Neural Fingerprint Enhancement

Edward Raff
Booz Allen Hamilton
[email protected]

Abstract—Biometrics fingerprint matching has been done with developed software to generate realistic fingerprint images,
a heavily hand-tuned and designed process of classical computer allowing the user to set and control different properties of the
vision techniques for several decades. This approach has led fingerprint such as rotations, noise, damage, and more. In this
to accurate solutions for solving crimes today and, as such,
little effort has been devoted to using deep learning in this work, we will leverage Cappelli’s software to generate training
domain. Given that convolutional neural networks have shown data for a denoising convolutional auto-encoder. Specifically,
dominance for most other image-based problems, we re-evaluate we use a neural network to learn the inverse of the noising
their potential for improving the fingerprint matching process. By process that converts the ground-truth ridge patterns into a
leveraging synthetic data generators, we show that one can train noisy and realistic fingerprint. The goal will be to then apply
a neural fingerprint enhancer to improve matching accuracy on
real fingerprint images. Our approach is both simple in design the learned function as a neural fingerprint enhancement (NFE)
and for potential deployment and adoption in real world use. technique before the minutiae extraction process, similar to
Index Terms—fingerprints, neural networks, biometrics how Gabor filters [25] and 2-D Fourier Transforms [5] are
often used for fingerprint enhancement today.
I. I NTRODUCTION We specifically look at enhancing images before processing
with existing minutiae extractors and matchers because we
Deep Learning has become a popular technique for many
believe it is a viable method to achieving real-world adoption.
biometric tasks such as face, iris, handwritten signature, and
Adoption by law enforcement will be preceded by FBI certifi-
gait recognition, yet these newer neural network techniques
cation and guidance, court admissibility, extended validation,
have found relatively little use in fingerprint matching and
and a host of other issues. By keeping all of the already
recognition [29]. Instead, two decades of classical image
accepted technologies in place, and adding one pre-processing
and signal processing techniques have been applied to the
step with our NFE, we minimize the risk and effort for law
problem with considerable success [23, 24]. By 2006, before
enforcement agencies to test and adopt our technique. It can
any resurgence of neural networks in the machine learning
be used concurrently with any existing product or methods to
community, these systems where already obtaining low error
do continuous testing and, if accepted, can simply sit in-front
rates on all but the most challenging examples [3] and became
of already trusted tools. If a failure case occurs, users can
integral to the criminal investigation process.
simply re-process the fingerprint without our NFE to see if
Fingerprint matching and recognition is generally based
their original tooling would have succeeded.
around minutiae extraction and matching [22]. The minutiae of
The rest of our paper is organized as follows. We will
a fingerprint are the locations where the ridges of a print either
address related work in using neural networks in the fingerprint
ends or combines with another ridge. Minutiae are characterized
domain in section II. In section III we will review our network
by their relative location to one another, type, and orientation.
design and our methodology to train and evaluate it. Then we
These are often represented as a graph, and graph-matching
will present results showing our NFE improves matching on
used to compare fingerprint minutiae [11]. At a high level, the
both unseen real fingerprint images, as well as a challenging
standard process for minutiae extraction takes the following
synthetic test set, in section IV. Finally we will conclude in
steps:
section V.
1) Pre-process the image to counter various forms of noise
and quality issues. II. R ELATED W ORK
2) Estimate information about the orientation of the ridges The general theme of our work is to use neural networks to
in a fingerprint. model problems that have been understood well enough that
3) Apply a process to identify minutiae locations, and use computer generated simulations are currently at or near the
pre-processed image and orientation data to characterize point of fully specifying the problem space. Shrivastava et al.
minutia. [28] used Generative Adversarial Networks (GANs) to increase
4) Post-process minutiae to remove common spurious, incor- the realism of rendered images of eyes to train a better eye
rect, and unreliable identifications. tracking system. By generating the data themselves, they know
In this work, we seek to re-evaluate the potential for neural exact ground truth labels. Tompson et al. [31] used CNNs to
networks to add value to the fingerprint matching process by reproduce the output of classical 2D and 3D fluid simulations.
exploiting the deep understanding and modeling that has been Their network could produce predictions fast enough for real-
done with fingerprint images. In particular, Cappelli [4][2] time simulations, where the classical methods are too slow for

