded with various noises. The network
aims to predict an enhanced version of the finger-
print image by removing noise while maintaining and
restoring the ridge structure. This ensures the network
learns minute details of fingerprint structure and en-
hances it where possible, aiding robust feature repre-
sentation extraction later. We term it self-supervision
as we use supervision from the enhancement task in-
directly. This leverages a smaller amount of labeled
Enhancement-Driven Pretraining for Robust Fingerprint Representation Learning
data with limited impressions and identities.
Table 1 are the results of the first stage of our net-
work where we are pre-training the U-Net model for
enhancement task. Results of this pre-training stage
are demonstrated in Table 1.
2.2 Learning Fingerprint
After the self-supervised pre-training, we conduct the
probing experiments using the pre-trained networks.
These experiments aim to assess the usefulness of the
learned representations for the task of fingerprint ver-
ification. For this, we add a 3-layer MLP projection
head on top of the frozen encoder part of the pre-
trained network. We then train this model using a
Sentence-BERT-like (Reimers and Gurevych, 2019)
siamese architecture, with a limited amount of labeled
data for the fingerprint verification task. We con-
catenate the fingerprint representations u and v of the
image pair with the element-wise difference |u v|
and then pass it through the linear layers and train
it for binary-classification objective as illustrated in
Figure 1. By keeping the encoder part frozen, the
model learns to adapt the existing representations for
the verification task, without changing the underlying
learned patterns. This approach allows us to leverage
a large amount of unlabeled data to learn initial repre-
sentations and a limited amount of labeled data for su-
pervised adaptation. Note that in the supervised fine-
tuning, allowing modifications in the encoder weights
can lead to higher performance on the end task, which
is the future scope of this work. As our goal here
is to examine the robustness of the learned represen-
tations by different pre-training techniques, we keep
the encoder frozen. In summary, the combination
of self-supervised pre-training with supervised fine-
tuning offers a promising learning framework for fin-
gerprint biometrics. Our methodology aims to lever-
age the strengths of both self-supervised and super-
vised learning, offering a pathway towards robust, ef-
ficient, and data-savvy fingerprint biometrics systems.
In this section, we discuss the experiments performed
to evaluate our proposed approach’s efficacy. We
cover the specifics of our experimental setup, includ-
ing the datasets used, the training details, and the eval-
uation metrics employed.
3.1 Datasets and Preprocessing
This study employs datasets comprising synthetic and
real-world fingerprint images from SFinGe (Cappelli,
2004), FVC (Maio et al., 2002a; Maio et al., 2002b;
Maio et al., 2004), and NIST SD-302 (Fiumara et al.,
2019). SFinGe simulates real-world challenges, while
FVC and NIST SD-302 offer large-scale, realistic fin-
gerprint data for generalizability. This dataset com-
bination enables model training and evaluation under
diverse conditions. Synthetic data provides scalability
and control, while real-world data ensures applicabil-
ity. During self-supervised pre-training, only train-
ing dataset fingerprint images are used, without la-
bels. Ground truth is needed for the enhancement
task, obtained from clean images for SFinGe and gen-
erated for NIST SD-302 and FVC using a classical ap-
proach (Hong et al., 1998). The next phase involves
a binary classification task for fingerprint verification.
Data augmentation, vital for self-supervised learning,
employs random transformations like rotation, color
jitter, resize, crop, and Gaussian blur.
3.2 Implementation Details
We perform experiments using the PyTorch (Paszke
et al., 2017) framework on an Nvidia GeForce RTX
2080 Ti GPU for training.
Our proposed enhancement-based pre-training
utilizes the U-Net architecture. This U-Net encoder
is employed consistently for pre-training with other
self-supervised methods to ensure fair comparison.
The U-Net has a depth of 5 layers, each with 2
convolutions, and expects 512 x 512-pixel grayscale
fingerprint images. The encoder outputs a 4096-
dimensional vector bottleneck, reduced to 512-d with
an MLP projection head. Depth-wise convolutions
minimize parameters. We use L
loss for U-Net’s
enhancement-based pre-training, adopting losses de-
scribed in respective papers for other techniques.
