See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4153574

Signature Verification and Forgery Detection System

Conference Paper · September 2003

DOI: 10.1109/SCORED.2003.1459654 · Source: IEEE Xplore

CITATIONS READS

12 7,153

2 authors:

Mohd Hafizuddin Mohd Yusof Vamsi Krishna Madasu

Multimedia University The University of Queensland
12 PUBLICATIONS 248 CITATIONS 59 PUBLICATIONS 977 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mohd Hafizuddin Mohd Yusof on 29 January 2014.

The user has requested enhancement of the downloaded file.

Student Conference on Research and Development (SCOReD) 2003 Proceedings, Putrajaya, Malaysia

Signature Verification and Forgery Detection System

Mohd Hafizuddin Mohd Yusof' and Vamsi Krishna Madasu'

'Faculty of Information Technology, Multimedia University, 63 100 Cyberjaya,
Selangor, Malaysia
'School of Infosation TechnoIogy and Electrical Eng.,
University of Queensland, Brisbane, QLD 4072, Australia
e-niail: ha l i z u ~ l d i i i . v u s o l i i r ! n a n u . i ~madasii(crlitcc.tiq.~~iii.~ii
~~,

Abstruct- This paper presents an innovative approach verification systems designed up to 1993 appear in [ 1,2,
for signature verification and forgery detection based on and 33. Another survey article [4] has summarized the
fuzzy modeling. The signature images are binarized and approaches used for off-line signature verification from
resized to a fixed size window and are then thinned. The 1993-2000. We present here a review of a few papers in
thinned image is then partitioned into a fixed number of this field, which have not been covered in the survey
eight sub-images called boxes. This partition is done articles. The emphasis of these papers is mainly on
using the horizontal density approximation approach. fuzzy based techniques for off-line signature
Each sub-image is then further resized and again verification.
partitioned into twelve further sub-images using the
uniform partitioning' approach. The features of An off-line signature system consisting of signature
consideration are normalized vector angle (a) and recognition and verification is proposed in [ 5 ] . In this,
distance (y) from each box. Each feature extracted from the recognition phase is based on the multi-stage
sample signatures gives rise to fuzzy sets. Since the classifier and a combination of global and local features
choice of a proper fuzzification hnction is crucial for whereas the verification is done using fuzzy concepts.
verification, we have devised a new fuzzification HMM based approach in [6] derives dynamically and
function with structural parameters, which is able to automatically the author dependent parameters to set up
adapt to the variations in hzzy sets. This function is an optimal decision rule for off-line verification process.
employed to develop a complete forgery detection and Here the cross validation principle is used to derive not
verification system, only the best HMM models, but aho an optimal
acceptation/ rejection threshold for each author. This
Keywords threshold leads to a high discrimination between the
Signature verification, Forgery detection, Box-method, authors and impostors in the context of random
Fuzzy modeiing. forgeries.

I. INTRODUCTION Signature verification is also attempted using the

Pseudo- Bacterial Genetic Algorithm (PBGA) [ 141
Signature verification and forgeIy detection relate to the which was applied for the discovery of fuzzy rules. The
process of verifying signatures automatically and rules are units themselves and they are constituted by
instantly to determine whether the signature is genuine several parameters to be optimized, however, the
or forged. There are two main types of signature performance of a fuzzy system is obtained
verification: static and dynamic. Static, or off-line synergistically as a sum of the outputs of several rules.
verification is the process of verifying an electronic or The PBGA was then applied for the extraction of
paper signature after it has been made, while dynamic or personal features for signature verification,
on-line verification takes place as a subject creates his
signature on a digital tablet or a similar device. The A pseudo-outer product based fuzzy neurat network
signature in question is then compared to previous drives the signature verification system in [7]. This
samples of the signer's signattire, which constitute the system is primarily used for verifying skilled forgeries.
database or knowledge base. In the case of an ink Signature verification using TS model is reported in [PI
signature on paper, the computer requires the sample to and features for this model are drawn from the box
be scanned for analysis, whereas a digital signature is approach o f [SI. In the present work, we follow the
already stored in a data format that signature same features as in [SI but the TS model is modified to
verification can use. enhance its capability for the detection of forgeries.

