Professional Documents
Culture Documents
Introduction
Handwritten signature can be defined as the name and surname written by own hand.
The specification of human signatures is one of the greatest problems in the designing
the credibly classifiers which work in the identification or verification modes. Both
identification and verification are important in biometrics. Repeatability of signatures
even the same person characterizes a large discrepancy. For example signatory may
put signature every time with different velocities, pen pressures, accelerations, etc.
The distribution of pen pressure can be so different in each signature, that the
determination of a pattern variation can be very difficult [1], [2]. If handwritten
signature will be appropriate measured then it can be treated as a biometric
characteristic. Signature belongs to behavioral biometrics and modernly is widely
acceptable and collectable biometric characteristic. Currently, there are many
measures of determining signatures similarity. Lack of repeatability of the signature
features causes problems with arbitrary indication which features should be analyzed.
The selection of the signature features is a well-known and frequently described
problem. It should be also noticed, that in all previous solutions, only selection of
signature features was used [3], [5], [10], [13], [14]. The algorithm, presented in this
work, not only performs the selection of the signature features but also indicates the
best similarity measures (from the set of available measures), which minimize
signature verification error. For every person different signature features and different
similarity measures can be chosen. Mentioned algorithm is based on the statistical
N. Chaki and A. Cortesi (Eds.): CISIM 2011, CCIS 245, pp. 128136, 2011.
Springer-Verlag Berlin Heidelberg 2011
129
Signature Acquisition
In the first step, the two sets of signatures are created for every person. The set
GS={S1,S2,...Sno} contains only original signatures, while the set GS={S1,S2,...Snf}
contains a forged signatures of the same person. As aforementioned, the signature
verification can be carried out on the basis of different signature features and different
similarity measures. The set of all available signature features that can be used in the
signature analyze process will be denoted as F={f1,f2,...fn} and the set of all available
methods of signature recognition will be denoted as: M={m1,m2,...mk}. It is assumed
that each signature feature can be analyzed by means of arbitrary, known method. In
this case, the set FM containing all possible combinations of pairs feature-method
used to verify a signature is defined as follows:
FM={(f,m)i: f F, m M}, i=1,,(n k)
where:
(f,m)i
n
k
(1)
In the next stage, the training sets are created. The training sets are necessary for a
proper operation of the classifier. Based on the learning sets, classifier performs
verification of a new, unknown signature. In the proposed method, the learning sets
contain the two matrices U(1) and U(2). Thanks to them, the classifier is able to
distinguish the original signatures of a given person from other forged signatures. The
correct creation of the training set (the set of examples) is very important and
significantly affects the classifier effectiveness. The matrix U(1) is built on the basis of
a set GS of original signatures of the person. The matrix contains the values of the
similarity coefficient Sim calculated between pairs of original signatures. As the
similarity coefficient Sim any, normalized to the interval [0,1], symmetric similarity
measure can be taken. The matrix U(1) is built as follows:
130
(2)
where:
Si, Sj
the i-th and j-th original signature,
no
the number of original signatures.
The matrix U(1) consists of the columnar vectors. The each columnar vector contains
the values of similarity coefficients between two signatures. The similarity
coefficients are calculated successively using all combinations of pairs feature
method. The first columnar vector of the matrix U(1) is shown below:
Sim( S1 S2 )( f1 ,m1 )
Sim( S1 S2 )( f1 ,mk )
[ S1 S2 ] =
Sim( S S )( f n ,m1 )
1
2
Sim( S S )( f n ,mk )
1
2
#( FM )1
(3)
where:
( f ,m )
Sim( S a Sb ) i j similarity coefficient of the i-th feature fi in the signature Sa
and Sb. Similarity was determined by means of the j-th method mj.
In order to create a matrix U(2), the set GS of original signatures and the set GS of
forgery signatures are required. The matrix U(2) is built similarly as before as follows:
U (2) = S1 S '1 ,..., S1 S 'nf ,..., Sno S 'nf
(4)
where:
nf number of the all unauthorized (forged) signatures,.
The individual columns of the matrix U(2) can be constructed similarly as in the
matrix U(1):
Sim( S1 S '1 )( f1 , m1 )
Sim( S1 S '1 )( f1 , mk )
[ S1 S '1 ] =
Sim( S S ' )( fm , m1 )
1
1
' ( f m , mk )
Sim( S1 S 1 )
#( FM )1
(5)
where:
Si
Sj
Sim( Sa S 'b )
( fi , m j )
131
132
K=
1 2
Ki ,
N 2 i =1
K i = U*(i ) U*(i )T ,
N=n1+n2
(6)
where:
(1) and U
(2) , respectively
the number of columns in the reduced matrices U
n1, n2
The algorithm of reduction of the matrices dimension will be executed in the
successive several steps:
1) The discriminant measure is calculated by means of the Hotellings statistic [4]:
T 2 ( y1 , , y p ) =
1 2
T
ni (ui u ) K 1 (ui u )
N 2 i =1
(7)
where:
yi
the i-th pair featuremethod,
p
the current dimension of the data vector.