1
 Max-Pooling

Max-Pooling

Max-Pooling
Batch-Norm

Batch-Norm

Batch-Norm
Finger Print
Conv 32x7x7 Conv 64x5x5 Conv 128x3x3

ReLU

ReLU

ReLU
Conv 256x1x1
Dilated 32x7x7 Dilated 64x5x5 Dilated 128x3x3

Sigmoid Deconv Batch-Norm Deconv Batch-Norm Deconv Batch-Norm

Activation 1x4x4 ReLU 64x4x4 ReLU 128x4x4 ReLU

Fig. 1: Diagram of network architecture. Red boxes indicate that two operations are done on the same input, and their results
are concatenated together as the final output.

such scenarios. Similar to these prior works, we will be using shallow architecture allowing for fast execution and handling
synthetically generated fingerprints to show the potential for variable input sizes. Training will be done on synthetically
neural networks to learn a more accurate fingerprint image generated fingerprints, and evaluation will occur on both real
enchantment. and synthetic prints.
The most similar prior work to our own is by Sahasrabudhe
A. Neural Network Architecture and Training
and Namboodiri [26], who trained a convolution Restricted
Boltzman Machine (RBM) to try and enhance fingerprint The overall design of our denoising neural network is
images. They trained in an unsupervised fashion on a hand- given in Figure 1 and is fully convolution. This allows our
picked subset of higher quality fingerprint images from three network to be applied to inputs of any size. This is important
different databases. Unfortunately, their approach did not since there is no fixed standard size of fingerprint images.
generalize to new fingerprints, and most analysis was restricted Our network starts with of four layers of convolutions, each
to the same three datasets that were used for training. We avoid followed by the use of batch-normalization [9], the ReLU
this overfitting problem by using synthetic data and instead non-linearity [18], and then max-pooling. For each set of
train in a supervised fashion to directly learn the denoising convolutions, we split them between the use of standard and
goal, and will show that our approach improves performance dilated convolutions [37]. We start with a total of 64 filters split
on new data. between each convolution type, doubling the number of filters
Several works have looked at using multiple neural networks after each round of pooling. The outputs of these two forms
to perform minutiae extraction, combined with additional hand of convolutions are concatenated, and then given to the next
tuned techniques to post-process results by still using classical layer of the network. We use the mix of dilation and standard
enhancement techniques like gabor filters [6, 13, 19]. None in order to capture both local fine-grained details (standard
of these prior works have evaluated the ability to perform convolutions), as well as farther non-local details via the
accurate matching with these minutia, which is the ultimate end dilation since our network is (relatively) shallow compared to
goal. In our work we measure the performance with respect modern design. After three rounds of convolution and pooling,
to the goal of accurate matching, and explicitly remove as we apply 1-by-1 convolutions to perform information sharing
much domain knowledge and hand-coded image processing across the filters. Then we apply deconvolution [38] several
from our approach. Work in fingerprint liveness detection (i.e., times to get to the correct output size, followed by the sigmoid
is the fingerprint from a real finger) has followed a similar activation to clamp values into the range of [0, 1] that are
path of using considerable hand-designed computer vision pre- appropriate for our images.
processing [20, 33]. B. Training Data
Most other works we are aware of that have used neural We use Cappelli’s [3] Synthetic Fingerprint Generator
networks in the fingerprint domain tackle specific sub-goals (SFinGe) version 4.1 to generate our training data. The SFinGe
in larger hard-coded pipelines. Olsen et al. [21] used Self software allows us to specify a number of parameters regarding
Organizing Maps to evaluate the quality of sub-portions of an the generation process, including to save ground-truth images
image. Yong-xia et al. [36] used neural networks to find the of where fingerprint ridges start and end. An example of a
core of a finger print. Sarbadhikari et al. [27] used a simple synthetic fingerprint and the ground-truth image is given in
fully connected network atop FFT based features to classify Figure 2. The goal is to train a neural network to take the raw
prints into types (left & right loop, whorl, twin loop, and plain input and learn to produce the ridge pattern. We then use the
arch). neural network’s prediction of the ridge pattern as an enhanced
image to use with standard minutiae matching and extraction
III. M ETHODOLOGY
approaches.
We will now review the methods by which we construct, For our training data we collected 600,000 image pairs
train, and evaluate our network. The network will be a relatively for training using SFinGe with all combinations of the

