Você está na página 1de 5

Offline Signature verification using Trace Transform

Abstract In this paper the application of Trace Transform


for signature verification has been investigated. Trace Transform
is used for the first time for signature verification and to binary
images. Trace Transform is 2D feature and is not unique and its
functional dependent. Features of lower dimensions are derived
from Trace Transform by applying functional. When functional
of certain properties are applied, affine invariant features can be
obtained from the trace transform function. At present here,
certain pairs of functional have been used to derive trace
transform and from it, its circus function which is function of
orientation of lines has been used as a feature after normalizing
it. Similarity scores between the features is computed and is used
for verifying the authenticity of signatures.
KeywordsBehavioral biometrics; signature verification; trace
Transform;circus function;

I.

INTRODUCTION

In modern day world, there is a growing demand and


necessity for personal identification and verification systems.
There is an active research going on in this field to build
robust sophisticated verification systems. Biometric systems in
this field play a major role as they have genuine advantages
over other systems like token based systems and password
based systems.
Biometrics is the traits or features acquired by an individual
by birth or adopted during the course of life. A person has
physiological features such as fingerprint, iris, palm print,
hand geometry etc., are called physiological biometrics. Traits
such as signature, gait, typing rhythm of a person forms the
behavioral biometrics or behaviometrics.
Signature is one of the behavioral biometric which is
socially as well as legally accepted form of personal
verification and identification. Signature verification finds
applications in credit card validation, bank cheque
authentication, certificates, bonds, contracts verification etc. In
signature identification, the system has to identify the author
of the test signature where as in signature verification the
system has to verify whether the test signature is a genuine or
a forgery. Forgeries can be divided as random, simple and
skilled.
Automatic signature verification can be of two types, online method and offline method. Online methods measure the
sequential data such as coordinates of writing points, pen
pressure, angle and direction of the pen. This data in the form
of electronic signals is obtained from an acquisition device
such as tablet, throughout the signature apposition process.
Offline methods use a camera or scanner that performs
acquisition of signature images from papers or documents
after the completion of signature apposition process. Since
online data has dynamic information, online systems are more
reliable and efficient than offline systems. But there are more
real time applications for offline systems and is also cost
effective.

Generally in offline methods, features are extracted from


image after preprocessing and these features are subsequently
used for classification, identification or verification. The types
of features used in literature can be classified as local, global,
Structural, Geometrical or Statistical Features. Global features
like Gabor transform, wavelet transform and projection
profile, extract the characteristic features from whole image by
treating as a single unit. Local features extract more detailed
information from the small regions of signature.
There is an active research [8] going on in this field over
the past two decades because of its potential applications.
Sigari et al., 2011 [7] used 2D Gabor transform that detects
the edges that match with the Gabor wavelet. Since global
features do not provide much useful information and to reduce
the computational complexity, a grid is superimposed on the
signature image and coefficients are computed at grid points.
Coefficients are computed by varying the rotation angle and
wavelength of the Gabor wavelet.
Coetzer et al., 2004 [5] used discrete Radon transform, a
matrix whose columns are projections of signature image at
various angles, as a feature. All the projections are interpolated
or decimated to have same length. Even though there is a
repetition of features from 1810 to 3600 angles, total
orientations are considered to achieve rotation invariance and
to make the feature feasible to use with ring topological
Hidden Markov Model (HMM) for classification. This feature
may be considered as a Global feature.
Pourreza et al., 2009 [4] detected line segments and their
widths in signature images using Radon Transform. The peak
of the projection value exceeding the threshold, after dividing
by the window size validates the existence of the line segment.
A histogram of line segments of various widths detected over
orientations 00 to 1790 is considered as feature vector. The
performance of the system is varied by changing width
combination and orientation resolutions. SVM classifier with
RBF kernel has been used for classification.
Signature verification field is still challenging because both
inter class and intra class variances are present. Intra class
variance is due to the fact that any person cannot writer his
signature same all the time and inter class variance is due to
the variation in forgery to genuine signatures. More inter class
variance and less intra class variance enables to build efficient
systems. The efficiency of the signature verification system
can be decided by inspection two types of errors, False
Acceptance Rate (FAR) and False Rejection Rate (FRR).
The rest of the paper is organized as follows.

II.

