Fingerprint Large Classification Using Sequential Learning on Parallel
Environment
Nicol
´
as A. Reyes-Reyes
1 a
, Marcela C. Gonz
´
alez-Araya
2 b
and Wladimir E. Soto-Silva
3 c
1
Programa de Doctorado en Sistemas de Ingenier
´
ıa, Facultad de Ingenier
´
ıa, Universidad de Talca, Campus Curic
´
o,
Camino a Los Niches km 1, Curic
´
o, Chile
2
Departamento de Ingenier
´
ıa Industrial, Facultad de Ingenier
´
ıa, Universidad de Talca, Campus Curic
´
o, Camino a Los
Niches km 1, Curic
´
o, Chile
3
Departamento de Computaci
´
on e Industrias, Facultad de Ciencias de la Ingenier
´
ıa, Universidad Cat
´
olica del Maule,
Avenida San Miguel 3605, Talca, Chile
Keywords:
Fingerprint, Large Classification, Sequential Learning, Extreme Learning Machine, Graphics Processing Unit.
Abstract:
Fingerprint classification allows a biometric identification system to reduce search space in databases and
therefore response times. In the literature, fingerprint classification has been addressed through different
approaches where deep learning techniques such as convolutional neural networks have been gaining attention.
However, the proposed approaches use extremely small data sets for large-scale real-world scenarios that
could worsen accuracy rates due to interclass and intraclass variations in fingerprints. For this reason, we
proposed a fingerprint classification approach that allows us to address this problem by considering millions
of samples. For this purpose, a classifier based on neural networks trained using online sequential extreme
learning machines was developed. Likewise, to accelerate the training of the classifier, the matrix operations
inside it was run in a graphic processing unit. In order to evaluate our proposal, the approach was tested on
three datasets with more than two million synthetic fingerprint image descriptors. The results are similar in
terms of accuracy and computational time to recent approaches but using more than 2.5 million samples.
1 INTRODUCTION
One of the most widely used biometric techniques for
identifying people today corresponds to the finger-
print. This is because it contains the necessary infor-
mation about unique characteristics of a person (Zia
et al., 2019). The fingerprint has applications in many
areas, and is mainly used in security and control sys-
tems such as entry to mass events, police control, na-
tional registration of people, assistance in companies,
etc (Galar et al., 2015a; Zabala-Blanco et al., 2020b).
In a fingerprint recognition system, to determine
a person’s identity, their fingerprint is searched in
a database until a match is found. However, this
is quite computationally expensive if the database
is large. For this reason, a fingerprint classifica-
tion is commonly performed, reducing the domain
or search space in the database. Fingerprint images
have structural characteristics based on the pattern
a
https://orcid.org/0009-0002-1100-6187
b
https://orcid.org/0000-0002-4969-2939
c
https://orcid.org/0000-0002-2203-9366
of the ridges, and according to their morphological
structure they are usually classified into five classes
(known as Henry’s classification), which are the fol-
lowing: Arch, Left Loop, Right Loop, Tented Arch
and Whorl. Furthermore, these classes present a high
level of imbalance between them, like other human
biological characteristics (Peralta et al., 2018; Saeed
et al., 2018; Zabala-Blanco et al., 2020b).
Several approaches have been presented in the lit-
erature to solve this challenging fingerprint classifica-
tion problem. For instance, Rajanna et al. (2010); Cao
et al. (2013) and Luo et al. (2014) proposed classifiers
based on k nearest neighbors (kNN) and support vec-
tor machines (SVM), and assembled as hierarchical
classifiers. Furthermore, to improve the performance
of the classifiers in terms of accuracy, the authors used
feature extractors such as orientation maps, curvelet
transform along with gray level co-occurrence matri-
ces, among others. Rule-based classifiers have also
been applied by Liu (2010); Guo et al. (2014); Galar
et al. (2014) and Dorasamy et al. (2015) as deci-
sion trees (DT), and these in combination with adap-
222
Reyes-Reyes, N., González-Araya, M. and Soto-Silva, W.
Fingerprint Large Classification Using Sequential Learning on Parallel Environment.
DOI: 10.5220/0012355300003636
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Conference on Agents and Artificial Intelligence (ICAART 2024) - Volume 2, pages 222-230
ISBN: 978-989-758-680-4; ISSN: 2184-433X
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