2
 enhancement. Significant prior testing has shown FFT and
Gabor based processing to deliver similar high performance
[30]. SourceAfis uses its own custom enhancement and ridge
smoothing algorithm developed and refined since 2009. Again,
our goal is to see that learning the image enhancement process
can improve results, thus simplifying the need for fine tuning
and maintaining such complex systems for fingerprint matching.
D. Evaluation on Real Fingerprints
The first set of datasets we will use to validate our
approach are real fingerprints from the Fingerprint Verification
Competitions (FVC) that were run in 2000 [14], 2002 [15],
(a) Fingerprint generated with (b) Ground truth ridge-line pat- and 2004 [16]. Each of these competitions has four constituent
“variable” quality setting. tern of generated print. “databases” (DBs) of prints collected from real people using
Fig. 2: Example of synthetic fingerprint from SFinGe. Left differing fingerprint readers, and under varying conditions that
image shows the final output (input to the neural network), would introduce noise (such as drying or moistening the fingers)
right shows the ground-truth ridge pattern (target output of the in order to increase the challenge of matching. The fourth
neural network). database from each of these competitions was synthetically
generated using earlier versions of SFinGe. Since our concern
is with generalization to real fingerprints, the we exclude the
following settings: Background ∈ {None, Optical, Capacitive} fourth DB from these evaluations. The FVC2004 competition,
and distribution as either “Varying quality and perturbations” or according to the authors, was meant to be “markedly more
1
“Very Low Quality”. Fingerprints were generated against 50,000 difficult than FVC2002 and FVC2000” . A summary of the
simulated fingers for each combination, with 2 impressions for databases under test is given in Table I.
each finger. Across all “Very Low Quality” generated prints we
TABLE I: Fingerprint Verification Competition (FVC) datasets.
set the horizontal and vertical translation to ±5% and rotation
to ±25°. All other parameters were left with their default Year DB# Sensor Type Image Size DPI
options, which includes an image dots per inch (DPI) of 500. 1 Low-cost Optical 300x300 500
2000 2 Low-cost Capacitive 256x364 500
C. Minutiae Extraction and Matching 3 Optical 448x478 500
We are applying neural networks only to the task of 1 Optical 388x374 500
cleaning/enhancing the input image. As such, it is still necessary 2002 2 Optical 296x560 569
3 Capacitive 300x300 500
to use other tools to perform the minutiae extraction and
matching process. If our neural fingerprint enhancement is 1 Optical 640x480 500
2004 2 Optical 328x364 500
effective, we should see performance improve with the use of 3 Thermal 300x480 512
any of the tooling. As such, we make use of three tools that
are publicly available. For each database from each competition year, the finger-
First, we use the NIST MINDTCT program for extraction prints of 10 persons are made publicly available. Each person
and BOZORTH 3 for matching [34]. We will refer to this pair as had their fingerprint taken 8 different times, resulting in a
NIST for short. Despite the age of these tools, a recent study total of 80 images. Due to the small number of total images
has found them to still be competitive in terms of matching and persons, we will perform Leave-One-Out (LOO) cross
accuracy [22]. validation on each of the FVC databases from each year. We
Second, we will use the SourceAfis [32] project, which also will record two metrics of interest, the 1-nearest neighbor
has a minutiae extraction and matching algorithm. Because (1-NN) error (i.e., how often do we fail to return another
SourceAfis is compatible with ISO/IEC 19794-2 [10], we can fingerprint from the same person) and the Area Under the ROC
use the SourceAfis matching algorithm with a different minutiae curve (AUC). Because each database has only 10 participants,
extractor. DigitalPersona, Inc. released their FingerJetFX com- it is easy for ties to occur. As such the AUC will be our
mercial solution for minutiae extraction under an open source primary focus. The AUC is the sum of area under the trade-off
license [8] . We will denote the use of just SourceAfis for both between a false positive rate and a true positive rate, and can
extraction and matching as SA, and the use of FingerJetFX for be interpreted as the quality of the ranking produced by each
extraction with SourceAfis for matching as FJ+SA. method.
Both the NIST and FJ extractors uses a 2D block FFT Evaluating on real fingerprints from a variety of sensors
in a similar style as Chikkerur et al. [5]. As such, if we helps to determine if our method is truly viable and could
see NFE provide an improvement over FJ+SA and NIST, generalize. Toward this end, we note a number of artifacts that
then we can infer that our approach improves upon the
hand turned classical computer vision approach to fingerprint 1 See http://bias.csr.unibo.it/fvc2004/databases.asp for quote.

