2016 15th International Conference on Frontiers in Handwriting Recognition

Multiple Generation of Bengali Static Signatures

Moises Diaz∗ , Sukalpa Chanda† , Miguel A. Ferrer∗ , Chayan Kr. Banerjee‡ , Anirban Majumdar‡ ,
Cristina Carmona-Duarte∗ , Parikshit Acharya§ and Umapada Pal§
∗ IDeTIC Institute, Universidad de Las Palmas de Gran Canaria, Spain. Email: {mdiaz, mferrer, ccarmona}@idetic.eu
† Alice, University of Groningen, The Netherlands. Email: sukalpa@ieee.org
‡ Chennai Mathematical Institute, Chennai, India. Email: {chayankb, majumdar}@cmi.ac.in
§ Indian Statistical Institute, 203 B.T. Road, 700108 Kolkata, India. Email: {pacharya, umapada}@isical.ac.in

Abstract—Handwritten signature datasets are really necessary

for the purpose of developing and training automatic signature
verification systems. It is desired that all samples in a signature
dataset should exhibit both inter-personal and intra-personal (a) Real Bengali signature (b) Synthetic Bengali signature
variability. A possibility to model this reality seems to be obtained
through the synthesis of signatures. In this paper we propose a Fig. 1: Real and synthetic examples in Bengali script.
method based on motor equivalence model theory to generate
static Bengali signatures. This theory divides the human action
to write mainly into cognitive and motor levels. Due to difference
between scripts, we have redesigned our previous synthesizer [1], could be generated using automatic algorithms taking into
[2], which generates static Western signatures. The experiments account a wide variety of realistic factors such as collecting
assess whether this method can approach the intra and inter- signatures in several sessions; handwriting variations over
personal variability of the Bengali-100 Static Signature DB from time or aging effects; simulation of signatures affected by
a performance-based validation. The similarities reported in some behavioral disorders, neurodegenerative diseases or other
the experimental results proof the ability of the synthesizer to cognitive impairment and so on. In this context, the scientific
generate signature images in this script.
literature has proposed two ways to synthesize the signa-
tures [7], [8]:
I. I NTRODUCTION
1) Generation of duplicated samples. In this case algo-
Beyond a behavioral biometric trait, handwritten signatures rithms create artificial intra-personal variability from reference
can be considered as a human response which expresses specimens. These reference specimens are the seed of these
personal identity. They are well-accepted in many cultures algorithms. Therefore, they do not create new individuals but
for legal validation of documents, e.g. last wills or testa- possible signatures executed by the signers. Both in on-line [9],
ments, corporative tax statements, financial transfers and so on. [10], [11], [12] and off-line [13], [14], [15], [16], [17], [18],
However, the unexpected dissimilarity between two specimens [19], [20], the algorithms are able to generate new samples of
executed by the same signer makes complex and sometimes a real user. Duplicated signatures present several advantages
unpredictable the use of related technology in civil and indus- such as training improvement in ASV, statistically meaningful
trial applications nowadays, especially when many documents evaluations, enlarging the number of samples in a database,
require to be automatically processed. etc.
Nevertheless, common benchmarks and competitions are
2) Generation of new identities. Algorithms are designed
being continuously organized in the most specialized confer-
in this modality by taking into account different rules and
ences to test the Automatic Signature Verifiers (ASV) designed
logics in order to simulate realistic effects. In [21] a model
mainly by industries and academias under the same conditions.
is described to generate dynamic signatures based on visual
Some of the signatures considered in recent competitions in
characteristics extracted from the time domain. The method
automatic signature verification are executed in several scripts1
was evaluated from a visual validation. In Galbally et al. [7],
such as Chinese, Dutch, Japanese, Western or Bengali (e.g. [3],
[8] the full generation of on-line flourish-based signatures was
[4], [5]).
carried out through two algorithms based on spectral analysis
To test different designed ASVs it is mandatory to use and the kinematic theory of rapid human movements. The
huge number of signatures. In fact, the more signatures, the first attempt to generate static flourishes in context of static
better the evaluation from a statistical point of view. However, signature image is due to [22]. Two follow-up works were
collecting signatures is by far one of the most tedious tasks presented in [1], [2] where authors coped with the generation
because of the legal obstacles to make the dataset publicly of Western static and dynamic signatures (text and flourish) as
available. Moreover it is a time consuming and a boring well as skilled forgeries under a neuromotor approach.
process. These drawbacks are being actively addressed by the
use of synthetic signature databases, which is supposed to Although the vast majority of contributions cope with artifi-
be one of the recent emerging issue in biometric signatures cially Western signatures generation, the necessity of statistical
verification [6]. These databases (of practically infinite size) meaningful databases is found equality in all scripts. Beyond
Western signature databases, databases in other scripts such as
1 The set of letters or characters (i.e. symbols) used for writing a particular Arabic [23], Chinese [3], Persian, Russian, Devanagari [24] or
language is know as script. Bengali [24] do exists. The lack of enormous databases for