tive boosting (AdaBoost). Likewise, to achieve bet-
ter classification results, feature extractors based on
singular points, orientation maps and directional pat-
terns were used. On the other hand, one of the most
used techniques to classify fingerprints is SVM. Thus,
studies presented by Cao et al. (2013); Saini et al.
(2013); Galar et al. (2015b); Gupta and Gupta (2015);
Huang et al. (2015) and Alias and Radzi (2016), pre-
sented SVMs as single classifiers and in combination
with other techniques such as probabilistic neural net-
works (PNN), and also as part of a hierarchical clas-
sifier. In this case, the authors improved the perfor-
mance of the classifiers by using feature descriptors
based on orientation fields, designs based on Wavelet
transforms, singular points, minutiae extraction, and
Fourier spectrum features. Also, classifiers have been
proposed based on adaptive genetic neural networks
(AGNN) by Borra et al. (2018), on naive-bayes by
Vitello et al. (2014), and on conditional probability
by Jung and Lee (2015).
In recent years, deep learning (DL) techniques
have focused the attention of researchers for finger-
print classification. Hence, multiples convolutional
neural networks (CNN) have been proposed by Wang
et al. (2016); Michelsanti et al. (2017); Ge et al.
(2017); Shrein (2017); Peralta et al. (2018); El Hamdi
et al. (2018), and Zia et al. (2019), considering differ-
ent amounts of convolutional layers, or auto-encoder
(AE) layers, and even pretrained networks like VGG-
F and VGG-S. In this case, CNNs calculate feature
descriptors through internal mechanisms, and other
techniques were not required by the authors. How-
ever, CNNs require large datasets to achieve high ac-
curacy and high computational effort to train due to
their complexity. For this reason, neural networks
trained using extreme learning machines (ELM) are
an alternative to CNNs since they can achieve simi-
lar accuracy rates and their training is extremely fast
(Huang et al., 2006). Thus, ELM neural networks
have been proposed as a classifier in human recogni-
tion from images by several studies (An et al., 2015;
Li et al., 2015; Zhang et al., 2018; Deng et al., 2019;
Lu et al., 2019).Specifically, it has been used for fin-
gerprint classification by Saeed et al. (2018) in its ba-
sic form with radial activation function, considering
the extraction of features in fingerprint images using
orientation fields with histograms of oriented gradient
(HOG). Also, Zabala-Blanco et al. (2020b) proposed
classifiers based on these neural networks in their ba-
sic and weighted form to treat unbalanced classes.
These authors used different feature extractors such
as Hong08, Liu10, and Capelli02, which are the most
common extractors currently for fingerprint classifi-
cation.
Although the approaches presented in the litera-
ture have achieved high accuracy rates (even above
0.95) and training times less than 5,000 seconds
(mostly), the datasets used in them are extremely
small. Furthermore, almost all of them resort to
publicly accessible datasets such as NIST-DB4 OR
FVC2000, 2002, 2004 that do not exceed 4,000 fin-
gerprint samples. Sometimes the SFinGe software
was used to generate synthetic fingerprints, but these
do not exceed 30,000 samples either. These condi-
tions are limiting for the classification of fingerprints
in biometric systems intended for large populations
of people because the fingerprints that best represent
each class must be chosen before performing the clas-
sification. Therefore, this choice is subject to human
error even though it is carried out by experts due to
inter-class similarities and intra-class differences in
fingerprints (Zia et al., 2019; Zabala-Blanco et al.,
2020b). For this reason, in this study we propose an
approach for large fingerprint classification in reason-
able computational times for practical scenarios. For
this purpose, a classifier based on neural networks
trained using the online sequential extreme learning
machine (OS-ELM) is developed. Furthermore, train-
ing of the classifier is accelerated by running its ma-
trix operations to a graphics processing unit (GPU).
Additionally, specialized feature extractors are used
to obtain input data from synthetic fingerprint images.
To validate the proposal, this approach is applied on
three data sets with millions of fingerprint descriptors.
The rest of this document is organized as follows.
Section 2 describes the proposed approach for large
fingerprint classification using neural networks and
sequential learning. Section 3 reports and discusses
the results of computational experiments carried out
on three fingerprint descriptor datasets. Finally, the
4 section presents the main conclusions of this study
and future work.
2 PROPOSED APPROACH
The proposed approach for large fingerprint classifi-
cation is summarized in Figure 1.
Each stage of the approach is briefly described in
the following subsections.
2.1 Large Datasets Generation
Three feature descriptor datasets are generated from
synthetic fingerprint images. To generate fingerprint
images with characteristics and distribution that sim-
ulate real fingerprints, the SFinGe v4.0 software is
used. The Capelli02, Hong08 and Liu10 extractors
Fingerprint Large Classification Using Sequential Learning on Parallel Environment
223