3
exist within the FVC databases that are not modeled by our additional pre-processing, and the last two columns show the
version of SFinGe. For example, FVC2000 DB1 and DB3 have same results but with our NFE pre-processed images.
circular artifacts from the capture device present in each image.
FVC2002/DB2 is at a 13% higher DPI than what SFinGe TABLE II: Comparing fingerprint matching on multiple real
produces and FVC2004/DB3 is using a completely different fingerprint corpora. First column is the year and database
type of capture device than what SFinGe can currently model. number of the FVC competitions. Second column is the base
Finally, all images in all databases are at differing image sizes matching algorithm being used. Third and fourth column shows
compared to the 256x336 images created by SFinGe. the error rate and AUC when using each algorithm. Last two
columns show error and AUC when the image is first pre-
E. Evaluation on Synthetic Fingerprints processed with our neural finger enhancement. Best results in
Due to the intrinsic personal nature of fingerprints, there each row for each metric shown in bold, ties show in italics.
do not currently exist large corpora from which their behavior Standard NFE
can be studied. This makes determining smaller impacts on
Year/DB# Algorithm Errors AUC Errors AUC
performance difficult. Solving this problem is part of the
purpose of SFinGe’s development, as it allows us a safe way NIST 7 92.77 1 99.00
2000/1 SA 11 83.93 0 99.46
to generate testing corpora of arbitrary sizes. FJ+SA 8 88.95 0 99.87
As such, we use SFinGe to generate a test corpus with the NIST 7 96.12 4 97.30
“Very Low Quality” settings as described in subsection III-B. 2000/2 SA 3 94.48 4 95.60
We generate 100,000 total fingerprints, each with two im- FJ+SA 0 98.23 3 95.49
pressions. The first impression of all 100,000 prints will NIST 2 95.05 7 96.26
be enrolled, and querying will be done with the first 1,000 2000/3 SA 3 88.06 8 86.35
FJ+SA 2 93.70 10 86.81
second impressions. This gives us a larger sample size to more
accurately determine if NFE improves upon the 1-NN error rate NIST 0 97.59 3 97.97
2002/1 SA 0 99.02 0 99.12
when matching, and if NFE improves matching for the most FJ+SA 0 97.46 0 99.51
challenging of fingerprints. We make sure to use a different seed
NIST 1 99.41 0 99.18
for the test dataset than used in all the training set generations 2002/2 SA 0 99.88 0 99.23
so that the network is not attempting to classifying the “same” FJ+SA 0 99.84 0 99.12
finger. NIST 1 96.56 1 96.75
2002/3 SA 1 97.35 0 99.23
IV. R ESULTS FJ+SA 1 96.24 0 97.92

