International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
383
REVIEW
ON OFFLINE SIGNATURE VERIFICATION METHODS
BASED ON ARTIFICIAL INTELLIGENCE TECHNIQUE
Imran Hussain, Vikash Shrivastava, Vivek Kr. Shrivastava
1
IT Department, ITM Bhilwara, Rajasthan, India; 2 I-Gate Global Solution Limited, Noida, U.P, India; 3 IT Department, ITM Bhilwara,
Rajasthan, India
Email: imran.hussain1983@gmail.com
ABSTRACT
Signature can be used as a biometric is implemented in various systems as well as every signature signed
by each person is distinct at the same time. It is very important to have a computerized signature
verification system. In case of offline signature verification system dynamic features are not included
obviously, but we can use a signature as an image and apply image processing techniques to make an
effective offline signature verification system. Author wants to illustrate two techniques reviewed by him
on Offline signature Verification. Those techniques are mixed of Energy with Angle and Energy with
Chain Code.
Keywords : Neural Network, Angle, Energy Density, Chain code, FAR and FRR.
I. INTRODUCTION
or person identification, the usage of biometrics is very
general and important in daily routine. Signature can be used
as a biometrics as every signature is distinct in nature. As
signature has already taken and accepted as an identification
of the person who signed in so many systems, it is very much
important to keenly observe the signature before having any
conclusion about the signee. This gives opportunity to develop
a computerized signature verification system. But, in many
cases only image of a signature is available so offline
signature verification seems more important than online
signature verification system. The main thing of consideration
is that a signature of a person may vary according to the
mood, health etc. even the genuine signer may not sign in the
same manner as his/her signature as it is. A change is
observed every time. Then this seems somewhat difficult to
distinguish between genuine signature and a forgery.
signatures, because of the age, geographic location, illness,
and perhaps the emotional state of the person, accentuates the
problem. All these joined together, cause large intra-personal
variation. A robust system has to be developed which should
not only be able to consider these factors but also able to
detect various types of forgeries. The systems should neither
be too coarse nor too sensitive. It should have an acceptable
trade-off between a low False Rejection Rate (FRR) and a low
False Acceptance Rate (FAR). The proposed systems should
also find an optimal storage and comparison solution for the
extracted feature points.
IJOART
Basically, approaches to signature verification divided into
two categories according to the acquisition of the data: Online and Off-line. On-line signature records the motion of the
stylus while the signature is produced, and includes position,
and possibly velocity, acceleration and pressure of pen, as
functions of time. Online signature verification systems use
this information captured during acquisition. These dynamic
characteristics are specific to each and every individual and
sufficiently stable . Off-line signature is a 2-D image of the
signature. Processing Off-line is complex due to the absence
of dynamic characteristics. Difficulty also lies in the fact that
it is hard to break signature strokes due to highly stylish and
unconventional writing styles. The quality and the variety of
the writing pen may also affect the nature of the signature
obtained. The non-repetitive nature of variation of the
Copyright © 2013 SciResPub.
The problem is approached in two steps. Initially the scanned
signature image is preprocessed to be suitable for extracting
features. Then the preprocessed image is used to extract
relevant geometric parameters that can distinguish signatures
of different persons. The next step involves the use of these
extracted features to verify a given image.
A. Motivation
The motivation behind the review is the growing need for a
full proof signature verification scheme which can
guarantee maximum possible security from fake signatures.
The idea behind the review is also to ensure that the
proposed scheme can provide comparable and if possible
better performance than already established offline
signature verification schemes. There may be a case where
the type of verification system used for training differs from
classification using network. Though the test sample is of a
genuine person, it might not be possible to prove with either
of these systems alone.
B. Research Objectives
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
Signature verification is an important research area in the
field of personal authentication. The recognition of human
handwriting is important concerning about the improvement
of the interface between human-beings and computers. If
the computer is intelligent enough to understand human
handwriting it will provide a more attractive and
economic man-computer interface. In this area signature is
a special case that provides secure means for authentication,
attestation authorization in many high security environment.
The objective of the signature verification system is to
discriminate between two classes: the original and the
forgery, which are related to intra and interpersonal variability.