For pre-training with existing self-supervised
methods, a grid search identifies optimal hyperparam-
eters. Models are pre-trained for 50 epochs with early
In probing experiments, MLP projection head
weights are adapted for verification while keeping
encoder weights fixed. We create 1:3 positive-to-
negative pairs for training and testing verification sets
from each dataset. After training, models are eval-
uated on test sets, reporting metrics like verification
accuracy, precision, recall, and F1-score. Results are
presented in two ways: 1) using the binary classifier
over the MLP projection head (Figure 1-a) and 2) uti-
lizing representations with thresholds on cosine sim-
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
Table 1: Enhancement pre-training stage results with U-Net architecture.
Raw Images 0.595 113.23 6.53 33.42
Enhancement U-Net 0.903 39.38 16.72 51.26
Figure 3: Degraded (top row) and Enhanced (bottom row) image pairs on FVC dataset from enhancement pre-training.
Table 2: Verification accuracy on SFinGe test dataset with genuine and imposter pairs.
SFinGe - Accuracy
Classification Similarity
Imposter Genuine Entire Data Imposter Genuine Entire Data
SimCLR 0.968 0.881 0.946 0.982 0.749 0.923
SimSiam 0.972 0.362 0.819 0.888 0.648 0.828
MoCo 0.963 0.881 0.942 0.955 0.845 0.927
BYOL 0.96 0.825 0.926 0.963 0.718 0.901
Ours 0.982 0.886 0.958 0.975 0.847 0.943
Table 3: F1 score on SFinGe test dataset with genuine and imposter pairs.
SFinGe - F1 score
Classification Similarity
Imposter Genuine Entire Data Imposter Genuine Entire Data
SimCLR 0.98 0.8 0.803 0.98 0.78 0.781
SimSiam 0.96 0.44 0.442 0.92 0.47 0.469
MoCo 0.98 0.79 0.785 0.97 0.74 0.737
BYOL 0.97 0.74 0.742 0.97 0.69 0.689
Ours 0.99 0.86 0.858 0.98 0.81 0.821
ilarity (Figure 1-b). The first method evaluates the
model as an end-to-end verification network, while
the second explores the potential of learned represen-
tations for similarity search and recognition tasks.
3.3 Results
The models are first pre-trained to learn finger-
print representations using the enhancement-based
approach and various self-supervised learning strate-
gies. Because these representations are not explic-
itly trained for fingerprint verification or identifica-
tion, using them directly for evaluation is inappropri-
ate. To gauge the stability and usefulness of these
learned representations, we add linear layers to the
frozen pre-trained encoders and then train the mod-
els for fingerprint verification tasks. The encoders re-
main frozen, allowing only the weights of the MLP
to adjust to the task, keeping the original representa-
tions unchanged. This setup aids in comparing the ef-
ficacy of different self-supervised learning techniques
against our method. The results of our probing exper-
iments are presented in Table 4 (Verification Accu-
racy) and 5 (F1-score). The verification accuracy and
F1-score on the SFinGe test set are shown in Tables 2
and 3 respectively. Figure 3 shows a few sample pairs
of input and predicted images from the pre-trained U-
Net model on the enhancement task used in our ap-
Enhancement-Driven Pretraining for Robust Fingerprint Representation Learning
Table 4: Verification accuracy on FVC test dataset with genuine and imposter pairs.
FVC - Accuracy
Classification Similarity
Imposter Genuine Entire Data Imposter Genuine Entire Data
SimCLR 0.915 0.619 0.841 0.943 0.537 0.841
SimSiam 0.956 0.122 0.747 0.387 0.733 0.473
MoCo 0.902 0.522 0.807 0.896 0.56 0.812
BYOL 0.886 0.568 0.806 0.926 0.477 0.813
Ours 0.957 0.73 0.900 0.933 0.818 0.904
Table 5: F1 score on FVC test dataset with genuine and imposter pairs.
FVC - F1 score
Classification Similarity
Imposter Genuine Entire Data Imposter Genuine Entire Data
SimCLR 0.94 0.5 0.502 0.95 0.51 0.51
SimSiam 0.94 0.16 0.156 0.55 0.19 0.186
MoCo 0.93 0.42 0.417 0.92 0.43 0.431
BYOL 0.92 0.42 0.421 0.94 0.43 0.432
Ours 0.97 0.68 0.679 0.96 0.66 0.659
Our approach is compared with methods like Sim-
CLR v2, SimSiam, MoCo v2, and BYOL on the
SFinGe and FVC test sets for fingerprint verification.