Figure 1: Flow of the proposed approach for fingerprint
classification.
Table 1: Distribution of fingerprint classes in each dataset.
Fingerprint class
No. of samples
Arch 660,573
Left Loop 5,947,535
Right Loop 5,599,876
Tented Arch 515,757
Whorl 4,973,608
are used to calculate the feature descriptors on the
generated syntactic fingerprint images. These extrac-
tors are used since the descriptors they generate have
a high quality and capacity to represent the most im-
portant visual characteristics of a fingerprint image
(Zabala-Blanco et al., 2020b). Each dataset has a to-
tal of 17,697,349 fingerprint descriptors, and each de-
scriptor is labeled. The classes considered are: Arch,
Left Loop, Right Loop, Tented Arch and Whorl, as
shown in the following Figure 2.
The distribution of these five classes of finger-
prints is not uniform in the three datasets, representing
the following percentages of the total samples: Arch
(3.73 %), Left Loop (33.61 %), Right Loop (31.64
%), Tented Arch (2.91 %) and Whorl (28.10 %) (as
shown in Table 1).
These datasets are used considering balanced
classes to achieve better classification performance.
This balancing is carried out through random subsam-
pling.
2.2 Classification Using Sequential
Learning
The extreme learning machine (ELM) neural network
is a feed-forward network with a single hidden layer
(SLFN), with its respective input and output layer.
This neural network is classified as a network with
random weights. The training of this network consists
of randomly assigning the weights and biases of the
hidden layer, and analytically computing the weights
between the hidden layer and the output layer (Huang
et al., 2006, 2015; Cao et al., 2018).
In Huang et al. (2006), this ELM neural network
is formally defined as follows:
Given N different arbitrary samples (x
i
, t
i
), where
x
i
= [x
i1
, x
i2
, ..., x
in
]
T
∈ R
n
and t
i
= [t
i1
, t
i2
, ..., t
im
]
T
∈
R
m
, the standard single hidden layer feed-forward
neural network with L neurons and activation func-
tion g(x) is mathematically modeled as:
L
∑
i=1
β
i
g
i
(x
j
) =
L
∑
i=1
β
i
g(w
i
∗ x
j
+ b
i
) = o
j
, j = 1, ..., N
(1)
where w
i
= [w
i1
, w
i2
, ..., w
in
]
T
is the vector of weights
connecting the i-th hidden neuron and the input neu-
rons, β
i
= [β
i1
, β
i2
, ..., β
im
]
T
is the vector of weights
connecting the i-th hidden neuron and the output neu-
rons, and b
i
is the bias of the i-th hidden neuron.
w
i
∗ x
j
denotes the inner product of w
i
and x
j
. The
output neurons are chosen linear.
The standard single-layer feed-forward neural net-
work hidden with L neurons and with activation func-
tion g(x) can approximate these N samples with zero
mean error
∑
L
j=1


o
j
− t
j


= 0, that is, they exist
β
i
, w
i
and b
i
such that:
L
∑
i=1
β
i
g(w
i
∗ x
j
+ b
i
) = t
j
, j = 1, ..., N (2)
The N equations above in (2) can be written in
compact form as:
Hβ = T (3)
where the matrix H is:
H(w
1
, ..., w
L
, b
1
, ..., b
L
, x
1
, ..., x
N
) =



g(w
1
∗ x
1
+ b
1
) ·· · g(w
L
∗ x
1
+ b
L
)
.
.
. ·· ·
.
.
.
g(w
1
∗ x
N
+ b
1
) ·· · g(w
L
∗ x
N
+ b
L
)



NxL
,
(4)
the vector of weights between hidden layer and output
layer:
β =



β
T
1
.
.
.
β
T
L



Lxm
, (5)
and the vector of labels or targets of the N samples is:
T =



t
T
1
.
.
.
t
T
N



Nxm
(6)
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
224

Figure 2: Fingerprint classes.
The matrix H, is the output matrix of the hidden
layer of the neural network. The i-th column of H is
the output of the i-th hidden neuron with respect to
the inputs x
1
, x
2
, ..., x
N
.
Considering that both H as T are known, the
weight vector between the hidden layer and the output
layer β, is determined by:
β = H
†
T (7)
where H
†
is the Moore-Penrose inverse or also known
as the pseudoinverse of the matrix H (Greville, 1959).
This generalized inverse can be obtained by means of
the following expression.
H
†
= (H
T
∗ H)
−1
∗ H
T
(8)
In addition to equation (8) (known as orthogonal
projection), there are different algorithms that allow
calculating the generalized inverse of a matrix, among
these are algorithms based on singular value decom-
position (SVD), extended orthogonal projection (with
regularization parameter) (EOP), Cholesky factoriza-
tion, among others (Golub and Kahan, 1965; Cour-
rieu, 2008; Lu et al., 2015). Specifically, the EOP al-
gorithm performs a fast and effective calculation of
the generalized inverse, achieving a high level of pre-
cision (Lu et al., 2015). For this reason, this algorithm
is useful in training an ELM neural network. The EOP
is defined as the equation (9) below:
H
†
=