The variation among the signatures of same person
is called Intra Personal Variation. The variation between
originals and forgeries is called Inter Personal Variation. Our
review is concerned with the techniques of off-line signature
verification. The static information derived in an off-line
signature verification system may be global, structural,
geometric or statistical. We concern with offline signature
verification which is based on geometric centre and is useful
in separating skilled forgeries from the originals. The
algorithms used have given improved results as compared to
the previously proposed algorithms
C. Applications of Off-Line Signature Verification
The handwritten signature has many purposes
and meanings. It can be used to witness intentions (e.g.
signing of a contract), to indicate physical presence (e.g.
signing in for work), as a seal of approval or authorization
and as a stamp of authenticity. Thus, numerous applications
for the off-line signature verification are available.
384
In this method, two features are used for training. Aspect
ratio is used as a global feature and energy density is used as
local feature. Aspect ratio is the ratio of Height (maximum
vertical distance) to length (maximum horizontal distance) of
the signature. It is calculated after skew removal. Energy
density is defined as the total energy present in each
segment. 100 segments of each signature are done and
energy density is obtained by counting the total number of 1s
in each segment (i.e. Total White Pixels). Thus, the “feature
vector” of size 101X1, for energy density method, is final
database. This final database is fed to the neural network to
perform the desired function i.e. training or classification.
b.
Chain Code
Fig. 1 Connectivity of a Pixel
IJOART
Fig. 2. Direction Changes in a Part of a Signature
Chain code is based on the direction of the connecting
pixels. Each pixel is observed for next connected pixel and
their direction changes mapped by giving a numerical value
A signature is treated as an image transport a certain to every possible direction. There are generally 8
pattern of pixel that relate to a specific individual signature connectivity is taken into consideration as shown in the Fig.
verification problem , therefore is concerned with 1. But in this technique 4 connectivity i.e. 0, 1, 2 & 3 is used.
examining and formative whether particular signature truly As another 4 directions i.e. 4, 5, 6 & 7 are simply the
belongs to a person or not. signature verification is a different negation of 0, 1, 2 & 3 directions. To obtain chain-code top
pattern recognition problem as two genuine signature of a left corner is considered as origin and scanning is done left to
person are precisely the same the difficulty also sterns from right, top to bottom (refer Fig. 2). Each pixel has observed
the fact that skilled forgeries follow a genuine pattern separately and direction vector for that pixel is noted down.
unlike fingerprint which vary widely for two different This process is carried out until the last pixel has scanned.
person. signature verification can be divided into two classes, Now, the frequency of occurrence in a particular direction is
namely, off- and on-line verification.
calculated for each segment and the histogram for each
segment is used it to train the neural network.
II. BACKGROUND AND RELATED WORK
III. PROPOSED APPROACHES
Various approaches were done by various people to achieve
highest accuracy as well as FAR & FRR. Till now the
accuracy, FAR and FRR features were compared on the basis
of Energy Density, Angle, Chain Code and mix of Energy
Density with Angle and mix of Energy Density with Chain
Code features. The details of these techniques are depicted as
below.
a. Energy Density
Copyright © 2013 SciResPub.
c.
Angle
In this method first the Pre-processing image is resized
and partitioned into four portion or cell using the equal
horizontal method after that each partition(cell) are divided
in to 3 row and 3 column of equal size so we have total
nine sub cell of each cell. After that consider the sub cell
one by one and calculate the angle of each with pixels by
considering the bottom left corner after that calculate the
mean value of the angles this process is repeat for all the
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
sub cells. Once the value of angle, for each sub cell, is
found, then calculating the mean value from that, to determine
the value of angle for that cell or partition. This process is
repeat for the reaming three partitions, so at the end we
have the angle vector of size 1*4. This is given as an
input to the neural network. For example the data base used
consist 100 signature samples. For one sample we have
angle vector of size 1*4 so for all 100 sample we have
feature vector of size 100 *4 which is used as a final data
base for training the neural network and also for
classification.
385
original digitized signature image under consideration. Then
from (b) to (e) result of binarization, denoising, thining and
skew removal is shown respectively. (f) describes the result
of segmentation process for which 100 segments is made for
each signature and further processing is done with each
segments found after this step. After complete preprocessing
both the energy density and directional feature is computed
and fed to an artificially intelligent network for training and
classification.
IV. IMPLEMENTAION AND VERIFICATION
(a)
Implementation of the proposed approach is basically
divided in two parts i.e. Training and Classification. Training
belongs to preparing and training of neural network for doing
the classification work with an optimum accuracy. The
proposed signature verification system takes an image of the
signature as input and verifies whether the input image
matches with the genuine training signature image available
in the database or not. The system can be broadly categorized
on the basis of method used for pre-processing and feature
extraction from the image database and final input given to
the neural network.