2167-6445/16 $31.00 © 2016 IEEE 42

DOI 10.1109/ICFHR.2016.18
 C OGNITIVE L EVEL M OTOR L EVEL

C. Ballistic
B. Text D. Ink
A. Morphology Trajectory
Trajectory Plan Deposition Model
and Lexicon Generation
Definition

Fig. 2: Overview of the proposed system to generate a master signature

each script is an open issue nowadays. filters; and (D) A virtual ink deposition model is used to obtain
realistic Bengali signature images.
Taking into account the mentioned lacks, the main contri-
bution of this paper is a neuromotor-based model to generate
static signatures in Bengali script. Specifically, we use a neu- A. Morphology and Lexicon Definition
romotor inspired synthesizer [1] as the basis of the proposed
Bengali static signature generator. In this contribution, we gen- To define a realistic synthetic Bengali signature images, its
erate new identities and duplicates from synthetic specimens morphology and lexicon should be firstly modeled so as to gen-
covering the intra and inter-personal variability, respectively. erate human-like specimens. Apart from the script, one of the
Using the only available static Bengali database [24], we will most different characteristic compared to Western signatures is
evaluate the closeness of our synthetic database and the real that flourishes do not appear in Bengali signatures. They are
one through a performance analysis. An example of real and mainly composed of two words which are related to the name
synthetic Bengali signature images is shown in Fig. 1. and surname as it is shown in Fig. 1. Occasionally, a third word
The outline of the paper is as follows: Section II describes associated to the middle name could come out. By using the
the neuromotor synthesizer to generate Bengali signature im- Bengali static signature dataset [24], we have calculated the
ages. Section III is dedicated to the generation of a synthetic distribution of the number of words per signature and letters
Bengali database. Experiments are discussed in Section IV per word as they are shown in Table I.
and finally, the paper ends with some conclusions and future Similar to [27], we have also taken into account other
research ideas in Section V. characteristics which have been modeled through a Generalize
Extreme Value (GEV) function. This function allows to model
II. N EUROMOTOR APPROACH FOR B ENGALI S IGNATURE a natural phenomena with three parameters {ξi , σi , μi }, namely
G ENERATION location, scale and shape respectively. For this work we have
As a follow-up work of [1], [2], in this work we have modeled the number of letters, slant, skew, text width and text
redesigned such previous signature synthesizer, developed un- height. The parameters which define each GEV are shared in
der common concepts in neuroscience, to generate Bengali Table II. As the GEV distribution range is (−∞, +∞), we
signatures automatically. have also determined the minimum and maximum realistic
values of each distribution. It is worth pointing out that certain
A handwritten signature is the final result of a complex parameters have lower variability in Bengali script compared to
system which involves cognitive and fine motor controls. To Western one such as skew or slant. We refer interested reader
produce synthetic signatures, we have focused on the motor to [27].
equivalence theory, which suggests that the brain stores in
two levels the movements aimed at performing a single task On the other hand, the signature generator is not loaded
like signing. The first level is irrespective of the muscles for real names owing to private reasons. Instead, we generate
and stores in the brain the movement in an abstract form as random names. Among all the Bengali characters, this script
spatial sequence of points which define the trajectory plan. can be defined by 12 vowels and 40 consonants. In order to
The second level depends on the signer muscular skills as create realistic names, after defining the number of words and
it consists of a sequence of motor commands to execute letters per word, we decide which letter should be included
the final ballistic trajectory [25], [1], [2], [26]. This theory in the signature following the frequency in which each letter
has mainly inspired the design of our off-lne handwritten usually appears in Bengali handwritten. Moreover, this script
Bengali signature synthesizer. This signature generator is built does not have the concept of upper case or lower case letters,
following the different steps shown in Fig. 2, which are all letters have the same height, including the starting letter.
summarized as: (A) The morphology and lexicon of Bengali Finally, the letters are usually connected at the head line region
handwritten signatures is firstly defined; (B) At the cognitive in a word. Instead, letters are generally connected in lower
level, letters engrams as fundamental part of the trajectory plan parts in Western style. To decide if a single letter is connected
are defined by hexagonal letter grids; (C) At motor level, the or not with the following one, we have used a simple random
final ballistic trajectory is defined through several kinematic function.

43
TABLE I: Distribution of words and letters in the signature (%) It was an approach inspired in the production of handwriting,
but it does not necessarily mean that we model exactly the
Word per Number of letter per word
Word
signature 1 2 3 4 5 6
muscles action during the handwritten.
1st 46.95 0.00 12.0 51.0 30.0 7.00 0.00 One of the main differences with respect to Western
2nd 46.95 4.00 48.0 43.0 4.00 1.00 0.00
signature is that Bengali signatures do not include any flourish.
3rd 6.10 0.00 69.2 23.0 0.00 0.00 7.80
This leads to reduce the number of inertial filters in this work
as follows: i) The first filter is associated to the finger effect.
So, we have programmed the filter in order to start and stop
TABLE II: Parameters to model the Bengali signature mor- stroke by stroke. It corresponds to neuromuscular commands
phology through GEV function {ξi , σi , μi } previously segmented in the text trajectory plan. This filter is
Parameter ξi σi μi min max most relevant to generate ballistic trajectories since Bengali
Number of letters -0.03 1.03 5.59 4.00 12.00
script is composed of a lot of small strokes. ii) The second
Slant (degrees) -0.26 12.32 -4.75 -28.50 33.69 filter applied to the signature is the so called wrist inertial
Skew (degrees) -0.17 2.03 -0.94 -5.06 6.12 filter. Because the typology in Bengali signatures, this filter
Text width (mm) -0.12 30.54 148.9 98.82 251.8 was reduced to the minimum. It heuristically imitates the wrist
Text height (mm) -0.06 6.55 31.11 19.07 65.50 movement from left to write when they sign as therefore, it is
applied to all the engram. Additionally, it has been noticed that
certain strokes end in the middle of other strokes producing a
B. Text Trajectory Plan contact point in the middle of them. To guarantee the continuity
between these kind of strokes, according to this observation,
To approach the cognitive level, our synthesizer takes into the filtered trajectory is constrained to pass by these contact
account the writing sequence of the text trajectory plan by the points.
composition of each letter trajectory plan. It consists in two
steps: To preserve the human-like appearance of the synthetic
signature, several inertial filters could be used such as Bezier
1) A hexagonal grid is initially defined. Its horizontal curves or Savitsky-Golay filter [22]. However, we found the
and vertical distances are defined as they are kept use of Kaiser filters [1], [2] more flexible, whose finite impulse
constant so as to design a specific writing style per response ht [n] is defined as:
signer [1], [2]. Since Bengali characters are more
complex than Western ones, we had defined a denser
grid in this work. Such denser grid is designed by ⎧   2

aumenting its resolution. ⎪
⎪ I0 πβ 1−( N2n
−1 −1)
⎨ 0≤n≤N −1
2) Next, stroke limits are defined over the grid as well as I0 (πβ)
ht [n] = (1)
the writing strokes order, which are kept constant per ⎪
⎪
writer. To build the text trajectory plan, we connect ⎩
0 otherwise
directly each end and initial point of each consecutive
letter. In Fig. 2 we could see the points to define the
text trajectory plan of a word with the stroke limits Where β is a shape factor fixed to 0.1 and N the filter
highlighted on the grid. length, which is related to the signing velocity. Being Nf =
Finally, to give more realism to Bengali synthetic signa- Lf v 2 for the finger inertial filter length and Nw = Lw v 2 for
tures, the upperline of a word is kept constant in the grid, the wrist inertial filter. Both Lf and Lw represents inertial
i.e., the upperline points’ letters, which compose the trajectory variables. Finally, v is associated to the signing velocity and it
plan, are located in the same row. It should be noted that the was modeled through a GEV function whose parameters are:
points defined in this section are not necessarily related to the {ξi , σi , μi } = {−0.07, 0.65, 2.63} in the range (2.00, 5.00).
so called target points in the cortex action plan. This part of the
synthesizer is loosely based on the motor equivalence model
theory but it does not pretend to model or simulate it. D. Ink Deposition Model
To obtain the off-line version of the signatures, we have
C. Ballistic Trajectory Generation 8-connected the signatures (i.e. the skeleton) with the Bre-
To approach the motor level we have used the inverse senham’s lines algorithm and then we apply the virtual ink
model proposed for static Western signature generation [1], [2]. deposition model [22], [1], which use three kind of inks
In Western signatures generation a multi-level motor scheme and different ballpoint pen sizes. Finally, signature image
was used based on several kinematic filters to simulate the resolution is adjusted at 600 dpi.
inertial action produced by three main parts: finger, forearm
and wrist. Briefly, in Western signature generation, the finger
was approached by an inertial filter applied to the whole III. B ENGALI DATABASE GENERATION
engram related to the text trajectory plan; the forearm was
In this section, we describe the procedure to adjust the
simulated by a filter applied to the flourish engram; and the
proposed synthesizer in order to approach both the inter and
wrist by a filter applied to the whole signature because the wrist
intra-personal variability in Bengali signatures similar to each
moves continuously when writing both the text and flourish.
other [1], [2].

44
A. Generation of New Identities intra-personal variability achieved with this method compared
to real signatures.
The inter-personal variability is introduced at morpholog-
ical, cognitive and motor levels in order to create master Finally, some of the parameters used in the Bengali signa-
signatures. On the morphological level, we modify the signer ture generator are not strictly related to this script. For further
full name according to the distributions given in Section II-A. details in these parameters, reader can refer to [1], [2].
On the cognitive level, the distance between rows and columns
of the signer grid was set for each writer so as to design a text IV. E XPERIMENTS
height and width letter style. On the motor level, the inertial The aim of the experiment is to evaluate the similarity
variables (Lf , Lw ) were randomly chosen from three to four between the synthetic and real signatures when they are
times the distance between the grid nodes and kept constant examined from two completely different Automatic Signature
for each writer. Verifiers (ASVs). Thus, as real database we have used the
Finally, because there are more stable signers than others, Bengali-100 Static Signature DB [24], composed of 100 user
i.e. signers which repeat their own signatures with a high and 24 repetition per individual (2400 signatures in total).
similarity among them, a stability parameter s is defined for As ASVs, we have used a geometric features followed by a
each writer. A visual example of the capacity to generate HMM classifier [28] and a texture-based followed by an SVM
several synthetic human-like signatures is shown in Fig. 3. classifier [29]. Both ASVs are based in completely different
principles in order to obtain more generic conclusions.
B. Generation of Duplicated Samples Each ASV was trained by the following well-known proce-
dure with T randomly selected signatures. Then, the remaining
Two handwritten specimens are never similar among other. signatures are used for testing the False Rejection Rate (FRR).
Personal signatures executed by the same signer have particular To calculate the False Acceptance Rate (FAR), we have used
modifications due to the pose, writing tool, emotional condi- the genuine test samples from all the remaining users. All
tion, etc. So, once the master signature is defined, the synthe- the experiments are repeated 10 times (r = 10). As such,
sizer generates new samples of the same master signature by the FRR is built with (ns − T ) × nu × r and the FAR
some modifications. by (ns − T ) × nu × (nu − 1) × r. Where the number of
On the morphological level, although the name is constant, signatures per user is ns = 24 and 100 the number of total
parameters like skew, slant, average space between letters and users nu . So, in the case of T = 5, FRR is composed
words are slightly varied [27], [2]. Let p be the parameter value of (24 − 5) × 100 × 10 = 19000 scores and the FAR of
and mp be its range (the maximum value minus the minimum). (24 − 5) × 100 × 99 × 10 = 1881000 scores. To imitate several
Then, let s be the stability parameter, the value p is worked out realistic scenarios with few or many enrolled samples, T varies
for every generated sample as p+u, u being a random variable as: (2, 3, 5, 8, 10). Finally, the results are given in terms of
which follows an uniform distribution U(−mp/2, mp/2). If Equal Error Rate (EER) since this represents the operative
the modified value exceed the parameter range it is set up to points where FRR and FAR curves coincide.
the minimum or maximum value, accordingly. Experimental results are shown in Fig. 5 through Detection
On the cognitive level, the intra-personal variability is Error Tradeoff (DET) curves, since they allow to analyze the
due to two parts: 1) We set up a circumference centered in behavior at several FAR and FRR points. We can qualitatively
each grid node, being its radius a source of variability per observe a tendency to achieve similar DET curves between the
signer. In each signature repetition, the grid node is moved synthetic and the real Bengali database. Also, this phenomena
inside such circumference. 2) The signature engram nodes is repeated in both systems (SVM and HMM) reinforcing
of each pen-down are horizontally and vertically displaced the feasibility of the synthesizer to approach such behavior.
a distance randomly chosen between zero and half the grid Moreover, it is worth mentioning that excellent results are
distance between nodes. Such engram nodes refer to the limits achieved when few signatures are used for training. As ad-
of synthetic strokes. 3) Similarly to [19], [20], a sinusoidal ditional observation, it is noted that the DET curves are pretty
transformation is applied to the engram nodes to simulate similar in all operative points in the case of HMM, however
the synchronism variability in the muscles. Let’s assume that higher similarity is seen in SVM for low values in the FAR
(xp , yp ) be the perceptual points coordinates; (Lx , Ly ) be the curves. Regarding performance measures, we could compare
signature width and height; (Ax , Ay ) be the amplitude of sine the EER of real and synthetic signature in each case. Once
and; (Px , Py ) be the period, then this transformation is applied again we could see that better similarities are achieved by
as follows: HMM-based system, especially while dealing with reduced
training sets. Nevertheless, promising results are also obtained
with the SVM, which seems to be much sensitive to our
x̂p = Ax sin (2πPx xp /Lx ) ŷp = Ay sin (2πPy yp /Ly ) (2) synthesizer than HMM. Finally, the standard deviation is also
given in the Fig. 5. It is calculated according to the 10 EER,
obtained in each repetition of the experiment. A low value in
where (x̂p , ŷp ) are the new perceptual point coordinates. all cases is being observed, which suggests the validity of the
The amplitudes are defined randomly between zero and twice EER in all experiments.
the distance between grid points whereas the periods are
random values in the range U (0, 2). V. C ONCLUSION
On the motor level, the parameters Lf and Lw are modified This paper presents a novel solution to generate human-like
as the morphological ones. Visually, Fig. 4 shows artificial synthetic static Bengali signatures. The signature synthesizer is

45
 Fig. 3: Bengali examples of inter-personal variability. Synthetic signatures are market by an asterisk.

Fig. 4: Bengali examples of intra-personal variability. Synthetic signatures are market by an asterisk.

Fig. 5: DET curves to compare the closeness between a real and a synthetic Bengali database with two completely different
ASVs: HMM and SVM.

46
inspired by motor equivalence theory which states that human [9] M. Munich and P. Perona, “Visual identification by signature tracking,”
action can be divided into a cognitive level and a motor level. IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 25,
no. 2, pp. 200–217, 2003.
Briefly, in this work we have redesigned a previous synthe-
[10] J. Galbally, J. Fierrez, M. Martinez-Diaz, and J. Ortega-Garcia, “Im-
sizer [1], [2] which generates static Western signatures. This proving the enrollment in dynamic signature verfication with synthetic
redesign implies major changes in the morphology definition of samples,” in 10th Int. Conf. on Document Analysis and Recogn. (IC-
the language, the introduction of a letter engram and hexagonal DAR), 2009, pp. 1295–1299.
grid to mimic the signature trajectory plan and the filters used [11] M. Song and Z. Sun, “An immune clonal selection algorithm for
to produce the ballistic trajectories. synthetic signature generation,” Mathematical Problems in Engineering,
vol. 2014, pp. 1–12, 2014.
To validate a proper signature generation, we took into [12] M. Diaz, A. Fischer, R. Plamondon, and M. A. Ferrer, “Towards an
account their realistic appearance as the comparison of the automatic on-line signature verifier using only one reference per signer,”
closeness performance regarding a real static Bengali database. in Proc. IAPR Int. Conf. on Document Analysis and Recogn. (ICDAR),
Generating realistic signatures in appearance could be not 2015, pp. 631–635.
mandatory for ASVs. However, keeping such realistic appare- [13] K. Huang and H. Yan, “Off-line signature verification based on ge-
ometric feature extraction and neural network classification,” Pattern
ance in synthetic signatures leads to gain a better understanding Recogn., vol. 30, no. 1, pp. 9 – 17, 1997.
of biological generation processes. [14] B. Fang, C. H. Leung, Y. Y. Tang, P. C. K. Kwok, K. W. Tse, and Y. K.
Due to the lack of research on Bengali script signatures, Wong, “Offline signature verification with generated training samples,”
IEE Proceedings - Vision, Image and Signal Processing, vol. 149, no. 2,
this static Bengali synthesizer emerges as an opportunity pp. 85–90, 2002.
to create datasets to test systems, for security evaluations, [15] E. Frias-Martinez, A. Sanchez, and J. Velez, “Support vector machines
scalability studies, etc. Our future work ideas rely on the multi- versus multi-layer perceptrons for efficient off-line signature recogni-
script generation by redesigning some stages in the synthesizer tion,” Engineering Applications of Artificial Intelligence, vol. 19, no. 6,
if it turns out necessary. Also additional experimental processes pp. 693 – 704, 2006.
could be validated in future research such as visual Turing [16] C. Rabasse, R. Guest, and M. Fairhurst, “A new method for the synthesis
tests, the use of real signatures to train and synthetic to of signature data with natural variability,” IEEE Trans. on Systems, Man,
and Cybernetics, Part B: Cybernetics, vol. 38, no. 3, pp. 691–699, 2008.
test and vice versa, and so on. Finally, research on dynamic
[17] R. Guest, O. Hurtado, and O. Henniger, “Assessment of methods
signature generation is being also taking into account reporting for image recreation from signature time-series data,” IET Biometrics,
promising results. vol. 3, no. 3, pp. 159–166, 2014.
[18] M. Diaz-Cabrera, M. Gomez-Barrero, A. Morales, M. A. Ferrer, and
ACKNOWLEDGMENT J. Galbally, “Generation of enhanced synthetic off-line signatures based
on real on-line data,” in Int. Conf. on Frontiers in Handwriting Recogn.,
M. D. is supported by a PhD fellowship from the ULPGC. ICFHR, 2014, pp. 482–487.
This study was funded by the Spanish governments MCINN [19] M. Diaz-Cabrera, M. Ferrer, and A. Morales, “Cognitive inspired model
TEC2012-38630-C04-02 research project and European Union to generate duplicated static signature images,” in Int. Conf. on Frontiers
FEDER program/funds. in Handwriting Recogn. (ICFHR), 2014, pp. 61–66.
[20] M. Diaz, M. A. Ferrer, G. S. Eskander, and R. Sabourin, “Generation
R EFERENCES of duplicated off-line signature images for verification systems,” IEEE
Trans. Pattern Anal. Mach. Intell., 2016, In press.
[1] M. Ferrer, M. Diaz-Cabrera, and A. Morales, “Static signature synthesis:
A neuromotor inspired approach for biometrics,” IEEE Trans. Pattern [21] D. V. Popel, Signature analysis, verification and synthesis in pervasive
Anal. Mach. Intell., vol. 37, no. 3, pp. 667–680, 2015. environments. World Scientific, 2007, vol. 67, ch. In Synthesis and
Analysis in Biometrics, pp. 31 – 64.
[2] M. A. Ferrer, M. Diaz, C. Carmona, and A. Morales, “A behavioral
handwriting model for static and dynamic signature synthesis,” IEEE [22] M. Ferrer, M. Diaz-Cabrera, and A. Morales, “Synthetic off-line signa-
Trans. Pattern Anal. Mach. Intell., 2016, In press. ture image generation,” in IAPR Int. Conf. on Biometrics (ICB), 2013,
pp. 1–7.
[3] D.-Y. Yeung, H. Chang, Y. Xiong, S. George, R. Kashi, T. Matsumoto,
and G. Rigoll, “SVC2004: First international signature verification [23] M. Ismail and S. Gad, “Off-line arabic signature recogn. and verifica-
competition,” in Biometric Authentication, ser. LNCS. Springer, 2004, tion,” Pattern Recogn., vol. 33, no. 10, pp. 1727 – 1740, 2000.
vol. 3072, pp. 16–22. [24] S. Pal, “Multi-script off-line signature verification,” Ph.D. dissertation,
[4] M. Liwicki, M. I. Malik, C. E. v. d. Heuvel, X. Chen, C. Berger, Griffith University, October 2014.
R. Stoel, M. Blumenstein, and B. Found, “Signature verification com- [25] A. M. Wing, “Motor control: Mechanisms of motor equivalence in
petition for online and offline skilled forgeries (SigComp2011),” in Int. handwriting,” Current Biology, vol. 10, no. 6, pp. R245 – R248, 2000.
Conf. on Document Analysis and Recogn. (ICDAR), 2011, pp. 1480– [26] A. Marcelli, A. Parziale, and R. Senatore, “Some observations on
1484. handwriting from a motor learning perspective.” in 2nd Workshop on
[5] M. I. Malik, S. Ahmed, A. Marcelli, U. Pal, M. Blumenstein, L. Alewi- Automated Forensic Handwriting Analysis (AFHA), 2013, pp. 6–10.
jns, and M. Liwicki, “ICDAR2015 competition on signature verifi- [27] M. Diaz-Cabrera, M. A. Ferrer, and A. Morales, “Modeling the lexical
cation and writer identification for on- and off-line skilled forgeries morphology of western handwritten signatures,” PLoS ONE, vol. 10,
(SigWIcomp2015),” in Int. Conf. on Document Analysis and Recogn. no. 4, p. e0123254, 2015.
(ICDAR), 2015, pp. 1186–1190. [28] M. Ferrer, J. Alonso, and C. Travieso, “Offline geometric parameters
[6] M. Diaz-Cabrera, A. Morales, and M. A. Ferrer, “Emerging issues for automatic signature verification using fixed-point arithmetic,” IEEE
for static handwritten signature biometric,” in Advances in Digital Trans. on Pattern Analysis and Machine Intelligence, vol. 27, no. 6, pp.
Handwritten Signature Processing. A Human Artefact for e-Society, 993–997, 2005.
2014, pp. 111–122. [29] M. Ferrer, J. Vargas, A. Morales, and A. Ordonez, “Robustness of offline
[7] J.Galbally, R. Plamondon, J. Fierrez, and J. Ortega-Garcia, “Synthetic signature verification based on gray level features,” IEEE Trans. on
on-line signature generation. Part I: Methodology and algorithms,” Information Forensics and Security, vol. 7, no. 3, pp. 966–977, 2012.
Pattern Recogn., vol. 45, pp. 2610–2621, 2012.
[8] J.Galbally, J. Fierrez, J. Ortega-Garcia, and R. Plamondon, “Synthetic
on-line signature generation. Part II: Experimental validation,” Pattern
Recogn., vol. 45, pp. 2622–2632, 2012.

47