TRACE TRANSFORM

Trace Transform is generalization of Radon Transform. The


Radon Transform is the 2D Representation of the image in
coordinates of and , where each cell ( , ) is the
integral of the image along a line with parameters and
as shown in figure (1). Radon Transform (R) is defined as in
equation (1).
R( , )

f ( x, y ) ( x cos y sin ) dxdy

(1)
Trace Transform is similar to Radon in the sense that, a
functional is calculated along line over the image and
functional may not necessarily be integral. The Radon
transform is a special case of trace transform.

Trace Functional ( T )

Diametric Functional (

Fig. 1. Definition of Parameters of line tracing the image

Line can be represented as l ( , , t ) , where -Length of


the normal drawn from origin to the line l , -Angle made
by the normal drawn to the line with x-axis, t - is a parameter
along the line. If an image is represented as a function
f ( x, y ) and the set of all lines is represented as L ( , )
using a functiona1 T , the trace transform ( g ) is calculated
as
g (T , f , , ) T ( f ( , , t ))
(1)
The functional used to calculate Trace transform is
called Trace functional T . Another functional taken along the
columns of Trace Transform called the Diametric function P
converts the 2D Trace transform function g to 1D vector
which is only a function of called circus function h,
h( P, f , ) P ( g ( , ))
(2)
A functional applied on this circus function h , called
circus functional converts this 1D function to a single
number called Triple feature .So,
(, f ) ( h( ))
(3)
( f ) ( P (T ( f ( , , t ))))
(4)
The value of Triple Feature is influenced by the
functional T , P and . So, we can choose functional
selectively such that we can construct affine invariant features.

Circus Functional ( )
Triple Feature (
)
Fig.2. Steps in obtaining Triple Feature. Pre - Processed Signature at top. Below
is the Trace Transform-computed for angles 0 to 179 degrees, circus function
and finally the triple feature.

III.

AFFINE INVARIANT FEATURES

Petrou et al., in [1] and [6] proposed the construction of


affine invariant features and affine parameter estimation with
the help of Trace Transform. They also derived face signatures
or object signatures from trace transform for authentication.
Nasruddin et al., in [2] used affine invariant features from the
Trace Transform for jawi character recognition. They also
computed triple features, moments and angular radial
transform and reported that Trace signature feature derived
from Trace Transform has high recognition rate.
Assuming the intra class variations of signatures are only
affine distortions i.e., straight lines are preserved; we find
affine invariant features from the trace transform which will

be used as a feature. Functional are identified with properties


such that they are invariant to certain transformations of the
images. For constructing affine invariant features, shift
invariant functional is required. From here on they are simply
referred as invariant functional.
A functional is invariant if
( ( x b)) ( ( x )) b R
(5)
An invariant functional may also have the following
properties:
Scaling the independent variable by a , where a 0 , Scales
the result by a
the functional:

, where k is a number that characterizes

( ( ax)) a k ( ( x )) a 0

(6)

Scaling the function by c , where c 0 , scales the result by

c , where is another number that characterizes the


functional:

(c ( x)) c ( ( x )) c 0

(7)
In this paper Features extracted from Trace
Transform of signatures have been used for signature
verification for the first time. The Trace Transform is not
unique. It depends on the functional used along tracing lines
when computing the transform. Invariant functional that can
be used as Trace functional and diametric functional are given
in [6], [1]. Since circus function h( ) is a function over
when represented in polar form ( h( ), ) , looks like a
closed contour of certain shape. This circus function derived
for objects in the image is generally called objects signature
which is used as a feature vector. Here object signature is
derived for signature images and they may be referred as
signatures signature. Since this term is a bit confusing, from
here on we will refer it as circus functions only. Since circus
function is computed, the two functional used are
characterized by T , kT , P and k P .
In order to compare the circus functions of an image
and its affine distorted image, we convert the affine
relationship among circus functions to ones like in the case of
rotation, translation and scaling transforms. Since in cases of
distortion like rotation, translation and scaling the relationship
among the circus functions is just a scaling factor and we can
get rid of it easily as explained in [6]. To achieve that we
define a concept called associated circus ha ( ) and
normalize it.
1 v
ha ( ) sign(h( )) h( )
(8)
The value of

v kT P k P

is already known.

Now, Normalization of shapes (Closed contours) is


done by scaling them in such a way that each shape has the
same moment of inertia about all the lines which pass through
the origin of the axes. The 1D function ha ( ) can be

normalized by a well-known formula in polar co-ordinates as


explained in [1] and it does not involve slow 2D computations.
Finally if two normalized associated circus functions of and
image and its affine distorted version are hn1 ( ) and

hn 2 ( ) respectively then they are related by


hn1 ( ) ahn 2 ( )

Where a (a 0) and are some constants.