2) The discriminant measure, taking into consideration the i-th pair featuremethod
absence, is then computed [4]:
i {1, , # FM}
Ti 2 ( y1 , , y p ) = T 2 ( y1 , , yi 1 , yi +1 , , y p )
(8)
(9)
Ui
1+ T
( y , , y ) U
1
(10)
i
The value F from (10) is then compared with the table of critical values of the Fdistribution. In our approach, in the each loop of the algorithm for the value F
significance level = 0.05 was established. The F-distribution critical values were
taken from the well-known statistical tables. Because we have two classes, the degree
of the freedom is f = N-p-1, hence the F1, N p 1 distribution is checked.
If the value of F does not fall into the critical region, the current i-th row of the
(1) and U
(2) ) can be removed. In the successive step the parameter p is dematrices U
creased, and algorithm starts with the new value of parameter p from the beginning.
Hence matrices dimension can be reduced.
(1) and U
(2) might have a significant
Finally, remaining elements of the matrices U
influence on the classifiers work. The mentioned procedure is executed for the every
(Q)
person Q. Hence, the results of the selection are stored in the set FM . It means that
133
for the person Q we obtain the set FM ( Q ) FM in which the pairs ( f m)Q are
included. These pairs the best distinguish the genuine signature of the person Q from
his/her forged signatures. It should be also strongly noted that the previously described process of the data reduction have to be repeated again when the new signature will be collected. It is necessary because the main set FM will be changed. It
means that recognition process is always conducted in the closed set of data. Hence,
the analyzed signature always belongs to one of the classes class determined by the
(1) or the class determined by the matrix U
(2) .
matrix U
Signature Verification
classified signature S* the vector h=[h1,,hn] is created. Elements of the vector h are
determined as follows:
(Q )
hi = Sim( S *, S j )( f , m )i ,
where:
S*
Si
(Q )
hi = Sim( S *, S j )( f , m )i
j = 1,..., n
(11)
signature to be verified,
the i-th original signature,
similarity coefficient of the signatures S* and j-th original signatures Sj . Similarity was determined with use of
the i-th pair ( f , m)i(Q ) from the set FM(Q).
In the next stage, the distances d (h, h i ) between the vector h and the vectors h i are
successive calculated. If h i 1 , then d h, h i 1 . If h i 2 , then d h, h i 2 .
Among all distances the smallest k distances are selected. So k-NN classifier is applied [7], [15]. Selected distances may belong to the two classes: 1 (original signature) and 2 (forged signature). The verified signature S* is classified to the class
134
Experimental Results
Aim of the study was to evaluate the proposed method of signature verification based
on automatic selection of signature features and methods of their analysis. During the
study the three variants of the classifiers work have been compared:
VAR1 where only signature features selection was carried out, for one previously assumed method,
VAR2 where only signature features selection was used and the most relevant
methods of their analysis were chosen,
VAR3 in this mode, signature features were not selected.
The study was conducted on a set of real data. Signatures came from the SVC2004
database [16]. It contains both genuine signatures and skilled forgery signatures. All
calculations were realized in the Matlab environment. The set of all combinations of
the pairs featuremethod contained the #FM=18 pairs. Signature features and
methods their analysis, were selected from many well-known from literature
similarity coefficients and metrics. The researches were conducted with the use of a
database containing 800 signatures coming from 40 people. For each person 10
original signatures and 10 professionally forged were used. The proposed method
FMS was compared with other, known from the literature PCA [6], [11] and SVD
[12] methods, where reduction of the features has been also performed. Based on [6],
[11], [12] and proposed technique the different experiments have been carried out.
In the method FMS it is not possible to arbitrarily determine the dimension of the
reduced vector of features. This dimension was determined automatically. The conducted studies shown that in practice, the dimension of the reduced vector is between
3 to 12. In a case of using the PCA and SVD methods the boundary features reduction
should be defined a priori. Achieved level of the FAR and FRR errors, for different
methods of features reduction, presents Tables 1-2. In the Tables 1-2 only the smallest
FAR/FRR values have been shown. Those results were obtained for the different
number of signatures from the sets GS (original signatures) and GS ' (forged signatures) for the different number of reduced features. The FRR ratio describes the genuine signatures classified as forged, while the FAR represents the forged signatures
recognized as the genuine.