Training our NFE took only six epochs on a Nvidia Titan X NIST 3 95.53 3 92.88
2004/1 SA 0 96.60 0 96.93
GPU. Only six epochs where used due to a lack of compute FJ+SA 0 97.01 0 96.11
resources and time. A batch size of 64 used all 12 GB of
NIST 9 90.78 4 92.43
RAM, and took 17 hours to train. Reading in a fingerprint, 2004/2 SA 3 90.85 0 95.33
applying NFE, and saving it back to disk as PNG took an FJ+SA 7 88.83 1 96.04
average of 0.618 seconds per image. This time could be reduced NIST 4 92.43 0 99.22
significantly by applying weight quantization [12] and pruning 2004/3 SA 0 95.23 2 95.28
FJ+SA 1 96.04 0 98.83
[17], and integrating matching tools so that the output doesn’t
need to be written back to a slower HDD2 . However these
improvements are beyond our current scope. Instead we will To answer the overall question, “Does NFE improve the AUC
compare the performance of our three minutiae extraction and of fingerprint matching?”, we run a Wilcoxon signed rank test
matching algorithms on both real and generated test sets. [35]. The Wilcoxon test is preferable to the more common
t-test and Friedman test to compare the relative performance
A. Real Fingerprints of differing algorithms[1, 7]. Running the Wilcoxon signed
For our evaluation on real fingerprint images, we use the nine rank test to compare if the NFE AUC is greater than results
databases as discussed in subsection III-D. For each database without NFE gives us a p-value of 0.020492, a statistically
all three minutiae extraction and matching algorithms, NIST, significant result demonstrating that our new approach is an
SA, and FJ+SA, were run with LOO cross-validation twice — improvement.
first on the original fingerprint images, and then with our Neural Because of the small population size in these test sets, and
Fingerprint Enhancement. Our primary goal is to see that NFE generally high accuracy of current matching methods, focusing
improves the matching accuracy for all of these algorithms as on just the 1-NN results in a number of ties. Some of these ties
a pre-processing step. All of these results for each algorithm, occur with no 1-NN errors, but still have significantly differing
dataset, and approach can be found int Table II. The third AUC scores. This is because the AUC includes the relative
and fourth columns show 1-NN errors and AUC without any ranking of all seven enrolled fingerprints that belong to the
query print. For example, consider a person ζ with query print
2 The HDD overhead causes GPU utilization to average at only 40% ζq , and two enrolled prints ζ1 and ζ2 . If ζ1 is returned as the

4
closest fingerprint to ζq and ζ2 is deemed the farthest (i.e., (which affects ridge thickness) could be further increased. In
least similar) fingerprint to ζq , this will result in a lower AUC addition, we remind the reader that all of these databases
score. Ideally we would see ζ1 returned as the closest, with ζ2 contain properties that differ from SFinGe’s generation process.
returned as the second closest. The FVC2004/3 database in particular was collected using an
Looking at individual tests, we see that NFE provides the entirely different kind of sensor than what SFinGe models and
largest gains in matching performance on the most difficult is at a differing DPI, but NFE still improved the AUC for
corpora. In particular, FVC2000/1 was the most difficult all three algorithms and reduced the results of two extracttors
database for all algorithms, with up to 11 errors. NFE reduced down to zero errors.
the number of errors down to a maximum of 1 error and The overall conclusion is that NFE does improve fingerprint
the AUC improved by up to 15.5 whole points. The test on matching performance in the majority of cases, and could be
FVC2004/2 saw similar uniform improvements in performance. further improved by a more effective synthetic distribution
Most other datasets had low enough error rates that changes generation process. If NFE were deployed, users must ensure
are not necessarily meaningful or resulted in ties, but NFE still that the devices used to capture fingerprints are sufficiently
tended to improve AUC. similar to SFinGe’s output before use, or improve the generation
process to take into account the environmental differences.

B. Simulated Fingerprints
We now look at evaluating NFE’s impact on a larger test
corpus of low quality fingerprints. As described in subsec-
tion III-E we generate a test set of 100,000 “Very low” quality
fingerprints to enroll, and a query set of 1,000 fingerprint
images. This allows us to better quantify NFE’s performance.
The results are shown in Table III.

