Handwritten Text Verification on Mobile Devices
Nilson Donizete Guerin Júnior, Flávio de Barros Vidal and Bruno Macchiavello
Department of Computer Science, University of Brasilia, Brasilia, DF, Brazil
Keywords:
Handwriting Recognition, Biometrics, Mobile Device Authentication.
Abstract:
In this work we propose an online verification system for both signature and isolated cursive words. The
proposed system is designed to be used in a mobile device with limited computational capability. In the
proposed scenario it is assumed that the user will use either his fingertip or a passive pen, therefore no azimuth
or inclination information is available. Isolated words have certain desirable traits that can be more useful on
a mobile device. Different isolated words can be used to verify the user in different applications, combining
a knowledge-based security systems (i.e. passwords) with a behavioral biometric verification system. The
proposed technique can achieve 4.39% of equal error rate for signatures and 6.5% for isolated words.
1 INTRODUCTION
The use of biometric systems has significantly in-
creased in recent years. Applications of these systems
can vary from identity authentication during secure
transactions to granting physical access to certain lo-
cation. The goal of a biometric system is to recog-
nize an individual based on a set of unique attributes.
These attributes are inherent characteristics of the in-
dividual, which gives biometric systems an advantage
against systems based on knowledge (e.g. passwords)
that can be forgotten, or tokens (e.g. badges, IDs)
that can be lost (Impedovo and Pirlo, 2008). Bio-
metric recognition based on physical attributes, like
fingerprints, normally presents a higher performance
than recognition based on behavioral characteristics.
Nonetheless, behavioral attributes are less invasive,
and some of them, like a person signature, are widely
accepted as a legal mean to verify a person’s identity.
A handwritten text biometric system can be clas-
sified into two categories: offline and online. Offline
systems consist basically on the analysis of the shape
information contained in the input image (Bulacu and
Schomaker, 2007; Tang et al., 2013), while online
systems can access features like trajectory, pressure
and velocity which are more unique, making harder to
forge an identity (Fierrez-Aguilar et al., 2005; Nanni
and Lumini, 2008; Yanikoglu and Kholmatov, 2009;
Sae-Bae and Memon, 2014; Zalasi
´
nski and Cpałka,
2013; Cpałka and Zalasi
´
nski, 2014). In both cases, as
in most biometric systems, the users will be normally
asked to be enrolled by providing one or more sam-
ples. Recognition is then performed by comparing the
new sample to the previously stored. The recognition
can be performed either to verify or identify certain
individual. In verification mode the subject claims its
identity, which is then authenticated. In identification
task the subject’s identity is established among those
enrolled.
Feature extraction is an important part of a hand-
written text verification system. Global features are
related to the text as a whole, while local features,
also referred as function-based, are measured at each
point along the trajectory of the text. Systems based
on global features have higher error rates than the
function-based ones, on the other hand they have a
much lower computational load and can be used as a
first step in a hybrid (local-global) system (Yanikoglu
and Kholmatov, 2009).
Handwritten signature has a long history as an
identity authentication method, mostly due to the fact
that signatures exhibit considerable inter-writer vari-
ability. Nowadays there is a growing need for security
application in mobile devices (like touch screen lap-
tops), where signature verification can substitute or
complement other authentication methods. However,
if the signature is compromised it is not easy for a
person to create a new and different signature. There-
fore, the use of isolated words for writer authentica-
tion in mobile devices can be more appropriate. Iso-
lated words are easy to modify once the previous one
has been compromised. The user can be requested to
use a specific word or sentence, combining behavioral
biometrics with knowledge-based systems. Further-
more, different words can be used for authentication
during different actions. Nevertheless, isolated words
26
Guerin Júnior N., de Barros Vidal F. and Macchiavello B..
Handwritten Text Verification on Mobile Devices.
DOI: 10.5220/0005355200260033
In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 26-33
ISBN: 978-989-758-091-8
Copyright
c
 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Data Acquisition -
x, y and timestamp
Extraction of multi-
histogram features
Manhattan Classifier based on
Fourier features and multi-
histogram analysis
Begin
Decision (complete or
cascade approach)
Profile creation –
combination of features
and statistic parameters
Enrollment?