(H
T
∗ H +
I
C
)
−1
∗ H
T
, i f N > L
H
T
∗ (H ∗ H
T
+
I
C
)
−1
, otherwise
(9)
where I and C in (9), correspond to an identity matrix
and a regularization parameter respectively.
Several extensions of the basic ELM have been
developed, with features to solve different types of
problems and/or scenarios (Huang et al., 2015). One
of these extensions is the online sequential extreme
learning machine, which is described below.
2.2.1 Sequential Learning
The online sequential extreme learning machine (OS-
ELM) is a feed-forward neural network training algo-
rithm, based on the standard ELM and the recursive
least squares (RLS) algorithm. Where the main dif-
ference with the standard, is that the OS-ELM does
not need to process all the data at once, since it can
perform the training by processing smaller batches of
data and update the output weights of the hidden layer
in the neural network iteratively. In OS-ELM, the
data batches to be processed can maintain a fixed size
(equal number of samples per batch) during training,
and can even vary in size, making the training process
even more flexible (Liang et al., 2006).
The OS-ELM achieves excellent generalizability
in a feed-forward neural network, and in extremely
short times, thus preserving the fundamental proper-
ties of the standard ELM (Liang et al., 2006; Huang
et al., 2015).
Algorithm 1 describes the procedure associated
with the OS-ELM:
To strengthen the calculation of the inverse matrix
in the equation (10) (in Algorithm 1), the calculation
of a usual inverse can be replaced by a generalized or
pseudoinverse (equation 9). This allows to avoid the
distortion produced in the inverse of a matrix, due to
the proximity of the original matrix to a singularity
condition (Liang et al., 2006; Huang et al., 2015).
2.3 Performance Evaluation
To evaluate performance, a simple cross-validation
scheme is applied. In this simple cross-validation
scheme, each original dataset is divided into two sub-
sets, one to perform the training and the other to test
the trained classifier and verify its performance, with
new data. In some cases, a third subset is consid-
ered to find good values for the hyperparameters of
a classifier (Raschka, 2018). Therefore, each dataset
is finally divided in a 80:20 ratio, obtaining a sub-
set for training-validation (with 80% of the total sam-
ples) and a subset for testing (with 20% of the total
samples), as performed by Peralta et al. (2018) and
Zabala-Blanco et al. (2020b).
Classification performance is measured through
accuracy (ACC) for all the experiments associated
with the three datasets. These performance measures
Fingerprint Large Classification Using Sequential Learning on Parallel Environment
225

Require: Initial training data block £
0
=
{
(x
i
, t
i
)
}
N
0
i=1
of a given training set
£ =
{
(x
i
, t
i
)|x
i
∈ R
n
, t
i
∈ R
m
, i = 1, ...
}
, an activation function g(x) and a number of neurons L, where
N
0
> L.
1. Initialization Phase:
Step 1: Randomly assign input weights and biases w
i
, b
i
, i = 1, ..., L.
Step 2: Calculate initial output matrix of hidden layer H
0
, where
H
0
=



g(w
1
∗ x
1
+ b
1
) ··· g(w
L
∗ x
1
+ b
L
)
.
.
. ·· ·
.
.
.
g(w
1
∗ x
N
0
+ b
1
) ·· · g(w
L
∗ x
N
0
+ b
L
)



N
0
xL
Step 3: Estimate the initial output weights β
0
= P
0
∗ H
T
0
∗ T
0
where P
0
= (H
T
0
∗ H
0
)
−1
and
T
0
= [t
1
, ..., t
N
0
]
T
.
Step 4: Set k = 0.
2. Sequential Learning Phase:
Present the (k + 1)-th batch of data
£
k+1
=
{
(x
i
, t
i
)
}
∑
k+1
j=0
N
j
i=(
∑
k
j=0
N
j
)+1
where N
k+1
denotes the number of samples in the batch k + 1.
Step 1: Calculate partial output matrix of hidden layer H
k+1
for batch data £
k+1
.
H
k+1
=




g(w
1
∗ x
(
∑
k
j=0
N
j
)+1
+ b
1
) ·· · g(w
L
∗ x
(
∑
k
j=0
N
j
)+1
+ b
L
)
.
.
. ·· ·
.
.
.
g(w
1
∗ x
∑
k+1
j=0
N
j
+ b
1
) ·· · g(w
L
∗ x
∑
k+1
j=0
N
j
+ b
L
)