Verification accuracy serves as the evaluation metric
for each method. The test data for fingerprint ver-
ification consists of a 1:3 ratio of positive to nega-
tive pairs, setting the random guess accuracy at 75%.
Verification accuracy is measured in two ways as de-
scribed before. This is presented in the below tables
under the ‘Classifier’ column. The second way is rep-
resented under the ‘Similarity’ column in the tables.
Moreover, we also report the ROC curves in Figure 4
for both datasets.
As seen from the results, our enhancement-based
pre-training method consistently outperforms other
self-supervised strategies across both test sets. Sim-
CLRv2 also consistently performs well. SimSiam
and BYOL methods show comparatively poor perfor-
mance. It is noteworthy that all models perform better
on the SFinGe test set than on the FVC test set. We
believe this is due to two primary factors: the train-
ing sets contain more data from SFinGe than FVC,
potentially resulting in a bias towards the former, and
SFinGe is a synthetic dataset while FVC consists of
real fingerprints, making the latter more challenging.
Hence, the performance of models on FVC data is the
real measure of the efficacy of models. Importantly,
our method also provides superior performance when
verification is based on the similarity of the represen-
tations, suggesting that the learned representations are
also useful for fingerprint recognition.
Despite promising results, our model demonstrates
greater efficacy on the synthetic SFinGe dataset than
on the real-world FVC dataset. This could be at-
tributed to potential bias from underrepresentation of
FVC data in training sets and complexities in real-
life fingerprint data. Another limitation is the lack
of specific training and evaluation for the fingerprint
recognition task. While our model shows potential, a
dedicated evaluation is essential for a comprehensive
understanding of its performance. The effectiveness
of self-supervised learning relies on data quality and
diversity, and our study used linear probing, leaving
room to explore alternative approaches like softmax
or ArcFace-based classification.
Future work should address limitations by in-
corporating a more diverse set of real-world finger-
print datasets during training. Exploring the option
of training the encoder with a smaller learning rate,
rather than freezing it, could enhance generalizabil-
ity and robustness. Specific training and evaluation
for the recognition task, investigating alternative lin-
ear probing techniques, and exploring various self-
supervised learning methods are valuable directions
for further optimization.
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
Figure 4: ROC curve based on similarity scores on SFinGe dataset(left) and FVC dataset (right).
In this study, we explored diverse self-supervised
learning techniques to pre-train a model for effective
fingerprint representations in recognition and verifi-
cation. A novel approach involved leveraging finger-
print enhancement as a self-supervised pre-training
method. Probing experiments assessed the effective-
ness of learned representations across various pre-
training strategies. Comparisons against SimCLR
v2, SimSiam, MoCo v2, and BYOL methods on
SFinGe and FVC datasets consistently showed our
method’s superior verification performance. Notably,
our model excelled in similarity-based verification,
underscoring its effectiveness in fingerprint recogni-
tion tasks. However, models performed better on
the synthetic SFinGe dataset, hinting at potential bias
in the training set and real-world data complexities.
Future work will expand to diverse real-world fin-
gerprint datasets, improving model generalizability.
We’ll also explore additional self-supervised methods
for enhanced adaptability to real-world complexities,
emphasizing the potential of self-supervised learning
in fingerprint biometrics while pointing to areas for
exploration and refinement.
Allen, R., Sankar, P., and Prabhakar, S. (2005). Fingerprint
Identification Technology, pages 22–61. Springer
London, London.
Cappelli, R. (2004). Sfinge : an approach to synthetic fin-
gerprint generation.
Cappelli, R., Ferrara, M., and Maltoni, D. (2010a). Minu-
tia cylinder-code: A new representation and match-
ing technique for fingerprint recognition. IEEE trans-
actions on pattern analysis and machine intelligence,
Cappelli, R., Ferrara, M., Maltoni, D., and Tistarelli, M.
(2010b). Mcc: A baseline algorithm for fingerprint
verification in fvc-ongoing. In 2010 11th Interna-
tional Conference on Control Automation Robotics &
Vision, pages 19–23. IEEE.
Cappelli, R., Maio, D., Lumini, A., and Maltoni, D. (2007).