The design of any signature verification system Automatic examination of questioned signatures did not
generally requires the solution of five sub-problem: come into being until the advent of computers in the
data acquisition, pre-processing, feature extraction, 1960s. As computer system became more powerful and
comparison process and performance evaluation [ 11. more affordable, designing an automatic forgery
Surveys of the state of the art off-line signature detection system became an active research subject.
Most of the work in off-line forgery detection, however,

0-7803-8 173-4/03/$17.00 02003 IEEE. 9

 has been on random or simple forgeries and less on 111. FEATURE EXTRACTION
skilled or simulated forgeries. Before looking into the
landmark contributions in the area of forgery detection, The pre-processed image is then partitioned into eight
we briefly explain the types of forgeries. ,
portions using the equal horizontal density method. In
this method, the bioarized image is scanned horizontally
The forgeries involved in handwritten signatures have from left to right and then from right to left and the total
been categorized based on their characteristic features number of dark pixels is obtained over the entire image.
f2]. The forgeries in signatures can be broadly
' The pixels are clustered into eight regions such that
classified into the following types: approximately equal number of dark pixels falls in each
1. . Rnndom Forgev - The signer uses the name of the region. This process is illustrated in Fig 3 and explained
'
victim in his own style to create a forgery known as in the following paragraphs,
the simple forgery or random forgery.
2. Unskilled Forgery - The signer imitates the Here, the total number of points (dark pixels) is 48. If
signature in his own style without any knowledge we partition these pixels into 4, we get 48/4=12 pixels
of the spelling and does not have any prior per partition. Since the partition is done column wise,
experience. getting exactly 12 points in each partition is quite
3. Skilled Forgery - undoubtedly professional impossible. Therefore, we will take approximately 12
impostors or persons who have experience in points in each partition using two-way scanning
copying the signature create the most difficult of all approach as described below. We scan the image from
forgeries. left to right till we reach the column where the number
of points in a particular partition is 12 or more. We then
Xn the 198O's, Ammar et al. [IO] worked on the repeat the same procedure while scanning the image
detection of skilled forgeries. They have calculated the from right Wleft direction. For left to right, we partition
'
statistics o f dark pixels and used them to identify the image and then for right to left we partition the, We
changes in the global flow of the writing. The later work then take the average of two cob" numbers in each
of Ammar [111 is based on reference patterns, namely partition
the horizonta1 and vertical positions of the signature
image. The projection of the questioned signature and Each partition is now resized to a fixed window (box of
the reference are compared using Euclidean distance. size 38 x 60 pixels) size and is thinned again. Each box
Guo et al. [12] have presented an algorithm for the is again partitioned into 4 rows x 3 columns,
detection of skdled forgeries based on a local constituting 12 boxes. In total we have 96 partitions for
correspondence between a questioned signature and a a single signature. We have considered only the angles
model obtained a priori. Writer-dependent properties are in the present work. We calculate the summation of the
measured at the sub-stroke level and a cost function is angles of all points in each box with respect to the
trained for each writer. bottom left corner of the box. The summation of angles
is normalized with the number of pixels in the box.
The original scanned signatures are pre-processed These angles constitute the feature database for a given
involving size normalization, binarization and thinning signature.
before features are extracted from each of them. These
features constitute the howledge base, which is then 111. PROPOSED SYSTEM
used for verifymg the genuine signatures and detecting
the forgeries. We now briefly explain the various stages Since the main thrust here is to establish the
in the signature verification system. genuineness of the signature thereby detecting the
forgeries, we go in for fuzzy modeling of angle features.
For the purpose of signature verification and detection
IL PRE-PROCESSING OF SIGNATURES of forgeries, we have employed the Takagi-Sugeno
model. In this, we are following the same concept as
The signatures were handwritten on a white sheet of outlined in [ 131 for considering each feature as forming
paper, using a black pen. The signature'images were a fuzzy set over large samples. This is because the same
then scanned at a resolution of 200 dpi and resized by feature exhibits variation in different samples giving rise
50% using a B-Spline filter. Some typical signatures to a fuzzy set. So, our attempt is to model the
along with their forgeries are given in Fig.1. uncertainty through a fuzzy model such as the TS
model.
Pre-processing of scanned signatures is necessary before
feature extraction. In the present system, all the Let .xk be. the kth feature in a fuzzy set Ak , so the kth
signature images are fist resized to a fixed window of
fuzzy rule IF THEN rule in TS model has the following
size (I20 x 60 pixels), then binarized and thinned using
form
the modified SPTA thinning algorithm [9]. Features are
then extracted fkom this pre-processed signature image Ruie k: IF xk is Ak
for the creatiori of the knowledge base.

10
Bach signature will have a rule so we have as many Accordingly, we have
rules as the number of features. The fuzzy set Ak is
represented by the following exponential membership
function that includes two structural parameters sk
andtk :

(2) (7)

a~ w ay hk
-=--..-.-
- as, ay a ~ ,at,
where x, is the mean 0;is the variance of k th fuzzy
set. Note that the inclusion of these parameters will
help track the variations in the handwriting of
signatures. When sk = 1 andtk = -1, the membership
function is devoid of structural parameters and hence it
is solely governed by the mean and variance+ The
justification for the modified membership function is
two-fold: (i) Easy to track variations aver mean and
variance, and (ii) no need of sophisticated learning l3J aJ dY aw,
-=-._-.-
technique. The numerator and denominator of
exponential €unction in (2) contain a constant term (i.e., at, ay hk at,
1) plus a function of parameter and the known variation
(i.e.,either change over mean or variance). This choice
is guided by the consideration of no role for parameters
if the signatures of a person don't change. But this need
not be the case for other applications. (9)
where d=Y-q, T = ( l + t k ) + t ~ ~ and
~
The strength of the rule in (1) is obtained as k denotes the rule number.
We use the gradient descent learning for the parameters
as follows
The output is expressed as

(4)

where L is the number of rules.