Table 1. False Accepted Rate for different features selection methods, different number of
reference signatures, and different dimension of vectors
FAR [%]
GS
GS
FMS
3
5
10
1
3
4
6.34
1.08
1.67
135
Table 2. False Rejection Rate for different features selection methods, different number of
reference signatures, and different dimension of vectors
FRR [%]
GS
GS
FMS
3
5
10
1
3
4
7.14
2.53
2.60
The Tables 1-2 show that the smallest FAR/FRR coefficients were obtained using
proposed in this work method. For this method FAR=1.08%, FRR=2.53%. These
results are significantly better than results obtained using the SVD or PCA methods.
Differences between obtained results in the PCA/SVD methods follow from the
precision of calculations. The smallest FAR/FRR ratio were obtained when the
number of original signatures in the set GS was 5 and the number of forged signatures
in the set GS ' was 3.
Conclusions
The originality of the proposed approach follows from the fact that classifier utilizes
not only extracted signatures features, but also the best (for the analyzed signature)
similarity measures. In the signature biometrics such statistical approach has not been
applied yet.
References
1. Al-Shoshan, A.I.: Handwritten Signature Verification Using Image Invariants and Dynamic Features. In: Int. Conf. on Computer Graphics, Imaging and Visualisation, pp. 173176
(2006)
2. Doroz, R., Porwik, P., Para, T., Wrobel, K.: Dynamic Signature Recognition Based on Velocity Changes of Some Features. Int. Journal of Biometrics 1, 4762 (2008)
3. Doroz, R., Wrbel, K., Porwik, P.: Signatures Recognition Method by Using the Normalized Levenshtein Distance. Journal of Medical Informatics & Technologies 13, 7377
(2009)
4. Anderson, T.W.: An introduction to multivariate statistical analysis. Wiley (1984)
5. Ibrahim, M.T., Kyan, M.J., Guan, L.: On-line Signature Verification Using Most Discriminating Features and Fisher Linear Discriminant Analysis. In: 10th IEEE Int. Symposium
on Multimedia, Berkeley CA, pp. 172177 (2008)
6. Ismail, I.A., Ramadan, M.A., Danf, T.E., Samak, A.H.: Automatic Signature Recognition
and Verification Using Principal Components Analysis. In: Int. Conf. on Computer
Graphics, Imaging and Visualization, pp. 356361 (2008)
136
7. Jzwik, A., Serpico, S.B., Roli, F.: A Parallel Network of Modified 1-NN and k-NN Classifiers-Application to Remote-Sensing Image Classification. Pattern Recognition Letters 19, 5762 (1998)
8. Kirkwood, B.R., Sterne, J.A.C.: Essentials of Medical Statistics, 2nd ed. Wiley-Blackwell
(2003)
9. Kovari, B., Charaf, F.: Statistical Analysis of Signature Features with Respect to Applicability in Off-line Signature Verification. In: 14th WSEAS Int. Conf. on Computers, vol. II,
pp. 473478 (2010)
10. Lei, H., Govindaraju, V.A.: Comparative Study on the Consistency of Features in On-line
Signature Verification. Pattern Recognition Letters 26, 24832489 (2005)
11. LI, B., Wang, K., Zhang, D.: On-line Signature Verification Based on PCA (Principal
Component Analysis) and MCA (Minor Component Analysis). In: Zhang, D., Jain, A.K.
(eds.) ICBA 2004. LNCS, vol. 3072, pp. 540546. Springer, Heidelberg (2004)
12. Moravec, P., Snasel, V.: Dimension Reduction Methods for Iris Recognition. In: Proc. of
the Int. Workshop on Databases DATESO 2009, pp. 8089 (2009)
13. Porwik, P., Wrobel, K., Doroz, R.: Signature Recognition Method by Means of the Windows Technique. Int. Journal Image Processing and Communication 14, 4350 (2009)
14. Richiardi, J., Ketabdar, H., Drygajlo, A.: Local and Global Feature Selection for on-line
Signature verification. In: Proc. of the 8th Int. Conf. Document Analysis and Recognition
ICDAR 2005, pp. 625629 (2005)
15. Shakhnarovich, G., Darrell, T., Indyk, P.: Nearest-Neighbor Methods in Learning and 15.
In: Vision: Theory and Practice (Neural Information Processing). The MIT Press (2006)
16. Willems, G., Pison, G., Rousseeuw, P. J., Van Aelst, S.: A Robust Hotelling Test, Vol. 55,
pp. 125138. Physica Verlag, An Imprint of Springer-Verlag GmbH (2002)
http://www.cse.ust.hk/svc2004/