Fingerprint image reconstruction from standard tem-
plates. IEEE transactions on pattern analysis and ma-
chine intelligence, 29(9):1489–1503.
Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P.,
and Joulin, A. (2021). Unsupervised learning of visual
features by contrasting cluster assignments.
Chang, S.-H., Cheng, F.-H., Hsu, W.-H., and Wu, G.-Z.
(1997). Fast algorithm for point pattern matching:
invariant to translations, rotations and scale changes.
Pattern recognition, 30(2):311–320.
Chen, T., Kornblith, S., Swersky, K., Norouzi, M., and Hin-
ton, G. (2020a). Big self-supervised models are strong
semi-supervised learners.
Chen, X., Fan, H., Girshick, R., and He, K. (2020b). Im-
proved baselines with momentum contrastive learn-
Chikkerur, S., Govindaraju, V., and Cartwright, A. N.
(2005). Fingerprint image enhancement using stft
analysis. In International Conference on Pat-
tern Recognition and Image Analysis, pages 20–29.
Darlow, L. N. and Rosman, B. (2017). Fingerprint minutiae
extraction using deep learning. In 2017 IEEE Interna-
tional Joint Conference on Biometrics (IJCB). IEEE.
Deshpande, U. U., Malemath, V., Patil, S. M., and Chau-
gule, S. V. (2020). Cnnai: a convolution neural
network-based latent fingerprint matching using the
combination of nearest neighbor arrangement index-
ing. Frontiers in Robotics and AI, 7:113.
Engelsma, J. J., Cao, K., and Jain, A. K. (2019). Learning
a fixed-length fingerprint representation. IEEE trans-
actions on pattern analysis and machine intelligence,
Fiumara, G., Flanagan, P., Grantham, J., Ko, K., Mar-
shall, K., Schwarz, M., Tabassi, E., Woodgate, B., and
Enhancement-Driven Pretraining for Robust Fingerprint Representation Learning
Boehnen, C. (2019). Nist special database 302: Nail
to nail fingerprint challenge.
Gavas, E. and Namboodiri, A. (2023). Finger-UNet: A u-
net based multi-task architecture for deep fingerprint
enhancement. In Proceedings of the 18th Interna-
tional Joint Conference on Computer Vision, Imag-
ing and Computer Graphics Theory and Applications.
SCITEPRESS - Science and Technology Publications.
Greenberg, S., Aladjem, M., and Kogan, D. (2002). Finger-
print image enhancement using filtering techniques.
Real-Time Imaging, 8(3):227–236.
Grill, J.-B., Strub, F., Altch
e, F., Tallec, C., Richemond, P.,
Buchatskaya, E., Doersch, C., Avila Pires, B., Guo,
Z., Gheshlaghi Azar, M., et al. (2020). Bootstrap your
own latent-a new approach to self-supervised learn-
ing. Advances in neural information processing sys-
tems, 33:21271–21284.
Gutmann, M. and Hyv
arinen, A. (2010). Noise-contrastive
estimation: A new estimation principle for unnormal-
ized statistical models. In Proceedings of the thir-
teenth international conference on artificial intelli-
gence and statistics, pages 297–304. JMLR Workshop
and Conference Proceedings.
Hong, L., Wan, Y., and Jain, A. (1998). Fingerprint image
enhancement: Algorithm and performance evaluation.
IEEE transactions on pattern analysis and machine
intelligence, 20(8):777–789.
Jain, A., Hong, L., and Bolle, R. (1997). On-line fingerprint
verification. IEEE transactions on pattern analysis
and machine intelligence, 19(4):302–314.
Jain, A., Ross, A., and Prabhakar, S. (2001). Finger-
print matching using minutiae and texture features. In
Proceedings 2001 International Conference on Image
Processing (Cat. No. 01CH37205), volume 3, pages
282–285. IEEE.
Jaiswal, A., Babu, A. R., Zadeh, M. Z., Banerjee, D., and
Makedon, F. (2020). A survey on contrastive self-
supervised learning. Technologies, 9(1):2.
Jing, L. and Tian, Y. (2020). Self-supervised visual feature
learning with deep neural networks: A survey. IEEE
transactions on pattern analysis and machine intelli-
gence, 43(11):4037–4058.