Raw Signature Database is gathered from 10 people and
110 Genuine & 110 Forgeries from each individual is
collected (i.e. 2200 Signature Samples) and digitized using
scanner, 300 dpi resolution. The first step for pre-processing
is Binarization. It is used to produce binary image i.e. to
convert colored (if any) image in black & white (i.e. in 0 or
1) format. In this paper global threshold is used for this
purpose. Noise is filtered out using median filter. Thinning is
done after noise filtering. Morphological operation is applied
(in MATLAB) to performs the desired thinning. Rotation of
signature patterns by a non predictive angle is one of the
major difficulties. In this paper simply the concept of
trigonometry is used for skew removal. Fig. 4 shows the
effect of different stages of pre-processing. The next step of
pre-processing is to extract the signature only from the whole
image, by removing the image of extra paper remained. After
signature extraction again resizing is carried out, and then
segmentation process is completed to extract the local
features of the signature (i.e. energy density or directional
feature of each segment according to the method under test).
Author has used 100 segment of each signature sample for
further processing. In energy density method author has also
used aspect ratio as a global feature to improve the
performance of the system. Aspect ratio is extracted just
before the resizing and segmentation.
(b) Binarization
Original Signature Image
(c) Denoising
(d) Thinning
(e) Skew Removal
IJOART
In proposed directional feature with energy method for
offline signature verification system author simply merge the
above two methods altogether to observe the effect of the
merging. Energy, aspect ration and directional feature of a
thinned and binarized signature image is calculated and fed
as an input to a feed forward back propagation neural
network for training and classification again this
preprocessing is done in the same manner. Fig. 3 shows the
effect of some of the preprocessing steps. (a) shows the
Copyright © 2013 SciResPub.
(f) Segmentation : 100 segments done for each signature Fig. 3. (a) to (f)
Output of Different Stages of Pre-processing
Fig. 4. Proposed Architecture of Neural Network for mix of energy
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
and Chain code method .
Nm,n shows Neuron where, m = Layer Number & n = Neuron
Number Wi,j,k represents Weight where, i=Layer No..,
j=Neuron No., k=Output No.of the particular Neuron
The proposed ANN scheme for mix of Chain code with
energy feature uses a multi layer feed forward network
employing a back propagation learning algorithm with 16
Neurons in input layer and 1 Neuron in output layer. One
hidden layer is present with 16 Neurons. The transfer
function used for all the three layers are Hyperbolic Tangent
Sigmoid (tansig). The proposed architecture of Neural
Network is shown in Fig. 4. Here, default value of bias is
chosen. Total 501 inputs i.e. 100 for energy of each segment,
400 as direction feature and 1 show aspect ratio is given to
this neural network.
In case of angle mix with energy feature the signature of
10 person is gathered which consist of 50 genuine and
50 forged signature of an individual person. (i.e. 1000
Signature Samples). Rest all process is same as the previous
technique.
386
increasing training samples for all of the methods. For
performance evaluation author used some common
parameters like time required for training, accuracy, FAR
(False Acceptance Ratio) i.e. the percentage of forgeries
accepted as genuine & FRR (False Rejection Ratio) i.e. the
percentage of rejecting the genuine signatures. A comparison
for all the three methods has done on the basis of above
mentioned parameters. Table I, II, III & IV respectively
elaborate the comparison of all the methods.
TABLE I:
Comparison of Energy Density and
Directional Features methods on the basis of Time
Required for Training
No. of
Training
Samples Elapsed Time
(50% (in sec.)
Genuine ( Energy
Sr. & 50% Density
No. Forgery) Method)
1 10
6.156
2 20
6.453
3 30
6.593
4 40
7.312
5 50
6.625
6 60
7.671
7 70
7.797
8 80
6.313
9 90
6.704
10 100
6.969
Result
Elapsed Time
(in sec.)
( Directional
Feature
Method)
7.375
7.89
8.39
9.282
10.718
8.172
8.281
8.594
9.188
9.125
IJOART
Fig.5 Proposed Architecture of Neural Network for mix of
energy and angle method
The proposed ANN scheme (Fig.5) for mix angle with
energy feature uses a multi layer feed forward network
employing a back propagation learning algorithm with 4
Neurons in input layer and 1 Neuron in output layer. One
hidden layer is present with 40 Neurons. The transfer
function used for all the three layers are Hyperbolic Tangent
Sigmoid (tansig). Total 4 inputs is given to this neural
network.