IV.

(9)

PREPROCESSING

The purpose of preprocessing is to make the images


ready for the feature extraction stage. In this case for
signatures images, signature is the main object in the image.
Grey level information is not significant for signatures. So,
we can binarize the images, like significant and nonsignificant information. Binarization is essentially done by
thresholding and threshold is decided by following otsus
algorithm. Before applying binarization the image is filtered
through 3X 3 averaging mask which is a low pass filter
used to obtain smooth edges for signatures. It also helps in
removing trivial noise if any.
After binarization, signature will be black and
background will be white. So, colors are inverted to
facilitate to measure information of the signature by any
mathematical computations. If any blobs or small group of
pixels may be formed during scanning which are not part of
signature are to be removed in preprocessing stage. A
bounding box is obtained around the signature and image is
resized to a fixed dimensions using scaling operation. Final
signature image after preprocessing looks like the one
shown in figure(2).
V.

EXPERIMENTS

For any two signature images under consideration, they are


compared and a similarity measure or score is obtained as
explained. Since functional are many, standard functional
given in [1] have been used. The combination of functional to
be used is found by trial and error. We are considering the
binary images of signatures, the median functional do not
capture much information. So, we avoid using these median
functions as Trace functional. So, these experiments are
carried out by taking G13, G17, and G19 as trace functional
and G4, G5 as Diametric functional. So, a six number of
circus functions are generated for each signature image. The
normalized associated circus functions
(10)
hnk ( ) k 1,2, 6.
are obtained from the circus functions as explained in section
III. So, the normalized associated circus functions are
calculated for both the signature images under consideration
using the same functional. Each circus function of signature is
compared with the corresponding circus function of other
image and a score of similarity is obtained. One of the two
circus functions is shifted continuously, and correlation
coefficient is found for all shift instances. This can be done by
using simple circular correlation. The maximum value of the
correlation coefficient is taken as similarity score for the two

different writer pairs is 16C1 X 16C1 256 . The


probability density function of the similarity scores obtained
for the considered writers. In ideal case the curves for same
writer and different writer should not overlap. In such cases it
is easy to define a decision threshold. In general, there is
always an overlap when the number of writers is more.
The performance of the signature verification system is
typically measured in terms of FAR, FRR and EER (Equal
Error Rate). False Acceptance is an imposter or a forgery is
being accepted as genuine and False Rejection is a genuine
signature is being rejected as a forgery. FRR and FAR are also
called Type-I and Type-II errors respectively. They are
defined as below:
FRR

Number of genuine signature rejected


X 100
Number of genuine signatures tested

FAR

Number of forgeries accepted


X 100
Number of forgeries tested

Same Writer / Different Writer PDF's


Same Writer
Different Writer

0.6
0.5
Probability ----- >

A. CEDAR Database
This database was built at CEDAR (Center of Excellence
for Document Analysis and Recognition), buffalo university.
Writers were asked to appose their signatures each in 2X2
space. 24 signature samples are collected from each of 55
writers. Some skilled forgers forged the signatures in the
database of 55 writers original signatures. 24 forgeries for
each writer are collected. So, the total size of the database is
2640 signatures.
For each writer, two-thirds of signatures are considered for
training and remaining one-third for testing. So, from 24
genuine and 24 forgeries of each writer, 16 genuine and 16
forgeries are used for training the system. Signature pairs are
formed as genuine-genuine and genuine-forgery pairs from the
training signatures to compute similarity score as explained
above. These signature pairs are called same writer pair and
different writer pairs respectively. Number of same writer
pairs per writer is 16C1 X 15C1 240 and number of

as in figure (3) with EER = 24.4% at a threshold of 0.069. An


ROC (Receiver Operating Characteristics) curve is drawn to
FAR and FRR as shown in figure (4), explains how FAR
changes with respect to FRR and vice versa. ROC Curve
enables to compare different Biometric system since it is
drawn between FAR and FRR thereby eliminating the
threshold. Figure (2) represents the Probability density
function of the similarity measures when comparisons of same
writer and different writers are considered. To normalize the
density function the amplitudes are divided by the total
number of scores considered.
By setting the threshold of the system at 0.069(EER),
testing the system with remaining test signatures an FAR and
FRR of 24.58% and 25.83% were obtained.

0.4
0.3
0.2
0.1
0

VI.