TABLE III: Results on test-set generated with SFinGen “Very

(a) Original print (b) NFE enhanced print low” quality. Second and third column show results without
Fig. 3: Example of NFE failure case on the FVC2000/3 NFE, last two show with NFE. Best results shown in bold.
database. The database contains a number of properties beyond
Standard NFE
what SFinGe generates, causing some errors to occur in the
Algorithm Errors AUC Errors AUC
enhanced print.
NIST 656 91.48 255 97.81
The notable failure of NFE was on the FVC2000/3 database, SA 721 81.91 165 96.30
FJ+SA 826 78.79 173 96.05
where NFE had a non-trivial increase in error rates. Manual
inspection of these results indicates the failure appears to be
caused by the accumulation of features and properties not In each case, a marked and dramatic improvement of every
captured by our version of SFinGe. An example is shown in metric occurs for every algorithm. The largest impact occurs to
Figure 3. In the corners of the image we can see that NFE the FJ+SA algorithm, reducing 1-NN errors by a factor of 4.8
creates some spurious ridges due to the circular receptive field and increasing the AUC by 17.26 whole points. We note that
of the fingerprint reader. We can also see spurious valleys these improvements in accuracy are on similar scales to those
be formed within some ridges. This appears to be caused by seen on real fingerprints in subsection IV-A on FVC2000/3
this dataset containing fingerprint ridges that are generally and FVC2004/2. This seems to indicate that NFE dramatically
thicker than what SFinGe produces. Last, we see a false ridge improves matching accuracy on the hardest of fingerprints.
generated at the bottom of the enhanced fingerprint. In the An example of one of these low quality fingerprints is given
original image there is a gap that appears to be caused by the in Figure 4, where we show the print, NFE’s enhancement, and
natural gap between the distal phalanx and middle phalanx. the ground truth output. This demonstrates that NFE reduces
NFE has partially filled in this gap, likely because SFinGe does the complexities and ambiguities that other algorithms and
not generate impressions corresponding to the middle phalanx. software would need to tackle for the majority of the print.
Despite these distributional differences between our gener- Comparing the result with the ground truth, the majority of
ated print and the FVC2000/3 database, it’s performance is minutiae are correctly formed. Most exceptions occur near the
not completely destroyed and the AUC was even improved edge of the fingerprint, or in the particularly low-quality area
for the NIST extractor/matcher. We also believe all of these near the delta of the fingerprint.
distributional differences are resolvable. Additional image While the AUC indicates uniform improvement of NFE on
backgrounds could be created from real fingerprint readers this corpus, fingerprint matching must often take into account
to add to SFinGe’s background generation process, the middle differing false positive rates. As such we show the ROC curve
phalanx could be simulated using the same kind of process used of each method with and without NFE in Figure 5. Here we
to generate the distal print, and the maximum pressure/dryness can see that not only does each algorithm’s ROC curve with

5
 R EFERENCES
[1] A. Benavoli, G. Corani, and F. Mangili. Should We
Really Use Post-Hoc Tests Based on Mean-Ranks?
Journal of Machine Learning Research, 17(5):1–10,
2016.
[2] R. Cappelli. SFinGe. In Encyclopedia of Biometrics,
pages 1–9. Springer US, Boston, MA, 2014.
[3] R. Cappelli, D. Maio, A. Lumini, and D. Maltoni.
Fingerprint image reconstruction from standard tem-
(a) Generated print (b) NFE enhanced print (c) SFinGe ground truth plates. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 29(9):1489–1503, 2007.
Fig. 4: Example of NFE success on SFinGe “very low” quality [4] R. Cappelli. Synthetic fingerprint generation. Handbook
test set. Left most image shows the generated fingerprint, of Fingerprint Recognition:203–232, 2003.
middle the result after NFE, and right most the target output. [5] S. Chikkerur, A. N. Cartwright, and V. Govindaraju.
Fingerprint enhancement using STFT analysis. Pattern
Recognition, 40(1):198–211, Jan. 2007. ISSN: 0031-
100 3203.
[6] L. N. Darlow and B. Rosman. Fingerprint minutiae ex-
traction using deep learning. In 2017 IEEE International
True Positive Rate

Joint Conference on Biometrics (IJCB), pages 22–30.