We define the performance function as

J = (Y, - Yy (5) where € 1 ,e2 , ~ are

3 the leaming coefficients such that
where,Y and denote the output of the fuzzy model € 1 , a~ n~d q > O .
and of the real system respectively. If r i s not
available, it can be assumed to be unity. Forged samples of a genuine signature are not readily
available as it is difficult to imitate the various styles of
In order to learn the parameters involved in the signatures by amateurs for producing the unskilled
forgeries and by professional impostors for producing
membership function (i.e., sk andt, ) and the the skilled forgeries. Keeping this point in mind and
consequent parameters ck0and c,, , (5) are partially considering the real-world scenario, we have trained our
system with only genuine signatures, i.e., none of the
differentiated with respect to each of these parameters.
forgeries were used for training the system.

11
 IV. IMPLEMENTATION AND RESULTS dJ
and -in order to update the structural parameters si
The proposed fuzzy modeling based on TS model ati
discussed above has been applied on a signature and ti ; i = 1,..,96. Using these values, we compute the
database, developed in the Graphics Visualization &
Games Development (GVGD) lab at the Multimedia membership functions for all the features. This process
is repeated for all training samples of a person. Here, we
University, Cybejaya, Malaysia.
have devised an innovative approach for the
classification of all signatures (i.e., test signatures and
The signature database consists of a total of 510
random, skilled and unskilled forgeries) of a person. In
handwritten signature images. Out of these, 255 were
order to know the extent of variation in the genuine
authentic signatures and others were forged ones. These
signatures, we determine the maximum and minimum
signatures were obtained Erom 17 volunteers with each
membership functions for each feature over all
person contributing 15 signatures as shown in Table 1.
signatures in the training set. The difference between
The signatures were collected over a period of a few
these two gives the inherent variation in the signatures
weeks to account for variations in the signature with
o f a person.
time. ?he forgeries of these signatures were collected
over a similar time frame. The random forgeries were
We now use the inherent variation to judge the test
obtained by supplying only the names of the individuals
signatures. We will also explain its utility in the testing
to the casual forgers who did not have any access to the
phase. For a particular feature, if the membership value
actual genuine signatures. The unskilled forgeries in
lies within the range of variation, which is given by the
turn, were obtained by providing sample genuine ,

difference of m i n i ” and maximum thresholds, it is

signatures to the forgers who were then allowed to
counted as ‘true’. The total number of ‘true’ cases for a
practice for a while before imitating them to create their
forgeries. Each volunteer had to provide five imitations particular signature is divided by the total number of
features (Le., 96) to get the percentage. For example, in
of any one of the genuine signatures, apart from his or
Fig. 3a, the test signature has 99% of its features lying
her own signatures. These samples constituted the set of
well within the threshold as can be seen from the
unskilled forged signatures for the set of genuine
membership function (i.e., 95 out of 96 features are with
signatures. We then requisitioned the services of a few
expert forgers who provided five forgeries of each
the range of inherent variation). The skill-forged and
unskilled forged signatures have corresponding figures
genuine signature in the test set to create the skilled
of 88.5% (Fig, 3b) and 82.3% (Fig. 3c) respectively. We
forged samples of all the persons.
set the minimum limit or acceptable percentage for
Table 1. Signature database genuine signature at 91% referring to the output result
of signature of signer. Signatures that have percentage
less than 91% are treated as forged signatures. Table 2
Training set Testing set Total summarizes the results of forgery detection using this
Genuine 17 x 10 17 x 5 I70+85= innovative approach.
I 1 I 255
Skilled I 1 17x5 I 85
forgery I
Unskilled I .I 17x5 I 85
1-

forgery I I 5
n
0.8
08-
.... ... Test (99%)
E 04-
02-

1 7 13 19 25 31 37 43 49 55 61 67 73 79 85 91
It is a well-known fact that any automatic signature Foatures
,

verification system requires a very small training set of

signatures. For this reason, we have set the number of t4
training signatures for each individual at ten. M m bershlp Graph of Signatures
1.2 ,
We now explain the process of recognition. If we take
= 1 and considering (15), then (5) becomes