Yes
End
No
Preprocessing –
mean removal and
time processing
Extraction of
Fourier features
Database
Figure 1: Architecture of the proposed system.
have fewer discrimination capacity than signatures,
therefore low error rates are harder to achieve. On-
line handwritten recognition based on isolated words
have received little attention in the literature. A previ-
ous work (Sesa-Nogueras and Faundez-Zanuy, 2012)
focused on block uppercase letters, identifying the
writer based on pen-down and pen-up strokes using
information from an active pen found in traditional
digitizers, that provides information like azimuth, and
inclination. However, in a regular mobile device an
active pen is not always available. A user will nor-
mally use a finger or a passive pen to write in his
mobile device. Histograms of angles combined with
pressure information has been used recently for sig-
nature verification on mobile devices (Sae-Bae and
Memon, 2014).
In this work, we propose a global feature system,
that works for both handwritten signature and isolated
words. It is assumed that the user will use cursive
handwritten, which is more natural than uppercase
letters, and that the system will work on a mobile de-
vice without the use of an active pen. The proposed
technique combines histogram of angles, velocity and
a Fourier Transform analysis. Different from (Sae-
Bae and Memon, 2014), we use one-dimensional
(1 − D) histograms of angles to create several levels
of discrimination. Moreover, pressure is not required
as a feature, since it can change depending on the use
of a passive pen or fingertip. An spectral analysis of
the online text using Fast Fourier Transform (FFT)
has been already proposed (Yanikoglu and Kholma-
tov, 2009). The main advantages of this technique
is its capacity to compactly represent an online text,
which leads to fast matching algorithms. However,
the error-rates of FFT descriptors are not as good as
other global methods. Here, a variation of the previ-
ously proposed FFT analysis is used along with his-
tograms of angles. Finally, the proposed technique
does not require any online training or elaborated
matching algorithm which can limit its use in a power-
constrained mobile device.
2 SYSTEM ARCHITECTURE
The general system architecture is composed of two
main tasks: the enrollment and verification. In the
first one, a person is registered in the system by donat-
ing M reference samples to create his profile. In the
second, the person will be authenticated in the sys-
tem by providing a sample and the user identification
which is claiming to be. The architecture is described
in Figure 1.
HandwrittenTextVerificationonMobileDevices
27

2.1 Data Acquisition
In this study, our main focus is handwritten text, sig-
natures or cursive isolated words, captured by a mo-
bile touch sensitive device without using any active
pen. In the case of isolated words, the user can pro-
vide samples of several words, which later can be
used to verify different actions.
To the best of our knowledge there’s no available
database for isolated cursive handwritten words in the
proposed scenario, therefore we created one. Our
database will be referred as LISA-01
1
(from the Lab-
oratory of Image and Audio Signals of the University
of Brasilia). The device used to collect the cursive
words was a Samsung tablet, model Galaxy Tab 7.0
Plus, with Operational System Android 4.1.2. The
database is composed of 50 writers, each one donating
10 samples of 4 different words: “love", “December",
“intelligence" and “pattern". Words could have few
discontinuities, even though they should be predom-
inantly cursive (see Figure 2). They were all written
by fingertip.
Figure 2: Sample of word "pattern".
2.2 Preprocessing
Previous works reported loss of discrimination dur-
ing recognition due to preprocessing tasks (Kholma-
tov and Yanikoglu, 2005). Therefore, we avoided any
filtering during preprocessing.
The collected data can be seen as a vector of
the form V
j
= [timeStamp
j
, ˆx
j
, ˆy
j
], where 1 ≤ j ≤
N, N is the number of points acquired for a certain
word/signature, timeStamp
j
is the time in millisec-
onds (since the application started) when the point j
was acquired and the pair ( ˆx
j
, ˆy
j
) represents the spa-
tial coordinates.
Initially, we obtain the difference time vector by
subtracting the first timestamp from all others: [t
j
] =
[timeStamp
j
−timeStamp
0
]. Then, we performed the
mean removal from the spatial coordinates: [x
j
,y
j
] =
[ ˆx
j
− Mean( ˆx), ˆy
j
− Mean( ˆy)]. This mean removal do
not interfere with the histogram computation and it is
1
The vectors of the dynamic features and images are
available at http://www.cic.unb.br/~fbvidal/htdb.
necessary in order to obtain translation invariance on
the Fourier descriptors.
2.3 Feature Extraction
Histograms are widely used as feature sets in or-
der to capture attribute statistics in several recogni-
tion tasks, for example, object recognition (Chaudhry
et al., 2009), human action recognition (da Rocha
et al., 2012) and off-line signature verification (Pirlo
and Impedovo, 2013). Works applying histograms
of features to on-line signature recognition can also
be found elsewhere (Sae-Bae and Memon, 2014;
Fierrez-Aguilar et al., 2005), however here we pro-
pose a novel use of histogram of angles along with
Fourier analysis for on-line verification.
Given the feature vector V
j
= [t
j
,x
j
,y
j
], 1 ≤ j ≤ N,
it is possible to derive other important information, by
instance the angle/direction θ
j
= tan
−1

y
j+1
−y
j
x
j+1
−x
j

, the
magnitude mag
j
=
q
x
j+1
− x
j
2
+ (y
j+1
− y
j
)
2
, the
velocity in both directions, (v
x
)
j
=
x
j
−x
j−1
t
j
−t
j−1
, (v
y
)
j
=
y
j
−y
j−1
t
j
−t
j−1
, and so on.
In our approach an on-line word or signature is
represented by a set of histograms of θ
j
, with values
limited to [−π,π], a velocity feature and Fourier de-
scriptors. The set of histograms is defined as H =
{h
1
,h
2
,...h
K
}, where each histogram h
i
consists of
angle counts divided into b
i
bins, with b
1
< b
2
< ... <
b
i
< ... < b
K
. For the remaining of the paper we will
refer h
i
as a histogram of level i. Since, the system
will work with different words/signatures from differ-
ent users, the optimal number of bins can vary. There-
fore, the use of several levels can improve the sys-
tem’s performance in the presence of intra-user vari-
ability.
The Fourier descriptors were obtained using a
technique inspired on a previous work (Yanikoglu and
Kholmatov, 2009). We apply FFT and obtain a coef-
ficient vector F (V
j
) = [F (x
j
),F (y
j
)]. These coeffi-
cients are normalize by their respective total magni-
tude, i.e.
N
∑
j=1
|F (y
j
)| for the vertical component and
N
∑
j=1
|F (x
j
)| for the horizontal component. We discard
half of the spectrum, due to symmetry, and the DC
component. The final step is to average two consecu-
tive descriptors, to account for variations in the neigh-
boring harmonics.
In order to obtain an equal number of Fourier
descriptors, prior to FFT computation, we pad each
sample with zeros. The size of the padded sample
varies for each user and word/signature. It is set to
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
28

1.5 times the longest word/signature in the enrolled
samples. Differently from (Yanikoglu and Kholma-
tov, 2009), this padding process will prevent for con-
stant cropping of future samples during verification,
thus avoiding loss of data. Also, we found no need
for drift estimation as done in (Yanikoglu and Khol-
matov, 2009). Since, the Fourier descriptors are used
along with histogram information, no significant gain
is obtained by performing drift correction. The first
p coefficients of each sample generates the descriptor
vector F = [ f
x
, f
y
].
As mentioned earlier, a velocity feature is also
used. It has already been showed that velocity is a
discriminative and helpful feature regarding signature
verification (Rashidi et al., 2012). Our feature is com-
puted as follows:
1. Compute the root mean square velocity of all
points in a sample:
RMS_Vel =
v
u
u
t
∑
N
j=1
(v
x
)
j
N
!
2
+
∑
N
j=1
(v
y
)
j
N
!
2
;
(1)
2. Evaluate the difference in velocity of two con-
secutive points, referred as di f f _Vel (this feature
will be related to acceleration);
3. Compute the root mean square of di f f _Vel, sim-
ilar to (1), referred as RMS_Di f
4. Combine these two metrics into one value:
velFeature = 0.6 × RMS_Di f + 0.4 × RMS_Vel.
2.4 Profile Creation
A profile is created when a new user registers M sam-
ples of a particular word/signature following the steps
described below:
1. Compute the mean between all of the M provided
samples of the velocity feature, meanVelFeature;
2. For each set H
i
=

h
i
1
,h
i
2
,..., h
i
M

with b
i
bins:
(a) Calculate a matrix, dMatrix of size M × M,
whose elements are given by d
m,n
= ||h
i
m
−h
i
n
||,
where || • || indicates the Manhattan distance
between two vectors. Each line of this matrix
represents a comparison between the jth sam-
ple with all other samples;
(b) Calculate the column vector, minVector of size
M, where each element is the mean of the re-
spective line in dMatrix, disregarding the self-
comparisons;
(c) Evaluate the following metrics: (i) the mean of
the minimum value in each line of dMatrix, de-
fined as minVal
i
, which indicates the average
distance between all nearest pairs of samples,
(ii) tempVal
i
= minimum(minVector), which
reflects the average distance of the sample that
is closest to all others, and (iii) maxVal
i
, the
mean of the maximum value in each line of
dMatrix, that measures the average distance
between all farthest pairs of samples;
(d) The sample that corresponds to tempVal
i
is also
identified. This template sample will be the
closest in average to all other samples and will
be denoted by tempSample
i
;
(e) The statistics vector P
i
H
, for
level i, is defined as P
i
H
=
{minVal
i
,tempVal
i
,maxVal
i
,tempSample
i
,H};
3. For the set of Fourier descriptors F =
{F
1
,F
2
,..., F
M
}, a similar approach is per-
formed and the statistics vector P
F
=
{minVal,tempVal,maxVal,tempSample,F}
is obtained.
The profile is then represented by the
set P = {meanVelFeature,P
H
,P
F
}, where
P
H
= {P
1
H
,P
2
H
,..., P
K
H
}
2.5 Classifier
This stage is performed during verification and is in-
spired on a previous work (Kholmatov and Yanikoglu,
2005). Considering the query sample, denoted by S,
and a profile P, for each set of histograms H
i
in the
profile and the corresponding query histogram H
i
S
, a
comparison is performed as follows:
1. Compute the column vector di f f Query, of length
M, whose elements are given by q
j
= ||h
i
S
− h
i
j
||,
where || • || indicates the Manhattan distance be-
tween two vectors;
2. Select three values from di f f Query: the mini-
mum (min
S
), the maximum (max
S
) and the differ-
ence to the corresponding tempSample (temp
S
);
3. Normalize those values by the respective com-
ponents of profile P at level i: min
0
S
=
min
S
/minVal
i
, tempS
0
= temp
S
/tempVal
i
and
max
0
S
= max
S
/maxVal
i
;
Note that min
0
S
, temp
0
S
and max
0
S
are highly cor-
related. Therefore, the three-dimensional vector can
be reduced to one dimension using PCA (Principal
component analysis). This will generate the resulting
distance value:
histDistance
i
= k
1
× min
0
S
+ k
2
×temp
0
S
+ k
3
× max
0
S
.
For the set of Fourier descriptors F, a similar ap-
proach is performed and a unique f ourierScore is
HandwrittenTextVerificationonMobileDevices
29

obtained. However, after several tests it was veri-
fied that the we could use fixed values k
1
= 0.3870,
k
2
= 0.3705 and k
3
= 0.2425 without any significant
loss in performance. Hence, differently from many
proposals (Rashidi et al., 2012; Sesa-Nogueras, 2011;
Sesa-Nogueras and Faundez-Zanuy, 2012; Zalasi
´
nski
and Cpałka, 2013; Maiorana et al., 2012; Kholmatov
and Yanikoglu, 2005), we do not require a training
phase during user enrollment.
In previous works several techniques for features
combination were presented (Damer et al., 2013; Gu-
davalli et al., 2012; Scheidat et al., 2011; Shekhar
et al., 2014). Here, we use a score value based on
a linear combination of the acquired features. More-
over, the proposed technique can be used in two
ways: (i) the “complete approach”, where the in-
formation obtained from all the different histograms
levels, histDistance, is computed as the mean of all
histDistance
i
and (ii) using the histDistance
i
values
at each level in a “cascade approach”. In the first
case, we calculate the score f inalDistance = k
4
×
histDistance + k
5
× f ourierScore + k
6
× velScore,
where velScore = |meanVelFeature − velFeature
S
|.
This method will yield better performance, the intra-
user variability is minimized due to the average of
the statistics, while the inter-user variability is max-
imized. After several empirical tests, we verified that
the constants can be set as k
4
= 0.29, k
5
= 0.57 and
k
6
= 0.14. This final score is compared to certain
threshold, and the sample is considered to belong to
the user if it is below that threshold.
The cascade approach can be used when the mo-
bile device has limited computational capabilities or
when it is operating in low-battery mode. A final
score is computed at each level (levelDistance
i
= k
4
×
histDistance
i
+ k
5
× f ourierScore + k
6
× velScore),
starting with the level with fewer number of bins.
Once again, this value is compare to a threshold, if
the sample is above that threshold we continue to the
next level, otherwise the sample is considered to be-
long to that profile. As it can be noticed, in this ap-
proach the computational effort during authentication
is reduced however it happens at the expense of error
rate increment.
2.6 Decision
The decision stage is basically the comparison be-
tween the metric obtained in the classifying stage with
a given threshold. Since we adopted a user-dependent
normalization, it’s possible to define a global thresh-
old instead of finding optimal thresholds for each pro-
file. In the cascade approach, there is a threshold for
each group of histograms at level i. Again, in or-
der to avoid specific optimization at each level, the
same threshold used in the complete approach can be
adapted to be used in the cascade approach. The set
of thresholds T
i
can be defined as T
i
= a
i
× T , where
0 < a
1
< a
2
< ... < a
i
< ... < a
K
= 1 and T indicates
the global threshold.
3 EXPERIMENTS AND RESULTS
As mentioned in previous works, comparison be-
tween methods for signature verification is not easy
(Yanikoglu and Kholmatov, 2009; Cpalka et al., 2014)
mainly due to the difference in databases and pro-
posed scenarios. Nevertheless, in this section we
present the results of the proposed technique and
compare them to previously reported results. For sig-
nature verification we used a widely adopted signa-
ture database, the MCYT-100 (Ortega-Garcia et al.,
2003).
3.1 Experiments Setup
We use the following setup for our simulations:
• The number of reference samples, M, was set to
5 . This number of enrollment samples was also
used in (Yanikoglu and Kholmatov, 2009; Cpałka
and Zalasi
´
nski, 2014; Zalasi
´
nski and Cpałka,
2013);
• We used the set of bins numbers B =
{6,8, 10,12,14, 16};
• We adopted p = 30 for the number of Fourier de-
scriptors;
• Simulations were repeated 5 times on each
database, raffling the samples used for compos-
ing the profiles and leaving the remaining ones for
testing. This is a common practice (Yanikoglu and
Kholmatov, 2009; Rashidi et al., 2012) in order to
better evaluate the system;
• Results were measured using EER (Equal Error
Rate), a widely used measure (Cpałka and Za-
lasi
´
nski, 2014; Kholmatov and Yanikoglu, 2005;
Maiorana et al., 2012; Rashidi et al., 2012; Sesa-
Nogueras, 2011; Sesa-Nogueras and Faundez-
Zanuy, 2012; Yanikoglu and Kholmatov, 2009;
Zalasi
´
nski and Cpałka, 2013);
• When using the LISA-01 database, we also raffled
samples of other users to compose the forgeries
ones;
• In MCYT-100 database, skilled forgeries for each
user are available. Therefore, we evaluated the
system considering that the forger has access to
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
30

the reference signature. Note, that in this scenario
the expected ERR’s are higher than using random
forgeries.
3.2 Histograms Analysis
First, we analyze the performance of the proposed
“complete” and “cascade” approaches. We ran sev-
eral tests and selected a fixed the number of bins that
yields the better overall results (referred as the “Best
Histogram” mode). Then, we performed verification
for all users, using only the histogram information.
Table 1 shows the results.
Table 1: Scenarios of histogram analysis.
Kind of analysis EER
Best histogram - 16 bins 14.17%
Complete approach 12.49%
Cascade approach 13.56%
From the results, it can be observed that both pro-
posed techniques present an improvement over the
use of a single histogram of angles.
3.3 Signature Evaluation
In Table 2 the results for signature verification using
the MCYT-100 database of the proposed techniques
and previous works are presented.
Even though the proposed techniques do not yield
the best EER’s, our results are adequate when com-
pared to previous studies. However, it is important to
notice that our technique does not require any com-
plex training phase (Maiorana et al., 2012), com-
plex 2− D histograms analysis (Sae-Bae and Memon,
2014), local features (Van et al., 2007) or the use of
features from active devices (Maiorana et al., 2012).
3.4 Cursive Isolated Words Evaluation
In Table 3 we present the results from isolated cur-
sive words using the LISA-01 database. In order
to provide a comparison, we implemented a previ-
ous proposal for signature verification (Yanikoglu and
Kholmatov, 2009) and applied to our isolated words
database. The implemented proposal is based only on
Fourier descriptors. Therefore, this comparison can
show how the use of histograms of angles can ade-
quately complement the use of Fourier descriptors.
As expected the ERR for isolated words is higher.
This is due to the fact that isolated words have less
discriminative power than signatures. Nevertheless,
as mentioned earlier, isolated words possess desirable
traits for use in mobile devices. The observed values
of EER for each word separately in the complete ap-
proach were 7.2%, 5.78%, 8.6% and 4.42% for the
words “December", “intelligence", “love" and “pat-
tern", respectively, when considering forgeries from
the same words. It means that the forger has al-
ready acquired the knowledge of the correct word.
These rates become 4.8%, 3.86%, 6%, 3.86%, re-
spectively, if the forgeries are raffled from random
words (which can include the same word). This case
leads to an EER of 4.63%. Note that the results
suggest that shorter words have fewer discriminative
power, as already noticed elsewhere (Sesa-Nogueras
and Faundez-Zanuy, 2012; Sesa-Nogueras, 2011). In
Fig. 3 the curves relating the False Acceptance Rate
(FAR) and False Rejection Rate (FRR) are depicted.
The point of intersection represents the Equal Error
Rate (EER) of the approaches. It’s related to the word
“pattern” in LISA-01 database, and it can be observed
that the proposed techniques have better performance
than the simple use of Fourier Descriptors.
3.5 Computational Complexity
The system proposed here is based on extraction of
global features. We performed an analysis on the
frequency domain in order to compose Fourier de-
scriptors vectors and a spatial analysis to create his-
tograms of angles. A velocity feature is also com-
puted. The asymptotic complexity for calculating the
velocity feature is O(N), since we pass over the fea-
ture vector a constant number of times.
In the case of histograms, we also perform O(N),
considering that we pass a constant number of times
over the vector to create a fixed number of histograms
with constant number of bins. The creation of the
Fourier descriptor involves the computation of the
FFT. The algorithm for this has asymptotic complex-
ity O(N log N), also performed a constant number of
times. The overall asymptotic system computational
complexity is O(N log N).
4 CONCLUSIONS
We presented a method for both handwritten signa-
tures and isolated cursive words to be used on mobile
devices. The technique does not require any complex
training or computation calculations, therefore it is
adequate for the limited capabilities of a mobile de-
vice. The results show a good performance for both
signatures and isolated words, without the use of fea-
tures like azimuth, inclination or pressure that are de-
pendent on the input device (active pen, passive pen or
HandwrittenTextVerificationonMobileDevices
31

0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6
0
10
20
30
40
50
60
Error Rate (%)
Threshold
(a) Fourier(Yanikoglu and Kholmatov, 2009)
0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6
0
10
20
30
40
50
60
Error Rate (%)
Threshold
(b) Complete approach
0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6
0
10
20
30
40
50
60
Error Rate (%)
Threshold
(c) Cascade approach
Figure 3: The relation of False Acceptance Rate (FAR, in red) and False Rejection Rate (FRR, in blue) related to a given
threshold. The point of intersection, given by the dashed lines represents the Equal Error Rate (EER). Each plot relates to
each of the approaches tested with respect to the word “pattern”.
Table 2: Signature verification on MCYT-100 database.
Proposal Features used EER for skilled forgeries
Histogram + Quantizer (Sae-Bae and Memon, 2014) x, y, time, pressure 4.02%
HMM + Likelihood (Van et al., 2007) All available 3.37%
Polynomial (Maiorana et al., 2012) All available 4.22%
Fourier (Yanikoglu and Kholmatov, 2009) All available 10.89%
DCT + WT (Nanni and Lumini, 2008) x, y, azimuth 9.80%
Our proposal (complete) x, y, time 4.39%
Our proposal (cascade) x, y, time 4.64%
Table 3: Results obtained in the LISA-01.
Proposal Features used EER for skilled forgeries
Fourier (Yanikoglu and Kholmatov, 2009) x, y 17.19%
Our proposal (complete) x, y, time 6.5%
Our proposal (cascade) x, y, time 7.18%
fingertip). Future work will focus on adopting a bet-
ter and systematic fusion method for combining the
different scores (histogram, Fourier descriptors, etc.).
Also it can be possible to analyze parts of the pro-
vided samples, instead of the complete word or sig-
nature, that may have more discriminative informa-
tion between different users. Moreover, an analysis of
the robustness of the system to external factors, such
as word inclination, or user movement while writing,
can be made. Finally, we can expand the LISA-01
database in order to include skilled forgeries.
REFERENCES
Bulacu, M. and Schomaker, L. (2007). Text-independent
writer identification and verification using textural and
allographic features. In IEEE Transactions on Pattern
Analysis and Machine Intelligence, pages 701–717.
Chaudhry, R., Ravichandran, A., Hager, G., and Vidal, R.
(2009). Histograms of oriented optical flow and binet-
cauchy kernels on nonlinear dynamical systems for
the recognition of human actions. In Computer Vision
and Pattern Recognition, 2009. CVPR 2009. IEEE
Conference on, pages 1932–1939. IEEE.
Cpałka, K. and Zalasi
´
nski, M. (2014). On-line signature
verification using vertical signature partitioning. Ex-
pert Systems with Applications, 41(9):4170–4180.
Cpalka, S., Zalasinski, M., and Rutkowski, L. (2014).
New method for the on-line signature verification
based on horizontal partitioning. Pattern Recognition,
47(8):2652–2661.
da Rocha, T., De Barros Vidal, F., and Romariz, A. R. S.
(2012). A proposal for human action classification
based on motion analysis and artificial neural net-
works. In Neural Networks (IJCNN), The 2012 In-
ternational Joint Conference on, pages 1–6.
Damer, N., Fuhrer, B., and Kuijper, A. (2013). Missing
data estimation in multi-biometric identification and
verification. In Biometric Measurements and Systems
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
32

for Security and Medical Applications (BIOMS), 2013
IEEE Workshop on, pages 41–45. IEEE.
Fierrez-Aguilar, J., Nanni, L., Lopez-Peñalba, J., Ortega-
Garcia, J., and Maltoni, D. (2005). An on-line signa-
ture verification system based on fusion of local and
global information. In Audio-and video-based biomet-
ric person authentication, pages 523–532. Springer.
Gudavalli, M., Babu, A., Raju, S., and Kumar, D. S. (2012).
Multimodal biometrics–sources, architecture and fu-
sion techniques: An overview. In Biometrics and
Security Technologies (ISBAST), 2012 International
Symposium on, pages 27–34. IEEE.
Impedovo, D. and Pirlo, G. (2008). Automatic signature
verification: the state of the art. Systems, Man, and
Cybernetics, Part C: Applications and Reviews, IEEE
Transactions on, 38(5):609–635.
Kholmatov, A. and Yanikoglu, B. (2005). Identity authen-
tication using improved online signature verification
method. Pattern recognition letters, 26(15):2400–
2408.
Maiorana, E., Campisi, P., La Rocca, D., and Scarano, G.
(2012). Use of polynomial classifiers for on-line sig-
nature recognition. In Biometrics: Theory, Applica-
tions and Systems (BTAS), 2012 IEEE Fifth Interna-
tional Conference on, pages 265–270. IEEE.
Nanni, L. and Lumini, A. (2008). A novel local on-line sig-
nature verification system. Pattern Recognition Let-
ters, 29(5):559–568.
Ortega-Garcia, J., Fierrez-Aguilar, J., Simon, D., Gonza-
lez, J., Faundez-Zanuy, M., Espinosa, V., Satue, A.,
Hernaez, I., Igarza, J.-J., Vivaracho, C., et al. (2003).
Mcyt baseline corpus: a bimodal biometric database.
IEE Proceedings-Vision, Image and Signal Process-
ing, 150(6):395–401.
Pirlo, G. and Impedovo, D. (2013). Verification of static
signatures by optical flow analysis. Human-Machine
Systems, IEEE Transactions on, 43(5):499–505.
Rashidi, S., Fallah, A., and Towhidkhah, F. (2012). Fea-
ture extraction based dct on dynamic signature verifi-
cation. Scientia Iranica, 19(6):1810–1819.
Sae-Bae, N. and Memon, N. (2014). Online signature veri-
fication on mobile devices. Information Forensics and
Security, IEEE Transactions on, 9(6):933–947.
Scheidat, T., Vielhauer, C., and Fischer, R. (2011). Com-
parative study on fusion strategies for biometric hand-
writing. In Proceedings of the thirteenth ACM mul-
timedia workshop on Multimedia and security, pages
61–68. ACM.
Sesa-Nogueras, E. (2011). Discriminative power of online
handwritten words for writer recognition. In Security
Technology (ICCST), 2011 IEEE International Carna-
han Conference on, pages 1–8. IEEE.
Sesa-Nogueras, E. and Faundez-Zanuy, M. (2012). Bio-
metric recognition using online uppercase handwrit-
ten text. Pattern Recognition, 45(1):128–144.
Shekhar, S., Patel, V. M., Nasrabadi, N. M., and Chel-
lappa, R. (2014). Joint sparse representation for ro-
bust multimodal biometrics recognition. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
36(1):113–126.
Tang, Y., Wu, X., and Bu, W. (2013). Offline text-
independent writer identification using stroke frag-
ment and contour based features. In Biometrics (ICB),
2013 International Conference on, pages 1–6.
Van, B. L., Garcia-Salicetti, S., and Dorizzi, B. (2007). On
using the viterbi path along with hmm likelihood in-
formation for online signature verification. Systems,
Man, and Cybernetics, Part B: Cybernetics, IEEE
Transactions on, 37(5):1237–1247.
Yanikoglu, B. and Kholmatov, A. (2009). Online signature
verification using fourier descriptors. EURASIP Jour-
nal on Advances in Signal Processing, 2009:12.
Zalasi
´
nski, M. and Cpałka, K. (2013). New approach for
the on-line signature verification based on method of
horizontal partitioning. In Artificial Intelligence and
Soft Computing, pages 342–350. Springer.
HandwrittenTextVerificationonMobileDevices
33