IEEE, Oct. 2017. ISBN: 978-1-5386-1124-1.
10−1 [7] J. Demšar. Statistical Comparisons of Classifiers over
NIST Multiple Data Sets. Journal of Machine Learning Re-
SA search, 7:1–30, Dec. 2006. ISSN: 1532-4435.
FJ+SA [8] FingerJetFX OSE, 2011. URL: https : / / github . com /
NFE NIST FingerJetFXOSE/FingerJetFXOSE.
10−2 NFE SA [9] S. Ioffe and C. Szegedy. Batch Normalization: Accel-
NFE FJ+SA erating Deep Network Training by Reducing Internal
Covariate Shift. In Proceedings of The 32nd Interna-
10−3 10−2 10−1 100 tional Conference on Machine Learning, volume 37,
False Positive Rate pages 448–456, 2015.
Fig. 5: Receiver operating characteristic curve for all three [10] ISO/IEC 19794-2:2011. Technical report, International
algorithms with and without NFE on the “very low” quality Organization for Standardization, 2011, page 93.
SFinGe test set. Both axes shown on log scale. [11] A. Jain, L. Hong, and R. Bolle. On-line fingerprint
verification. IEEE transactions on pattern analysis and
machine intelligence, 19(4):302–314, 1997.
NFE dominate the one without, all of the NFE augmented [12] D. Lin, S. Talathi, and S. Annapureddy. Fixed Point
algorithms dominate all non-NFE curves. Quantization of Deep Convolutional Networks. In M. F.
Balcan and K. Q. Weinberger, editors, Proceedings of
V. C ONCLUSION The 33rd International Conference on Machine Learn-
By using synthetic training data, we have demonstrated ing, volume 48 of Proceedings of Machine Learning
that it is possible to learn a fingerprint enhancement pre- Research, pages 2849–2858, New York, New York, USA.
processor without the complex pre-processing steps that have PMLR, 2016.
been used in prior works. Our new neural network fingerprint [13] Lu Jiang, Tong Zhao, Chaochao Bai, A. Yong, and Min
enhancement improves the AUC for multiple different minutiae Wu. A direct fingerprint minutiae extraction approach
extraction/matching algorithms on real fingerprints, and testing based on convolutional neural networks. In 2016 Inter-
on additional generated data shows that NFE can significantly national Joint Conference on Neural Networks (IJCNN),
improve results on the hardest of fingerprint images. We believe pages 571–578. IEEE, July 2016. ISBN: 978-1-5090-
our work shows the potential for further development of neural 0620-5.
networks that could simplify fingerprint processing systems [14] D. Maio, D. Maltoni, R. Cappelli, J. Wayman, and
today. In future work, we hope to re-visit some of the other A. Jain. FVC2000: fingerprint verification competition.
tasks which have previously designed neural networks with IEEE Transactions on Pattern Analysis and Machine
significant domain knowledge Intelligence, 24(3):402–412, Mar. 2002. ISSN: 01628828.
[15] D. Maio, D. Maltoni, R. Cappelli, J. L. Wayman, and
A. K. Jain. FVC2002: Second Fingerprint Verification