1 7 13 19 25 31 37 43 4B 55 61 67 73 79 85 91
dJ
With the above performance index, we compute - Features

dSi (b)

12
1 1.2 ,
M” bmrihlp Graph of Sianstursm

1
[3] R. Sabourin, R. Plamondon and G. Lorette, Off-
line identification with handwritten signature
images: Survey and Perspectives, Structured Image
Anaiysis, Springer-Verlag, New York, 1992, pp.
2 19-234.
[4] R. Plamondon and S.N. Srihari, “On-line and off-
line Handwriting Recognition: A Comprehensive
Survey”, IEEE Transactions on Pattern Analysis
I 7 13 le 25 31 37 43 49 55 61 67 73 79 85 91
Foatures and Machine Intelligence, Vol. 22, No. I, pp. 63-
x4,2000.
[5] M.A. Ismail and Samia Gad, “Off-line Arabic
Fig. 3. Membership graphs for a particular signature signature recognition and verification”, Pattern
Recognition, Vol. 33, No. 10, pp. 1727-1740, 2000.
Table 2. Results of verification [6] A. El-Yacoubi, E.J.R. Justino, R. Sabourin and F.
Bortolozzi, “Off-line signature verification using
Total Accepted Rejected HMMS and cross-validation”, Proceedings of the
I I I IEEE Workshop on Neural Networks for Signal
Genuine 1 85 I 85 (100%) 1 0 (0%) Processing, USA, 2000, pp. 859-868.
[7] C. Quek and R.W. Zhou, “Antiforgery: a novel
Skilled pseudo-outer product based fuzzy neural network
driven signature Verification system”, Pattern
Recognition Letters, Vol. 23, pp. 1795-18i6, 2002.
Unskilled
[XJ M. Hanmandlu, K.R. Murali Mohan, S.
Chakraborty, S. Goel and D. Roy Choudhury,
Random “Unconstrained handwritten character recognition
based on fuzzy logic”, Pattern Recognition, Vol.
36, NO.3, pp. 603423,2003.
[SI M. Hanmandlu, K.R. Murali Mohan, S. Chakraborty
V. CONCLUSIONS and G . Garg, “Fuzzy modeling based signature
verification system”, Proceedings of the sixth
An off-line signature verification and forgery detection International Conference on Document Analysis
system is modeled by TS model, which involves and Recognition, USA, 2001, pp. 110-1 14.
structural parameters in its exponential membership
’
[lo] M. Ammar, Y. Yoshida and T. Fukumura, “A new
function. The features consisting of angles are extracted effective approach for off-line verification of
using box approach. Each feature yields a fuzzy set signatures by using pressure features”, Proceedings
when its values are gathered from all samples because of the International Conference on Pattern
of the variations in handwritten signatures. In this recognition, 1986, pp, 566-569.
formulation, a single feature constitutes each rule. [ I I ] M. Ammar, “Progress in verification of skillfully
simulated handwritten signatures”, International
The efficacy of this system has been tested on a large Journal of Pattern Recognition and Artificial
database of signatures. The verification system is able to Intelligence, Vol. 5, pp. 337-351, 1991.
detect all types of forgeries: random, unskilled and [12] Jinhong K. Guo, D. Doermann and A. RosenfeId,
skilled with utmost precision. The choice of initial “Off-line skilled forgery detection using stroke and
parameters is important but not crucial. But, we need to sub-stroke properties”, Proceedings o f the
make a proper choice only once and it i s applicable to International Conference on Pattern Recognition,
all types of signatures. We haven’t used global learning 2000, pp. 355-358.
techniques for want of simpticity at the implementation [I31 M. Hanmandlu, K.R. Murali Mohan and Vivek
stage. Gupta, “Fuzzy logic based character recognition”,
Proceedings of the International Conference on
Image Processing, Santa Barbara, USA, pp.714-
VI. REFERENCES 717.
[14]Y. Xuhua, T. Furuhasbi, K. Obata, Y. Uchikawa,
[ 11 R. Plamondon and F. Leclerc, “Automatic signature Study on signature verification using a new
verification: the state of the art 1989-1993”, approach to genetic based machine learning.
lnternationai Journal of Pattern Recognition and Proceedings of the IEEE International Conference
Artificial Intelligence, Vol. 8, No. 3, pp. 643-660, on Systems, Man and Cybernetics, USA, 1995, pp.
1994. 4383-4386.
[2] R. Plamondon and G. Lorette, “Automatic signature
verification and writer Identification: the state of
the art”, Pattern Recognition, Vol. 22, No. 2, pp.
107-131, 1989.

13
 Genuine Skillcd forgery

Fig. 1. Some Examples of typical signatures and their forgeries

Table 3. Knowledge base for a particular signature

Fig. 2. Steps in pre-processing

14

View publication stats