N
k+1
xL
Step 2: Set T
k+1
=

t
(
∑
k
j=0
N
j
)+1
, ..., t
∑
k+1
j=0
N
j

T
.
Step 3: Calculate the output weights β
k+1
P
k+1
= P
k
− P
k
∗ H
T
k+1
∗ (I + H
k+1
∗ P
k
∗ H
T
k+1
)
−1
∗ H
k+1
∗ P
k
(10)
β
(k+1)
= β
k
+ P
k+1
∗ H
T
k+1
∗ (T
k+1
− H
k+1
∗ β
(k))
) (11)
Step 4: Set k = k + 1. Go to 2.
Algorithm 1: Online Sequential Extreme Learning Machine.
are calculated using the following equation (12).
ACC =
T P + T N
T P + FN + T N + FP
(12)
where the values of true positive (TP), true nega-
tive (TN), false positive (FP) and false negative (FN).
Equation (12) allows calculating the classification ac-
curacy considering the number of total hits on the
number of total samples. Although these equation
(12) is defined for binary classification, the exten-
sion to multi-class classification implies only in (12)
adding the correct answers for each class over the to-
tal samples of all classes.
To maximize ACC performance in the neural net-
work trained by the OS-ELM, a estimation of optimal
hyperparameter values is performed. The hyperpa-
rameters considered are the number of hidden neurons
(L) and C (regularization parameter). This estima-
tion consists of performing simple cross-validations
(training-validation), on a set of combinations of spe-
cific values for the two mentioned hyperparameters.
In the case of L, these values range from 500 to 5,000
neurons in steps of 500 neurons. On the other hand,
the value of the regularization parameter ranges from
10
−20
to 10
20
, in steps of 1 in the exponent. In this
way, a wide search space of 410 points is approached
for fingerprint classification problem (Zabala-Blanco
et al., 2020b). Furthermore, indicate that a smaller
portion of samples is used to estimate hyperparame-
ters of the classifier, since it is computationally ex-
pensive to use the total number of samples in this pro-
cess. For this estimation, a portion of samples will be
considered such as those used by Zabala-Blanco et al.
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
226

Table 2: Details of the server used in the experiments.
Item Description
CPU Intel
R
 Xeon
R
 Gold 6140
RAM 128GB
GPU Geforce GTX 1080ti
OS Debian 10 64 bit
(2020b) in base ELM neural networks. Thus, 30,000
samples are used in total: 18,000 for training, 6,000
for validation, and 6,000 for testing.
3 RESULTS
3.1 Software and Hardware
The OS-ELM algorithm was coded in C++ v11, us-
ing the CUDA v9.2, cuBLAS v9.2 and cuSOLVER
v9.2 libraries (from the NVIDIA platform) to acceler-
ate matrix operations such as multiplication and cal-
culation of inverses. It should be noted that the spe-
cific operations of the OS-ELM algorithm that were
brought to the GPU correspond to matrix multiplica-
tions, calculation of activation functions and calcula-
tion of pseudoinverse matrices. All experiments were
carried out on the server detailed in Table 2.
3.2 Experimental Results
It should be noted that in all the experiments carried
out, the weights W and biases b between the input
layer and the hidden layer of the neural network were
random values generated by a uniform distribution
in the intervals [-1 , 1] and [0,1] respectively Huang
et al. (2006). Likewise, the activation function used
in the hidden layer of the neural network corresponds
to a sigmoid function defined as: g(x) =
1
1+e
−x
, given
its universal approximation capability in a SLFN us-
ing algorithms based on ELM (Zabala-Blanco et al.,
2020a).
3.2.1 Hyperparameter Estimation
It can be seen in Figures 3, 4, and 5, that there is a
concentration of high ACC within the range 10
−10
to
10
10
in C, where the central values in this range reach
the ACC maximums. Likewise, there is a trend to-
wards better ACC when the number of hidden neu-
rons increases, especially in quantities greater than
3,000 neurons. However, in Figure 5 there are also
good values between 1,500 and 2,500 neurons. Ac-
cording to the above, it can be established that those
hyperparameter values that maximize ACC are within
the mentioned regions.
Figure 3: Accuracy for different hyperparameter values on
Hong08 descriptor dataset.
Figure 4: Accuracy for different hyperparameter values on
Liu10 descriptor dataset.
Figure 5: Accuracy for different hyperparameter values on
Capelli02 descriptor dataset.
The hyperparameter values that maximize ACC in
the neural network are presented in the following Ta-
ble 3.
The optimal hyperparameters indicated in Table 3,
were used in all experiments in Subsection 3.2.2.
Fingerprint Large Classification Using Sequential Learning on Parallel Environment
227

Table 3: Optimal hyperparameter and maximum ACC ob-
tained.
Dataset
Hyperparameter ACC
L C Validation Test
Hong08 4,500 10 0.92 0.92
Liu10 4,000 1 0.83 0.84
Capelli02 2,000 100 0.74 0.74
Table 4: Classification results on large fingerprint datasets.
Dataset
ACC Time (sec)
Train Test Train Test
Hong08 0.94 0.94 4,980.95 79.30
Liu10 0.83 0.83 4,655.93 64.38
Capelli02 0.75 0.75 2,638.78 47.14
3.2.2 Fingerprint Large Classification
An experiment was carried out with a large amount
of data to evaluate the performance of the neural net-
work trained by OS-ELM. This experiment consists
of a training and testing on datasets with balanced fin-
gerprint. For this purpose, 2,578,785 samples were
used for each dataset, where each one is divided into
two subsets, a training subset with 80% of the total
samples (2,063,025 samples) and a test subset with
20 % of the total samples (515,760 samples). Table 4
shows the results.
Table 4 shows that regardless of the data set
used, ACCs above 0.70 are achieved. However, it
is observed that Hong08 and Liu10 show superior-
ity in terms of ACC. On the other hand, when using
Capelli02 the computational times are reduced almost
by half compared to Hong08 and Liu10, mainly due
to the smaller size of the hidden layer in the neural
network. Although using Capelli02 requires less time
than using Hong08, with the latter it improves by al-
most 0.2 of ACC, showing superiority in fingerprint
classification.
In order to evaluate our proposal, the best results
achieved were compared with other studies in the lit-
erature. Table 5 presents some measures of interest
for comparative purposes.
It can be seen in Table 5 that our proposal is com-
petitive in terms of testing ACC. Because the differ-
ence in ACC is not greater than 0.05 compared to
other studies. Regarding training times, our proposal
shows notable superiority since it achieves a learn-
ing speed almost 21 times greater than the fastest
approach presented by Zabala-Blanco et al. (2020b).
The learning speed is given by the quotient: no. of
training samples/training time. Furthermore, it should
be noted that our proposal allows us to deal with fin-
gerprint classification for a large number of samples,
by processing almost 60 times more samples than the
largest scenarios presented in the literature. There-
fore, this approach is a suggestive alternative for prac-
tical application, since it is suitable to be introduced
into large-scale person identification systems such as
entry to mass events, entry into countries at airports,
police control, among others.
4 CONCLUSIONS
In this study, we developed an approach to deal for
fingerprint large classification problem. For this pur-
pose, a single hidden layer feed-forward neural net-
work trained by online sequential extreme learning
machine was used. This approach was applied on
three datasets with millions of descriptors (or sam-
ples) obtained from synthetic fingerprints. To achieve
high classification accuracy, specialized fingerprint
descriptors such as Hong08, Liu10 and Capelli02
were used. Regarding the results, our approach classi-
fied fingerprints with an accuracy of 0.94 when using
Hong08 descriptors. Furthermore, to achieve this ac-
curacy performance, less than 5,000 seconds of train-
ing time were required. This result is remarkable
because a classifier based on neural networks was
trained with millions of samples. Furthermore, this
proposal shows a significant improvement in learning
speed when compared to other approaches presented
in the literature. These improvements are suggestive,
since they allow us to address large-scale fingerprint
classification when introduced into biometric systems
intended for mass identification of people, which ac-
cording to the authors’ knowledge has not been ad-
dressed.
Finally, future work remains to address the iden-
tification of fingerprints with unbalanced classes, the
use of deep features to improve performance in terms
of accuracy, and extend the approach to process
raw fingerprint images creating a fully automatic ap-
proach.
REFERENCES
Alias, N. A. and Radzi, N. H. M. (2016). Fingerprint classi-
fication using support vector machine. 2016 Fifth ICT
International Student Project Conference (ICT-ISPC),
pages 105–108.
An, L., Yang, S., and Bhanu, B. (2015). Efficient smile de-
tection by extreme learning machine. Neurocomputing,
149:354–363.
Borra, S. R., Reddy, G. J., and Reddy, E. S. (2018). Clas-
sification of fingerprint images with the aid of morpho-
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
228

Table 5: Comparison with results achieved by other studies.
Approach
Total
samples
Testing
ACC
Training time
(sec)
Learning speed
(samples/sec)
Hong08 + OS-ELM (Our proposal) 2,578,785 0.94 4,980.95 414.18
Hong08 + W-ELM2 (Zabala-Blanco et al., 2020b) 30,000 0.94 880.00 18.18
Bayesian deep CNN (Zia et al., 2019) 3,300 0.96 4,393.00 0.38
Novel CNN (Peralta et al., 2018) 40,000 0.99 960.00 10.42
CaffeNet (Peralta et al., 2018) 40,000 0.99 2,306.00 4.34
Pretrained VGG-S (Michelsanti et al., 2017) 3,300 0.96 108,000.00 0.03
logical operation and agnn classifier. Applied computing
and informatics, 14(2):166–176.
Cao, K., Pang, L., Liang, J., and Tian, J. (2013). Finger-
print classification by a hierarchical classifier. Pattern
Recognition, 46(12):3186–3197.
Cao, W., Wang, X., Ming, Z., and Gao, J. (2018). A review
on neural networks with random weights. Neurocomput-
ing, 275:278–287.
Courrieu, P. (2008). Fast computation of moore-penrose
inverse matrices. arXiv preprint arXiv:0804.4809.
Deng, C., Han, Y., and Zhao, B. (2019). High-performance
visual tracking with extreme learning machine frame-
work. IEEE transactions on cybernetics, 50(6):2781–
2792.
Dorasamy, K., Webb, L., Tapamo, J., and Khanyile, N. P.
(2015). Fingerprint classification using a simplified rule-
set based on directional patterns and singularity features.
In 2015 International Conference on Biometrics (ICB),
pages 400–407. IEEE.
El Hamdi, D., Elouedi, I., Fathallah, A., Nguyen, M. K.,
and Hamouda, A. (2018). Fingerprint classification us-
ing conic radon transform and convolutional neural net-
works. In Advanced Concepts for Intelligent Vision
Systems: 19th International Conference, ACIVS 2018,
Poitiers, France, September 24–27, 2018, Proceedings
19, pages 402–413. Springer.
Galar, M., Derrac, J., Peralta, D., Triguero, I., Paternain,
D., Lopez-Molina, C., Garc
´
ıa, S., Ben
´
ıtez, J. M., Pagola,
M., Barrenechea, E., et al. (2015a). A survey of fin-
gerprint classification part i: Taxonomies on feature ex-
traction methods and learning models. Knowledge-based
systems, 81:76–97.
Galar, M., Derrac, J., Peralta, D., Triguero, I., Paternain, D.,
Lopez-Molina, C., Garc
´
ıa, S., Ben
´
ıtez, J. M., Pagola, M.,
Barrenechea, E., et al. (2015b). A survey of fingerprint
classification part ii: Experimental analysis and ensem-
ble proposal. Knowledge-Based Systems, 81:98–116.
Galar, M., Sanz, J., Pagola, M., Bustince, H., and Herrera,
F. (2014). A preliminary study on fingerprint classifi-
cation using fuzzy rule-based classification systems. In
2014 IEEE International Conference on Fuzzy Systems
(FUZZ-IEEE), pages 554–560. IEEE.
Ge, S., Bai, C., Liu, Y., Liu, Y., and Zhao, T. (2017). Deep
and discriminative feature learning for fingerprint clas-
sification. In 2017 3rd IEEE International Conference
on Computer and Communications (ICCC), pages 1942–
1946.
Golub, G. and Kahan, W. (1965). Calculating the singular
values and pseudo-inverse of a matrix. Journal of the
Society for Industrial and Applied Mathematics, Series
B: Numerical Analysis, 2(2):205–224.
Greville, T. (1959). The pseudoinverse of a rectangular or
singular matrix and its application to the solution of sys-
tems of linear equations. SIAM review, 1(1):38–43.
Guo, J.-M., Liu, Y.-F., Chang, J.-Y., and Lee, J.-D. (2014).
Fingerprint classification based on decision tree from
singular points and orientation field. Expert Systems with
Applications, 41(2):752–764.
Gupta, P. and Gupta, P. (2015). A robust singular point
detection algorithm. Applied Soft Computing, 29:411–
423.
Huang, G.-B., Zhu, Q.-Y., and Siew, C.-K. (2006). Extreme
learning machine: theory and applications. Neurocom-
puting, 70(1-3):489–501.
Huang, Q., Chang, S., Liu, C., Niu, B., Tang, M., and Zhou,
Z. (2015). An evaluation of fake fingerprint databases
utilizing svm classification. Pattern Recognition Letters,
60:1–7.
Jung, H.-W. and Lee, J.-H. (2015). Noisy and incom-
plete fingerprint classification using local ridge distribu-
tion models. Pattern recognition, 48(2):473–484.
Li, W., Chen, C., Su, H., and Du, Q. (2015). Local bi-
nary patterns and extreme learning machine for hyper-
spectral imagery classification. IEEE Transactions on
Geoscience and Remote Sensing, 53(7):3681–3693.
Liang, N.-Y., Huang, G.-B., Saratchandran, P., and Sun-
dararajan, N. (2006). A fast and accurate online sequen-
tial learning algorithm for feedforward networks. IEEE
Transactions on neural networks, 17(6):1411–1423.
Liu, M. (2010). Fingerprint classification based on adaboost
learning from singularity features. Pattern Recognition,
43(3):1062–1070.
Lu, L., Zhang, X., and Xu, X. (2019). Hypercomplex ex-
treme learning machine with its application in multispec-
tral palmprint recognition. PloS one, 14(4):e0209083.
Lu, S., Wang, X., Zhang, G., and Zhou, X. (2015). Ef-
fective algorithms of the moore-penrose inverse matrices
for extreme learning machine. Intelligent Data Analysis,
19(4):743–760.
Luo, J., Song, D., Xiu, C., Geng, S., Dong, T., et al. (2014).
Fingerprint classification combining curvelet transform
and gray-level cooccurrence matrix. Mathematical Prob-
lems in Engineering, 2014.
Fingerprint Large Classification Using Sequential Learning on Parallel Environment
229

Michelsanti, D., Ene, A.-D., Guichi, Y., Stef, R., Nasrollahi,
K., and Moeslund, T. B. (2017). Fast fingerprint classifi-
cation with deep neural networks. In International Con-
ference on Computer Vision Theory and Applications,
pages 202–209. SCITEPRESS Digital Library.
Peralta, D., Triguero, I., Garc
´
ıa, S., Saeys, Y., Benitez,
J. M., and Herrera, F. (2018). On the use of convolu-
tional neural networks for robust classification of multi-
ple fingerprint captures. International Journal of Intelli-
gent Systems, 33(1):213–230.
Rajanna, U., Erol, A., and Bebis, G. (2010). A comparative
study on feature extraction for fingerprint classification
and performance improvements using rank-level fusion.
Pattern Analysis and Applications, 13:263–272.
Raschka, S. (2018). Model evaluation, model selection, and
algorithm selection in machine learning. arXiv preprint
arXiv:1811.12808.
Saeed, F., Hussain, M., and Aboalsamh, H. A. (2018). Clas-
sification of live scanned fingerprints using histogram of
gradient descriptor. In 2018 21st Saudi computer society
national computer conference (NCC), pages 1–5. IEEE.
Saini, M. K., Saini, J., and Sharma, S. (2013). Moment
based wavelet filter design for fingerprint classification.
In 2013 International Conference on Signal Processing
and Communication (ICSC), pages 267–270. IEEE.
Shrein, J. M. (2017). Fingerprint classification using convo-
lutional neural networks and ridge orientation images. In
2017 IEEE Symposium Series on Computational Intelli-
gence (SSCI), pages 1–8. IEEE.
Vitello, G., Sorbello, F., Migliore, G., Conti, V., and
Vitabile, S. (2014). A novel technique for fingerprint
classification based on fuzzy c-means and naive bayes
classifier. In 2014 Eighth International Conference on
Complex, Intelligent and Software Intensive Systems,
pages 155–161. IEEE.
Wang, R., Han, C., and Guo, T. (2016). A novel finger-
print classification method based on deep learning. In
2016 23rd International Conference on Pattern Recogni-
tion (ICPR), pages 931–936. IEEE.
Zabala-Blanco, D., Mora, M., Azurdia-Meza, C. A., De-
hghan Firoozabadi, A., Palacios J
´
ativa, P., and Soto, I.
(2020a). Relaxation of the radio-frequency linewidth
for coherent-optical orthogonal frequency-division mul-
tiplexing schemes by employing the improved extreme
learning machine. Symmetry, 12(4):632.
Zabala-Blanco, D., Mora, M., Barrientos, R. J., Hern
´
andez-
Garc
´
ıa, R., and Naranjo-Torres, J. (2020b). Finger-
print classification through standard and weighted ex-
treme learning machines. applied sciences, 10(12):4125.
Zhang, Y., Wang, Y., Zhou, G., Jin, J., Wang, B., Wang, X.,
and Cichocki, A. (2018). Multi-kernel extreme learning
machine for eeg classification in brain-computer inter-
faces. Expert Systems with Applications, 96:302–310.
Zia, T., Ghafoor, M., Tariq, S. A., and Taj, I. A. (2019).
Robust fingerprint classification with bayesian convolu-
tional networks. IET Image Processing, 13(8):1280–
1288.
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
230