Kim, B.-G., Kim, H.-J., and Park, D.-J. (2002). New en-
hancement algorithm for fingerprint images. In 2002
International Conference on Pattern Recognition, vol-
ume 3, pages 879–882. IEEE.
Liu, M., Chen, X., and Wang, X. (2014). Latent fingerprint
enhancement via multi-scale patch based sparse repre-
sentation. IEEE Transactions on Information Foren-
sics and Security, 10(1):6–15.
Liu, M. and Qian, P. (2020). Automatic segmentation and
enhancement of latent fingerprints using deep nested
unets. IEEE Transactions on Information Forensics
and Security, 16:1709–1719.
Liu, X., Zhang, F., Hou, Z., Mian, L., Wang, Z., Zhang, J.,
and Tang, J. (2021). Self-supervised learning: Gener-
ative or contrastive. IEEE transactions on knowledge
and data engineering, 35(1):857–876.
Maio, D., Maltoni, D., Cappelli, R., Wayman, J. L., and
Jain, A. K. (2002a). Fvc2000: Fingerprint verification
competition. IEEE transactions on pattern analysis
and machine intelligence, 24(3):402–412.
Maio, D., Maltoni, D., Cappelli, R., Wayman, J. L., and
Jain, A. K. (2002b). Fvc2002: Second fingerprint ver-
ification competition. In 2002 International Confer-
ence on Pattern Recognition, volume 3. IEEE.
Maio, D., Maltoni, D., Cappelli, R., Wayman, J. L., and
Jain, A. K. (2004). Fvc2004: Third fingerprint ver-
ification competition. In Zhang, D. and Jain, A. K.,
editors, Biometric Authentication, pages 1–7, Berlin,
Heidelberg. Springer Berlin Heidelberg.
Maltoni, D., Maio, D., Jain, A. K., and Feng, J. (2009). Fin-
gerprint Matching, pages 167–233. Springer London,
Maltoni, D., Maio, D., Jain, A. K., and Feng, J. (2022).
Fingerprint Sensing, pages 63–114. Springer Interna-
tional Publishing, Cham.
Nguyen, D.-L., Cao, K., and Jain, A. K. (2018). Robust
minutiae extractor: Integrating deep networks and fin-
gerprint domain knowledge. In 2018 International
Conference on Biometrics (ICB), pages 9–16. IEEE.
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E.,
DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and
Lerer, A. (2017). Automatic differentiation in pytorch.
Qian, P., Li, A., and Liu, M. (2019). Latent fingerprint en-
hancement based on denseunet. In 2019 International
Conference on Biometrics (ICB), pages 1–6. IEEE.
Rahman, S. M., Ahmad, M. O., and Swamy, M. (2008).
Improved image restoration using wavelet-based de-
noising and fourier-based deconvolution. In 2008 51st
Midwest Symposium on Circuits and Systems, pages
249–252. IEEE.
Ratha, N. K., Karu, K., Chen, S., and Jain, A. K. (1996).
A real-time matching system for large fingerprint
databases. IEEE transactions on pattern analysis and
machine intelligence, 18(8):799–813.
Reimers, N. and Gurevych, I. (2019). Sentence-bert: Sen-
tence embeddings using siamese bert-networks.
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:
Convolutional networks for biomedical image seg-
mentation. In International Conference on Medical
image computing and computer-assisted intervention,
pages 234–241. Springer.
Sherlock, B., Monro, D., and Millard, K. (1992). Algorithm
for enhancing fingerprint images. Electronics letters,
Tang, Y., Gao, F., Feng, J., and Liu, Y. (2017). Fingernet:
An unified deep network for fingerprint minutiae ex-
traction. In 2017 IEEE International Joint Conference
on Biometrics (IJCB), pages 108–116. IEEE.
Wayman, J., Jain, A., Maltoni, D., and Maio, D. (2005).
An Introduction to Biometric Authentication Systems,
pages 1–20. Springer London, London.
Yang, J., Liu, L., and Jiang, T. (2002). Improved method
for extraction of fingerprint features. In Second Inter-
national Conference on Image and Graphics, volume
4875, pages 552–558. SPIE.
Zaeri, N. (2011). Minutiae-based fingerprint extraction and
recognition. Biometrics.
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications