ICDAR2015 Competition on Signature Verification and Writer Identification for On- and Off-line Skilled Forgeries ( SigWIcomp2015 )

Abstract

This paper presents the results of the ICDAR 2015 competition on signature verification and writer identification for on- and off-line skilled forgeries jointly organized by PRresearchers and Forensic Handwriting Examiners (FHEs). The aim is to bridge the gap between recent technological developments and forensic casework. Two modalities (signatures and handwritten text) are considered and training and evaluation data are collected and provided by FHEs and PR-researchers. Four tasks are defined for four different languages; Bengali offline signature verification, Italian off-line signature verification, German on-line signature verification, and English handwritten text based writer identification. In total, 40 systems have participated in this competition. The participants of the signatures modality were motivated to report their results in Likelihood Ratios (LRs). This has made the systems even more interesting for application in forensic casework. For evaluating the performance of the systems, we have used the forensically substantial Cost of Log Likelihood Ratios ( bCllr) in the case of signatures, and the F-measure in the case of handwritten text.

Full text

ICDAR2015 Competition on Signature Verification
and Writer Identification for On- and Off-line
Skilled Forgeries (SigWIcomp2015)
Muhammad Imran Malik
∗
, Sheraz Ahmed
∗
, Angelo Marcelli
†
,
Umapada Pal
‡
, Michael Blumenstein
§
, Linda Alewijns
¶
, Marcus Liwicki
k
∗
German Research Center for AI (DFKI GmbH), Kaisersautern, Germany (firstname.lastname@dfki.de)
†
University of Salerno Via Giovanni Paolo II, Fisciano (SA), Italy (amarcelli@unisa.it)
‡
CVPR Unit, Indian Statistical Institute, Kolkata, India (umapada@isical.ac.in)
§
School of ICT, Griffith University, Australia (m.blumenstein@griffith.edu.au)
¶
Netherlands Forensic Institute. The Hague, The Netherlands. (l.alewijnse@nfi.minvenj.nl)
k
University of Fribourg, Switzerland (marcus.liwicki@unifr.ch)
Abstract—This paper presents the results of the ICDAR 2015
competition on signature verification and writer identification
for on- and off-line skilled forgeries jointly organized by PR-
researchers and Forensic Handwriting Examiners (FHEs). The
aim is to bridge the gap between recent technological develop-
ments and forensic casework. Two modalities (signatures and
handwritten text) are considered and training and evaluation
data are collected and provided by FHEs and PR-researchers.
Four tasks are defined for four different languages; Bengali off-
line signature verification, Italian off-line signature verification,
German on-line signature verification, and English handwritten
text based writer identification. In total, 40 systems have par-
ticipated in this competition. The participants of the signatures
modality were motivated to report their results in Likelihood
Ratios (LRs). This has made the systems even more interesting for
application in forensic casework. For evaluating the performance
of the systems, we have used the forensically substantial Cost of
Log Likelihood Ratios (
b
C
llr
) in the case of signatures, and the
F-measure in the case of handwritten text.
I. INTRODUCTION
Most of the current research in the field of automatic
signature verification does not take the real needs of Forensic
Handwriting Experts (FHEs) into account [1], [3], [4], [5]. In
their casework, FHEs often need to have verification results in
the form of likelihood ratios. This is very important as it allows
one to combine the FHE’s evidence (from the results of an au-
tomated system) with other evidence presented in the courts of
law [6], [7]. We have organized the ICDAR SigWIcomp2015
competition where we asked the participants to produce a
comparison score (e.g., a degree of similarity or difference),
and the evidential value of that score, expressed as the ratio
of the probabilities of finding that score when the questioned
signature is a genuine signature and when it is a forgery (i.e.,
the likelihood ratio). Note that by these competitions we are
further strengthening a paradigm shift in automatic signature
verification (introduced in the SigComp 2011, and tested
heavily in 4NSigComp 2012 and SigWiComp2013) from the
“decision paradigm” to an “evidential value paradigm”. For
more details on this, please refer to [8], [9], [10].
TABLE I
ITALIAN OFF-LINE SIGNATURE DATA
Data Authors Genuine Forged Total
Training 50 250 0 250
Testing 50 229 249 478
TABLE II
BENGALI OFF-LINE SIGNATURE DATA
Data Authors Genuine Forged Total
Training 10 120 0 120
Testing 10 120 300 420
Furthermore, we have focused on the writer identification
and retrieval task from the view point of FHEs. Here the
defined task is: Given a handwritten text from an author,
retrieve all the texts written in different writing styles by the
same author. This is also very important when viewed from
the perspective of FHEs. They often have the scenarios where
they are provided with a piece of text handwritten by an author
in a different style than those texts which are present in the
dataset and are also written by the same author. This is a
usual situation when the same person, for example, writes
threat letters to more than one persons and tries to change
her/his handwriting style every time. The task here is to link
the different handwriting styles from the same writer; which
in fact is a real challenge so far. For evaluation of this task,
we used standard precision and recall measures and calculated
F-measure to declare the winning system.
II. DATA
We have the signature data available in Bengali, Italian,
and German and handwriting data in different writing styles
in English only. The details follow.
A. Italian Off-line Signatures
The Italian off-line signatures were collected at the Univer-
sity of Salerno from the University employees and students.

TABLE III
GERMAN ON-LINE SIGNATURE DATA
Data Authors Genuine Forged Total
Training 30 300 0 300
Testing 30 150 300 450
TABLE IV
ENGLISH OFF-LINE HANDWRITTEN TEXT DATA
Data Authors Pages/Author Total
Training 55 3 165
Testing 55 3 165
Two unique aspects of these data are; these are actual signa-
tures that were provided by individuals while filling various
forms and applications at the University, and the signatures
collection span over a period of time that is between 3 and 5
years depending on the subject. The forgeries are also made
by students where they were allowed to practice the forgery
as many times as they liked and then they produced skilled
forgeries. For training, a set of 50 specimen genuine authors,
with 5 reference signatures from each specimen author, was
provided to the participants in binarized form. For evaluation,
we used the same 50 specimen authors; with 10 correspond-
ing questioned signatures each: containing genuine signatures
and skilled forgeries, in binarized form. Detailed breakup is
provided in Table I.
B. Bengali Off-line Signatures
The signatures were collected from different parts of the
West Bengal, a state of India. The majority of the signatures
were contributed by students. A total number of 240 genuine
signatures were collected from 10 contributor (24 genuine sig-
natures by each). For each contributor, all genuine specimens
were collected in a single day’s writing session. In order to
produce the forgeries, the imitators were allowed to practice
forgeries as long as they wished. A total number of 300 (30
signatures, 10 individuals) forged signatures were collected.
The images were captured in 256 level Grey scale at 300 dpi
and stored in TIFF format (Tagged Image File Format). A
detailed breakup is provided in Table II.
C. German On-line Signatures
The German on-line signatures were collected at the Ger-
man Research center for Artificial Intelligence, Germany. A
unique aspect of these data is that the data have been collected
using a digitized pen rather than a tablet, i.e., by Anoto
Pen [11]. This pen specializes in providing the look and feel
of regular pens. It only demands to add Anoto dot pattern to
any paper and data can be digitized seamlessly. The Anoto
pattern makes it possible for the Anoto pen built-in camera to
detect strokes and record signatures. The signature data were
collected from employees of different financial institutions and
students of University of Kaiserslautern who also participated
in generation of skilled forgeries. The dataset consists of
ASCII files with the format: X, Y, and Pressure with sampling
rate of 75 Hz and resolution of 85 dpi. For training, 30
specimen genuine authors were selected, from whom each
one provided 10 genuine reference signatures. For evaluation,
data from the same 30 specimen genuine authors; with 15
corresponding questioned signatures each: containing genuine
signatures and skilled forgeries were used. Detailed breakup
is provided in Table III.
Note that for all of the signature verification tasks, we did
not provide any forgeries for training (see Tables I, II, and III).
This is important as; In the real world, one can never limit
the forgery set since every signature other than the genuine
signature written naturally by an authentic author can be a
forgery, and also when forgeries are used for training–there is
always a chance that an automatic system may learn to declare
signatures as forgeries when they are coming from the forgers
on whom the system is trained [12], [13].
D. English Handwritten Text Samples
The handwritten text samples for the writer identification
(based on different handwriting styles) task were off-line and
in English only. They were collected and prepared by the
Netherlands Forensic Institute for this competition/task. A pre-
printed paper was used with horizontal lines for writing. The
pre-printed paper was placed underneath the blank writing
paper. Four extra blank pages were added underneath the first
two pages to obtain a soft writing surface. The writings were
scanned at 400 dpi using the EPSON Expression 10000XL
scanner, with RGB color and saved as PNG images. Files were
randomly numbered, so that file numbers do not link to the
writers. The training set comprised of 165 samples written
by all the 55 authors in 3 different handwriting styles. The
evaluation set comprised of 165 samples written by the same
55 authors in 3 different handwriting styles (other than the
styles available in the training set). A detailed breakdown is
provided in Table IV.
III. SUBMITTED SYSTEMS
50 systems initially registered for this competition, how-
ever, 40 systems from 13 different institutions eventually
participated. The details follow: 9 systems for off-line Italian
signature verification, the same 9 systems optimized for off-
line Bengali signature verification, 12 system for the on-line
German signature verification, and 10 systems for the writer
identification task. Table V provides details about affiliations
of the participants. In the following we will describe each of
these systems briefly by providing references so that inter-
ested readers may follow. Note that some of the participants
preferred remaining anonymous after the results were declared.
A. Off-line Verification Tasks: Italian and Bengali
System 1: This system uses histogram of oriented gradients
(HOG) and local binary patterns (LBP) extracted from local
regions, together with the user-based and global SVM clas-
sifiers and a sophisticated classifier combination. The basic
system description can be found in [14].

TABLE V
OVERVIEW OF THE SUBMITTED SYSTEMS
System Modality Participant Mode
1 Signatures Sabanci University, Turkey Off-line
2-6 Signatures Tebessa University, Algeria Off-line
7,8 Signatures Qatar University, Qatar Off-line
9 Signatures Commercial System Off-line
10-14 Signatures Anonymous On-line
15 Signatures Anonymous On-line
16 Signatures Bahria University, Pakistan On-line
17-18 Signatures Cursor Insight On-line
19 Signatures Commercial System On-line
20 Signatures Qatar University, Qatar On-line
21
Signatures Sabanci University, Turkey On-line
22-26 Handwriting Tebessa University, Algeria Off-line
27-31 Handwriting Qatar University, Qatar Off-line
TABLE VI
RESULTS FOR TASK 1: ITALIAN OFF-LINE SIGNATURE VERIFICATION
System Participant
b
C
llr
b
C
min
llr
1 Sabanci University 0.655109 0.021358
2 Tebessa University 0.993189 0.893270
3 Tebessa University 1.065696 0.952499
4 Tebessa University 1.074474 0.880930
5 Tebessa University 1.065475 0.901003
6 Tebessa University 1.041895 0.901003
7 Qatar University 8.901864 0.972708
8 Qatar University 13.111064 0.960163
9 Commercial System 1.003786 0.988845
System 2: This system is based on edge directional based
features. The method starts with conventional edge detection
that generates a binary image in which only the edge pixels
are ”on”. Then each edge pixel is considered in the middle of
a square neighborhood and checked in all directions emerging
from the central pixel and ending on the periphery of the
neighborhood for the presence of an entire edge fragment.
For comparison purposes, the Manhattan Distance Metric is
used.
System 3: This system is based on the edge-hinge features
which estimate the joint distribution of edge angles in a
writer’s handwriting. They are constructed by performing an
edge detection using a Sobel kernel on the input images, and
subsequently, measuring the angles of both edge segments that
emanate from each edge pixel.
System 4: The system is based on multi-scale run length
features which are determined on the binary image taking into
consideration both the black pixels corresponding to the ink
trace and the white pixels corresponding to the background.
For details, refer to [15].
System 5: The system is based on the combination of
both types of features used by the previous two methods:
multi-scale edge-hinge features and multi-scale run-length
features [16].
System 6: This system is based on the combination of three
types of features used by the systems 2, 3, and 4: multi-scale
edge-hinge features, multi-scale run-length features and the
edge based directional features.
System 7 and System 8: Both of these system combine
through a logistic regression classifier the geometrical features
TABLE VII
RESULTS FOR TASK 2: BENGALI OFF-LINE SIGNATURE VERIFICATION
System Participant
b
C
llr
b
C
min
llr
1 Sabanci University 0.68895 0.052485
2 Tebessa University 0.933750 0.154390
3 Tebessa University 0.939850 0.116541
4 Tebessa University 0.923886 0.039721
5 Tebessa University 0.931253 0.060248
6 Tebessa University 0.929922 0.055556
7 Qatar University 1.161387 0.973043
8 Qatar University 2.839290 0.297842
9 Commercial System 0.998630 0.893893
described in [17]. Those features are based on number of
holes, moments, projections, distributions, position of barycen-
ter, number of branches in the skeleton, Fourier descriptors,
tortuosities, directions, curvatures, chain codes and edge based
directional features. The two systems differ in weighting the
features they are built on.
System 9: This is a commercial system and the participating
organization has requested to not mention their identity. The
following basic details are provided about this system. The
system has four main modules. The first one identify and
extracts the signature from the image page. The second module
applies filtering and other preprocessing on the extracted
signature. The third module extracts various spatial, geometric,
morphological, and statistical features. The last module allows
the combination of multiple classifiers and uses DTW, Pearson
Correlation, and Euclidean distance combined by an MLP
neural network to make the final decision.
The same nine systems (Systems 1-9) have participated for
both Italian and Bengali tasks (when optimized for the specific
task), their details are therefore omitted.
B. On-line Verification Task: German
System 10-14: Systems 10 to 14 are from the same aca-
demic organization and they have requested to not mention
their identity. All of these systems are based on DTW on
various time functions derived from the position trajectory and
the pressure information. These 5 systems differ in the specific
functions considered and the type of score normalization
applied.
System 15: The following details are available about the
working of this system, though the participating organization
has chosen to remain anonymous. This system is based on
the DTW algorithm and uses both global features (e.g., total
writing time) and local features (e.g., pressure in a given
point). If the total writing time of the signature to be verified is
acceptable, for each local feature f (such as X-Y coordinates,
pressure), the DTW distance between the time series related
to the authentic signature and to the signature to be verified
is computed. Then, DTW distances are combined with a
weighted sum, by giving a weight to each of the different kinds
of features. The similarity score is computed accordingly.
System 16: In this system, multiple time series represen-
tation of signature including time series of X-coordinate, Y-
Coordinate, Pressure, speed and centroid distance are com-

TABLE VIII
RESULTS FOR TASK 3: GERMAN ON-LINE SIGNATURE VERIFICATION
System Participant
b
C
llr
b
C
min
llr
10 Anonymous 0.948878 0.680014
11 Anonymous 0.944402 0.688105
12 Anonymous 0.855595 0.576762
13 Anonymous 0.874655 0.577098
14 Anonymous 1.262207 0.684602
15 Anonymous 0.996586 0.839468
16 Bahria University 1.025046 0.772073
17 Cursor Insight 1.243891 0.290158
18 Cursor Insight 1.496257 0.540563
19 Commercial System 0.807276 0.465681
20 Qatar University 3.345016 0.854712
21 Sabanci University 0.873383 0.304159
puted. Verification of genuine and forged signature is done
using adaptation of multivariate m Mediods based classifica-
tion and anomaly detection approach as presented in [18].
System 17-18: Both of these systems are commercial
products designed by Cursor Insight. These systems specialize
in calculating more than 70000 movement characteristics of
each handwritten sample. For further details, refer to Cursor
Insight
1
.
System 19: This system compares the signatures using
statistical data as well as a time based model. It extracts
the changes of several dynamic features (direction, pressure
and speed) over time. The statistical data contains static
information like duration of the signature or duration and
time of pen liftings. The reference signatures are compared
against each other to define the validity space of the model.
The similarities of the features are combined into a final score
using different weightings.
System 20: This system computes differences between the
features of the questioned signature and the reference signa-
tures at the signal level as well as the histogram level [19].
Those features include x and y coordinates, pressure, direc-
tions, angles, speed and angular speed.
System 21: This system only uses the ’x’ and ’y’ coordi-
nates as features and Dynamic Time Warping (DTW) as the
matching algorithm. Normalization of the query scores are
done based on reference signature statistics. Details of this
system can be found in [20].
C. Off-line Writer Identification Task: English
Systems 22-26: By the same participant (who provided
Systems 2-6), with the same methodology, as explained earlier,
optimized to consider handwriting.
Systems 27-31: The proposed methods combine through a
logistic regression classifier the geometrical features. Those
features are based on tortuosities, directions, curvatures, chain
codes and edge based directional features. Further details are
available in [17], [19].
IV. EXPERIMENTS AND EVALUATION
We defined two modalities, i.e., signatures and handwritten
text, and the following four tasks for this competition,
1
www.cursorinsight.com
TABLE IX
RESULTS FOR TASK 4: WRITER IDENTIFICATION BASED ON DIFFERENT
HANDWRITING STYLES.
System Participant Avg. F1-measure(%)
22 Tebessa University 17.37
23 Tebessa University 33.54
24 Tebessa University 32.53
25 Tebessa University 33.94
26 Tebessa University 33.54
27 Qatar University 30.71
28 Qatar University 21.01
29 Qatar University 21.01
30 Qatar University 20.80
31
Qatar University 19.60
• Task 1: Italian off-line signature verification
• Task 2: Bengali off-line signature verification
• Task 3: German on-line signature verification
• Task 4: Writer identification based on handwriting styles
Accordingly, the evaluations are made separately for each
of these tasks. In the case of signature verification tasks (the
first three: both on-line and off-line), the task was to determine
if a given questioned signature has been written by the author
of the n reference specimen signatures or if it was forged by
another writer. We evaluated the signature verification systems
according to the cost of the log-likelihood ratios
b
C
llr
using the
FoCal toolkit [21], and finally the minimal possible value of
b
C
llr
, i.e.,
b
C
min
llr
, as the final assessment value. Note that a
smaller value of
b
C
min
llr
denotes a better performance of the
method. An interesting observation about the
b
C
llr
and
b
C
min
llr
is that it is not only based on the number of errors made by
a system but also looks into the severity of errors by warping
the scores of automatic systems. This makes the metric (
b
C
min
llr
)
well suited for forensic applications where a severely mistaken
system (although may be having a very low error rate) may
lead a person to death, therefore considering the severity of
errors is substantial for forensic casework. We have focused
on this issue in the previous competitions of this chain, i.e., at
ICDAR 2011 and ICDAR 2103. For more details about this,
refer to [10].
Further important observations about the signature verifica-
tion tasks for different data are: For Italian off-line signatures,
the system using a combination of local and global features
outperformed all other systems (mostly based on global fea-
tures) by a large margin; for Bengali off-line signatures, most
of the systems performed good and there was not much
variance among the results which might indicate that the
systems which were originally designed for western languages
were able to adapt well on an Indic language provided better
training; for German on-line signatures, the systems had to
face difficulties as these data were collected by specialized
electronic pen-Anoto. The Anoto pen also records position and
pressure similar to standard data capturing tablets, however
with different internal data capturing mechanism. Since, the
use of electronic pens is increasing, the PR research should
also consider these data and optimize the systems accordingly.

The results of the first three tasks are given in Ta-
bles VI, VII, and VIII, respectively. As can be seen, different
systems performed better on different tasks. The winner of the
Italian off-line signature verification task, Task 1, is system 1
from Sabanci University, Turkey. The winner of the Bengali
off-line signature verification task, Task 2, is system 4 from
Tebessa University, Algeria. The winner of the German on-line
signature verification task, Task 3, is system 17 from Cursor
Insight (a company).
The results of the fourth task, i.e., writer identification
by linking various handwriting styles of the same writer are
provided in Table IX. In this task we used the standard
precision and recall measures to evaluate the systems and final
ranking is based on the best average F-measure value [22].
Note that in this case the values of precision and recall were
the same (and thereby the F-measure value–so only F-measure
value is given): we used the F1-measure, the harmonic mean
of precision and recall, where equal emphasis is given to both
the precision and recall. This is due to the scenario for the said
competition where for every author 3 documents handwritten
in different styles were available in the repository of all the
documents and the system must also retrieve only the 3 top
matching handwritten documents from that repository.
Furthermore, as given in Table IX, the systems are quite
low in performance on the F-measure scale. This represents
that the given task was really challenging and still requires a
lot of efforts from the PR-researchers. Nonetheless, System 25
performed comparatively better than the other systems and is
declared the winner for this task.
V. CONCLUSIONS AND OUTLOOK
In this paper we presented the results of the ICDAR2015
competition on signature verification and writer identification
for on- and off-line skilled forgeries (SigWiComp2015) jointly
organized by PR-researchers and FHEs. Note that it is a
continuation of previous signature verification competitions,
i.e., SigComp2011, 4NSigComp2012, and SigWiComp2013.
In the SigWiComp2015, we have defined three tasks for off-
line and on-line signature verification and provided signature
data (both on- and off-line) in multiple languages (Italian,
Bengali, and German). Furthermore, for evaluation, we rein-
forced our paradigm based on likelihood computations via the
b
C
llr
and
b
C
min
llr
values which are forensically substantial. In
addition to that, in the SigWiComp2015, we also defined one
task for writer identification and retrieval based on different
handwriting styles. The inclusion of this task has made the Sig-
WiComp2015 even more forensically relevant and interesting
for both the FHEs and PR-researcher.
In the future, we plan to make the datasets even larger
and more diverse. We also plan to include signature samples
written in different languages (bi-lingual signatures). Another
important aspect is the usability and application of automatic
systems in real forensic cases, which motivates us to evaluate
the usability of automatic systems as well. We are planning to
organize competitions focusing on the real world usability of
automatic verification systems in the future.
VI. ACKNOWLEDGEMENT
We are thankful to Antonio Parziale and Srikanta Pal who
provided us access to the signature data for SigWIcomp2015.
REFERENCES
[1] R. Plamondon and G. Lorette, “Automatic signature verification and
writer identification – the state of the art,” Pattern Recognition, vol. 22,
pp. 107–131, 1989.
[2] F. Leclerc and R. Plamondon, “Automatic signature verification: the state
of the art 1989–1993,” in Progress in Automatic Signature Verification,
R. Plamondon, Ed. World Scientific Publ. Co., 1994, pp. 13–19.
[3] R. Plamondon and S. N. Srihari, “On-line and off-line handwriting
recognition: a comprehensive survey,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 22, no. 1, pp. 63–84, 2000.
[4] D. Impedovo and G. Pirlo, “Automatic signature verification: The state
of the art,” IEEE Transactions on Systems, Man, and Cybernetics, Part
C (Applications and Reviews), vol. 38, no. 5, pp. 609–635, Sep. 2008.
[5] S. Ahmed, M. I. Malik, M. Liwicki, and A. Dengel, “Signature segmen-
tation from document images,” in ICFHR. IEEE, 2012, 423-427.
[6] J. Gonzalez-Rodriguez, J. Fierrez-Aguilar, D. Ramos-Castro, and
J. Ortega-Garcia, “Bayesian analysis of fingerprint, face and signature
evidences with automatic biometric systems,” Forensic Science Interna-
tional, vol. 155, no. 2-3, pp. 126–140, 2005.
[7] M. I. Malik and M. Liwicki, “From terminology to evaluation: Per-
formance assessment of automatic signature verification systems,” in
ICFHR. IEEE, 2012, pp. 613–618.
[8] M. Liwicki, M. I. Malik, C. E. van den Heuvel, X. Chen, C. Berger,
R. Stoel, M. Blumenstein, and B. Found, “Signature verification compe-
tition for online and offline skilled forgeries SigComp2011,” in ICDAR.
IEEE, 2011, pp. 1480–1484.
[9] M. Liwicki, M. I. Malik, L. Alewijnse, C. E. van den Heuvel, and
B. Found, “Icfhr2012 competition on automatic forensic signature ver-
ification 4NsigComp 2012,” in ICFHR. IEEE, 2012, pp. 780–785.
[10] M. I. Malik, M. Liwicki, L. Alewijnse, W. Ohyama, M. Blumenstein,
and B. Found, “ICDAR 2013 competitions on signature verification and
writer identification for on-and offline skilled forgeries (SigWiComp
2013),” in ICDAR. IEEE, 2013, pp. 1477–1483.
[11] M. I. Malik, S. Ahmed, A. Dengel, and M. Liwicki, “A signature
verification framework for digital pen applications,” in DAS. IEEE,
2012, pp. 419–423.
[12] M. I. Malik, L. Alewijnse, M. Liwicki, and M. Blumenstein, “Signa-
ture verification tutorial,” 2nd International Workshop and Tutorial on
Automated Forensic Handwriting Analysis (AFHA), 2013.
[13] M. Liwicki and M. I. Malik, “Surprising? power of local features for
automated signature verification,” in 15th IGS Conf. International
Graphonomics Society, 2011, pp. 18–21.
[14] M. B. Yilmaz, B. Yanikoglu, C. Tirkaz, and A. Kholmatov, “Offline sig-
nature verification using classifier combination of hog and lbp features,”
in Int. Joint Conf. on Biometrics, 2011, pp. 1–7.
[15] C. Djeddi, I. Siddiqi, L. Souici-Meslati, and A. Ennaji, “Text-
independent writer recognition using multi-script handwritten texts,”
Pattern Recognition Letters, vol. In press, 2013.
[16] C. Djeddi, L. Souici-Meslati, and A. Ennaji, “Writer recognition on
arabic handwritten documents,” in ICISP, 2012, pp. 493–501.
[17] A. Hassaine, S. Al-Maadeed, and A. Bouridane, “A set of geometri-
cal features for writer identification,” Neural Information Processing.
Springer Berlin Heidelberg, 2012.
[18] S. Khalid and S. Razzaq, “Frameworks for multivariate m-mediods based
modeling and classification in euclidean and general feature spaces,”
Pattern Recogn., vol. 45, no. 3, pp. 1092–1103, 2012.
[19] A. Hassane and S. Al-Maadeed, “An online signature verification system
for forgery and disguise detection,” Neural Information Processing.
Springer Berlin Heidelberg, 2012.
[20] A. Kholmatov and B. Yanikoglu, “Identity authentication using im-
proved online signature verification method,” Pattern Recognition Let-
ters, vol. 26, no. 15, pp. 2400–2408, 2005.
[21] N. Br
¨
ummer and J. du Preez, “Application-independent evaluation of
speaker detection,” Computer Speech & Language, vol. 20, no. 2-3, pp.
230–275, 2006.
[22] T. Fawcett, “An introduction to ROC analysis,” Pattern Recogn. Lett.,
vol. 27, pp. 861–874, June 2006.