V. RESULTS
(i)In case of Energy, Chain Code (as a directional feature)
and mix of Energy with Chain Code features:
Till now for comparison and performance evaluation of
the proposed methodology the author of the proposed
approches have used equal number of genuine and forgery
samples for training and 100 numbers (50 Genuine + 50
Forgeries) are used for classification. Author increased the
training signature samples from 10 (5 Genuine + 5 Forgeries)
to 100 (50 Genuine + 50 Forgeries) and observe the effect of
Copyright © 2013 SciResPub.
Elapsed Time
(in sec.) ( Mix
of Both)
7.672
8.718
9.141
9.594
11.64
8.5
8.5
11.688
9.531
9.532
Table: II Comparison on the basis FAR
Sr.
No
1
2
3
4
5
6
7
8
9
10
No. of Training
Samples (50%
Genuine + 50 %
Forgery)
10
20
30
40
50
60
70
80
90
100
Result (FAR in %)
Energy
Chain Mix of
Density
Code
Energy
Density
with Chain
Code
42
40
28
42
40
12
42
30
10
48
40
6
46
20
8
22
6
4
2
0
6
6
0
0
10
0
0
6
4
0
TABLE III: Comparison on the basis of Accuracy
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
2 5 o, 2 5 f
0 . 12
No. of Training
3 0 o, 2 0 f
0 . 03
Result (FRR in %)
Samples (50%
4 0 o, 1 0 f
0 .0 2 5
Genuine + 50 %
Forgery)
Energy
Chain
Code
Density
Sr.
No
Mix of
Energy
Density
with
Chain
Code
1
10
44
20
30
2
20
4
12
14
3
30
22
14
22
4
40
12
20
26
5
50
20
30
26
6
60
18
28
24
7
70
22
22
8
8
80
2
6
6
9
90
2
8
4
10
100
2
2
0
5 0 o, 5 0 f
387
0 . 08
0 . 16
0 . 20
0 . 30
0 . 04
0 . 03
0 .0 2 5
0 . 08
0 . 05
0 . 10
0 . 12
0 . 08
0 . 04
Table 3: (For Sample-C).
Methods of feature extraction
Energy
Angle
Density
Feature
No. of
samples
(Original
Forgery)
FRR
1 0 o,
2 0 o,
2 5 o,
3 0 o,
4 0 o,
5 0 o,
4 0f
3 0f
2 5f
2 0f
10f
50f
FAR
0 . 20
0 . 15
0 . 08
0 . 06
0 .05
0 . 18
FRR
0
0 . 06
0 . 20
0 . 25
0 . 20
0 . 24
FAR
0 . 15
0 . 10
0 . 06
0 . 06
0
0 . 16
0
0 .03
0 .12
0 .20
0. 1 0
0.20
Comparison on the basis of Time:
Num ber of Samples
Time required for
Feature Extraction
and Training
(Energy Density)
Feature required for
Feature Extraction
and Training
(Angle Feature)
50
21 .9 4 Sec.
49 .8 2 Sec.
TABLE IV: Comparison on the basis of FRR
(ii) In case of Energy, Angle (as a directional feature) and
mix of Energy with Angle features:
IJOART
100
36 .5 6 Sec.
92 .3 3 Sec.
Table 1: (For Sample-A)
Methods of feature extraction
Energy
Angle
Density
Feature
No. of
samples
(Original
Forgery)
1 0 o,
2 0 o,
2 5 o,
3 0 o,
4 0 o,
4 0f
3 0f
2 5f
2 0f
1 0f
5 0 o, 5 0 f
VI. CONCLUSION
Chain code method is reported very effective for recognition.
The mixed method of energy with chain code is giving
superior performance. Mixed of energy and angle is also
performed well. Due to variation in the signature there is
small change in performance. It is required to evaluate the
FRR
FAR
FRR
FAR
0 . 20
0 . 15
0 . 08
0 . 03
0 .0 2 5
0 . 10
0 . 06
0 . 16
0 . 20
0 . 20
0 . 15
0 . 10
0 . 08
0 . 06
0 .0 2 5
0 . 02
0 . 03
0 . 04
0 . 10
0 . 10
0 . 12
0 . 16
0 . 06
0 . 14
Table 2: (For Sample-B).