6
 Competition. Proceedings of the 16 th International FFT based features using MLP. Neural Computing &
Conference on Pattern Recognition (ICPR’02) Volume 3 Applications, 7(2):180–191, 1998. ISSN: 0941-0643.
- Volume 3. ICPR ’02, 24(3):402–412, 2002. [28] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W.
[16] D. Maio, D. Maltoni, R. Cappelli, J. L. Wayman, and Wang, and R. Webb. Learning from Simulated and
A. K. Jain. FVC2004: Third fingerprint verification Unsupervised Images through Adversarial Training. In
competition. In Biometric Authentication, pages 1–7. 2017 IEEE Conference on Computer Vision and Pattern
Springer Berlin Heidelberg, 2004. Recognition (CVPR), pages 2242–2251. IEEE, July 2017.
[17] D. Molchanov, A. Ashukha, and D. Vetrov. Variational ISBN : 978-1-5386-0457-1.
Dropout Sparsifies Deep Neural Networks. In Interna- [29] K. Sundararajan and D. L. Woodard. Deep Learning for
tional Conference on Machine Learning (ICML), 2017. Biometrics: A Survey. ACM Comput. Surv., 51(3):65:1–
[18] V. Nair and G. E. Hinton. Rectified Linear Units Improve 65:34, May 2018. ISSN: 0360-0300.
Restricted Boltzmann Machines. Proceedings of the 27th [30] E. Tabassi. Development of NFIQ 2.0. Technical report,
International Conference on Machine Learning:807–814, National Institute of Standards and Technology, 2012.
2010. [31] J. Tompson, K. Schlachter, P. Sprechmann, and K.
[19] D.-L. Nguyen, K. Cao, and A. K. Jain. Robust Minutiae Perlin. Accelerating Eulerian Fluid Simulation With
Extractor: Integrating Deep Networks and Fingerprint Convolutional Networks. In D. Precup and Y. W. Teh,
Domain Knowledge. In The 11th International Confer- editors, Proceedings of the 34th International Conference
ence on Biometrics, 2018, 2018. on Machine Learning, volume 70 of Proceedings of
[20] R. F. Nogueira, R. D. A. Lotufo, and R. C. Machado. Machine Learning Research, pages 3424–3433, Inter-
Evaluating software-based fingerprint liveness detection national Convention Centre, Sydney, Australia. PMLR,
using Convolutional Networks and Local Binary Patterns. 2017.
In Proceedings of the IEEE Workshop on Biometric [32] R. Važan. SourceAFIS, 2018. URL: https://sourceafis.
Measurements and Systems for Security and Medical machinezoo.com/.
Applications (BIOMS), pages 22–29, Rome, Italy, 2014. [33] C. Wang, K. Li, Z. Wu, and Q. Zhao. A DCNN Based
ISBN : 9781479951765. Fingerprint Liveness Detection Algorithm with Voting
[21] M. A. Olsen, E. Tabassi, A. Makarov, and C. Busch. Strategy. In J. Yang, J. Yang, Z. Sun, S. Shan, W. Zheng,
Self-Organizing Maps for Fingerprint Image Quality and J. Feng, editors, Biometric Recognition, pages 241–
Assessment. 2013 IEEE Conference on Computer Vision 249. 2015. ISBN: 978-3-319-25417-3.
and Pattern Recognition Workshops:138–145, 2013. [34] C. I. Watson, M. D. Garris, E. Tabassi, C. L. Wilson,
ISSN : 21607508. R. M. McCabe, S. Janet, and K. Ki. User’s Guide
[22] D. Peralta, M. Galar, I. Triguero, D. Paternain, S. to NIST Biometric Image Software (NBIS). Technical
García, E. Barrenechea, J. M. Benítez, H. Bustince, report, National Institute of Standards and Technology,
and F. Herrera. A survey on fingerprint minutiae- 2007.
based local matching for verification and identification: [35] F. Wilcoxon. Individual Comparisons by Ranking Meth-
Taxonomy and experimental evaluation. Information ods. Biometrics Bulletin, 1(6):80, Dec. 1945. ISSN:
Sciences, 315:67–87, Sept. 2015. ISSN: 0020-0255. 00994987.
[23] N. K. Ratha, S. Chen, and A. K. Jain. Adaptive [36] L. Yong-xia, Q. Jin, and X. Rui. A new detection
flow orientation-based feature extraction in fingerprint method of singular points of fingerprints based on
images. Pattern Recognition, 28(11):1657–1672, Nov. neural network. In Computer Science and Information
1995. ISSN: 00313203. Technology (ICCSIT), 2010 3rd IEEE International
[24] N. Ratha, K. Karu, Shaoyun Chen, and A. Jain. A real- Conference on, volume 1, pages 301–305, July 2010.
time matching system for large fingerprint databases. [37] F. Yu and V. Koltun. Multi-Scale Context Aggregation
IEEE Transactions on Pattern Analysis and Machine by Dilated Convolutions. In International Conference
Intelligence, 18(8):799–813, 1996. ISSN: 01628828. on Learning Representations, 2016.
[25] A. Ross, A. Jain, and J. Reisman. A hybrid fingerprint [38] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus.
matcher. Pattern Recognition, 36(7):1661–1673, 2003. Deconvolutional networks. In Computer Vision and
ISSN : 00313203. Pattern Recognition (CVPR), 2010 IEEE Conference
[26] M. Sahasrabudhe and A. M. Namboodiri. Fingerprint on, pages 2528–2535. IEEE, 2010.
Enhancement Using Unsupervised Hierarchical Feature
Learning. In Proceedings of the 2014 Indian Conference
on Computer Vision Graphics and Image Processing,
ICVGIP ’14, 2:1–2:8, New York, NY, USA. ACM, 2014.
ISBN : 978-1-4503-3061-9.
[27] S. N. Sarbadhikari, J. Basak, S. K. Pal, and M. K.
Kundu. Noisy fingerprints classification with directional