RESULTS

From the training signatures, three to four genuine


signatures are selected as reference templates. Similarity
scores obtained for remaining training signatures with respect
to the reference templates are used to draw the FAR and FRR
curves for the system. The FAR and FRR curves are obtained

0.1

0.15

0.2
0.25
0.3
0.35
Similarity Scores ----- >

0.4

0.45

0.5

FAR vs FRR Curves

100

FRR
FAR

90
80
70
60
50
40
30
20
10
0

These Rates should be as low as possible for a good


verification system and there is a Tradeoff between FAR and
FRR by varying the threshold. As FAR increases by varying
the threshold, FRR decreases and vice-versa.

0.05

Fig.3. The Probability density function of the similarity scores obtained


from the same writer pairs and different writer pairs

Percentage ----- >

signatures considered. To express in the form of distance


between the signatures, inverse cosine of the similarity score
is taken. In this way the distance is calculated for all six circus
function pairs and their sum is taken as a measure of
similarity. So, this measure will be low for most similar
signatures and high for dissimilar signatures.
.

0.05

0.1

0.15
0.2
Threshold ----- >

0.25

0.3

0.35

Fig.4. FAR vs. FRR Curves of the signature verification system using
trace transform: FRR corresponds to genuine signatures and FAR
corresponds to Forgery signatures.

VII.

CONCLUSION AND FUTUREWORK

REFERENCES
[1]

Receiver Operating Characteristics(ROC) curve

100
90

[2]

80

FRR ------ >

70
60

[3]

50
40

[4]

30
20

[5]

10
0

10

20

30

40

50
60
FAR ---- >

70

80

90

100

Fig.5. Receiver Operating Characteristic (ROC) curve: FAR vs. FRR


Definition
kT T
Code
G13

r f (r ) dr

T ( f (t ))

-2

-3

[6]

[7]

G14

T ( f (t ))

f ( r ) dr

[8]

1/ 2

G15

T ( f (t) median({ f (tk c) }tk 0,{ f (tk c) }tk 0 )

G16

T ( f (t) median({ (tk c1)( f (tk c1) }tk 0 ,{ f (tk c1) }tk 0 )

G17

T ( f (t ))

-1

i 4 ln r1

r10.5 f ( r1 ) dr1 -1.5

1/ 2

G18

T ( f (t ))

i 3 ln r1

i 5 ln r1

f ( r1 ) dr1

-1

T ( f (t ))

Code

Definition
R

r1 f ( r1 ) dr1

kP

G4

P ( g ( p ))

g ( p k 1 ) g ( p k ) 0

G5

P( g ( p)) median({g ( p k )}k ,{ g ( pk )}k ) 0

G7

P ( g ( p ))

Table2. Diametric functional

-0.75

14 dw

FourierTra nsf ( g ( p ))( w)

Petrou, Maria, and Alexander Kadyrov. "Affine invariant features from


the trace transform." Pattern Analysis and Machine Intelligence, IEEE
Transactions on26.1 (2004): 30-44.
Nasrudin, Mohammad F., Maria Petrou, and Leonidas Kotoulas. "Jawi
character recognition using the trace transform." Computer Graphics,
Imaging and Visualization (CGIV), 2010 Seventh International
Conference on. IEEE, 2010.
Kalera, Meenakshi K., Sargur Srihari, and Aihua Xu. "Offline signature
verification and identification using distance statistics." International
Journal of Pattern Recognition and Artificial Intelligence 18.07 (2004):
1339-1360
Pourreza Shahri, Reza, and Hamid Reza Pourreza. "Offline signature
verification using local radon transform and support vector machines."
International Journal of Image Processing 3 (2009).
Coetzer, Johannes, Ben M. Herbst, and Johan A. du Preez. "Offline
signature verification using the discrete radon transform and a hidden
Markov model." EURASIP Journal on Applied Signal Processing 2004
(2004): 559-571
Kadyrov, Alexander, and Maria Petrou. "The trace transform and its
applications." Pattern Analysis and Machine Intelligence, IEEE
Transactions on23.8 (2001): 811-828.
Sigari, Mohamad Hoseyn, Muhammad Reza Pourshahabi, and Hamid
Reza Pourreza. "Offline handwritten signature identification and
verification using multi-resolution gabor wavelet." International Journal
of Biometric and Bioinformatics 5 (2011).
Impedovo, Donato, and Giuseppe Pirlo. "Automatic signature
verification: the state of the art." Systems, Man, and Cybernetics, Part
C: Applications and Reviews, IEEE Transactions on 38.5 (2008): 609635.

Você também pode gostar