Methods of feature extraction
Energy
Angle
Density
Feature
No. of
samples
(Original
Forgery)
1 0 o, 4 0 f
2 0 o, 3 0 f
FRR
FAR
FRR
FAR
0 . 06
0 . 20
0 . 05
0 . 16
0 . 10
0 . 05
0
0 . 03
Copyright © 2013 SciResPub.
Sr.
No
No. of
Training
Samples (50%
Genuine +
50 % Forgery)
Result (Accuracy in %)
Energy
Chain
Mix of
Density
Code
Energy
Density with
Chain Code
1.
10
57
70
71
2.
20
77
74
87
3.
30
68
78
84
4.
40
70
70
84
5.
50
67
75
83
6.
60
80
83
86
7.
70
88
89
93
8.
80
96
97
97
9.
90
94
96
98
10.
100
96
97
100
mixed method performance on same signatures databases. The
IJOART
International Journal of Advancements in Research & Technology, Volume 2, Issue 5, M ay-2013
ISSN 2278-7763
388
feature and classifier performance will also be needed to
evaluate at small training dataset.
References
1. Bai-ling Zhang, “Off-Line Signature Recognition and Verification by
Kernel Principal Component Self-Regression,” Fifth International
Conference on Machine Learning and Applications (ICMLA'06)
10.1109/ICMLA.2006.37, p. 28 to 33.
2. Deepthi Uppalapati,“Integration of Offline and Online Signature
Verification systems,” Department of Computer Science and Engineering,
I.I.T., Kanpur, July 2007.
3. Rejean Plamondon and Sargur N. Srihari, “On-Line and Off-Line
Handwriting Recognition: A Comprehensive Survey,” IEEE transections
on pattern analysis and machine intelligence, Vol. 12 No. 1, January
2000, pp. 63 - 84.
4. The IEEE website. [Online]. Available: http://www.ieee.org
5. Department of Computer Science, IIT, Kanpur for precious imformations.
Available: http://www.cse.iitk.ac.in
6. H. N. Prakash and D. S. Guru, “Off-line signature verification: an
approach based on score level fusion,” International journal of computer
application, Vol. 1-No.18, 2010.
7. Samaneh Ghandali and Mohsen Ebrahimi Moghaddan, “Off-line Persian
Signature identification and verification based on image registration and
fusion,” Journal of Multimedia, Vol.-4 No.3, june 2009.
8. Imran Siddiqi and Nicole vincent, “A set of chain-code based features for
writer recognition,”10th International Conference on document analysis
and recognition, 2009.
9. Stephane
Armand,
Michael
Blumenstein
and
Vallipuram
Muthukkumarasamy, “Off-line signature verification using enhanced
modified direction feature and neural based classification,”
10. Andrew T. Wilson, “Off-line handwriting recognition using neural
network,”
11. Miguel A. Ferrer, Gesus B. Alonso and Carlos M. Travieso, “Off-line
Geometric parameters for automatic signature verification using fixed
point arithematic,” IEEE transections on pattern analysis and machine
intelligence, Vol. 27 No.6, June 2005, pp.993-997.
IJOART
12. Reena Bajaj and Shantanu Choudhari, “Signature Verification
using multiple neural classifiers,” Pattern recognition, Vol. 30 No.
1, 1997, pp. 1-7.
13. Emre Ozgunduz, Tulin Sentruk & M. Elif karsligil, “Off-line
signature verification and recognition by support vector machine”.
14. Thomas M. Breuel, “Representations and metrics for off-line
handwriting segmentation.
15. M. Blumenstein and B. Verma, “An artificial neural network based
segmentation algorithm for off-line handwriting recognition”.
16. Jamal
Fathi and Abuhasna, “Signature recognition using
conjugate gradient neural network”.
17. Minal Tomar & Pratibha Singh, “A Simpler Energy Density
method for Off-line. Signature Verification using Neural
Network”.
18. Minal Tomar & Pratibha Singh, “An Intelligent network for offline
signature verification system using chain code” published in
proceeding of The First International Conference on Computer
Science and Information Technology.
19. Rahul Sharma & Manish Shrivastava, “An Offline Signature
Verification System using Neural Network Based on Angle
Feature and Energy Density”
Copyright © 2013 SciResPub.
IJOART