Anda di halaman 1dari 3

ISBN: 978-15-08772460-24

Date: 8.3.2015
PROCEEDINGS OF INTERNATIONAL CONFERENCE ON RECENT INNOVATIONS IN ENGINEERING AND TECHNOLOGY

Comparison of Offline Signature Verification


Scheme using Pseudo-dynamic features
Saroj Ramadas1, Geethu P.C2
1

PG Scholar, Dept. Of CSE, Vidya Academy of Science and Technology, Thrissur, India
Asst. Professor, Dept. Of CSE, Vidya Academy of Science and Technology, Thrissur, India
sarojramadas@gmail.com, geethu.p.c@vidyaacademy.ac.in,

Abstract Pattern recognition is an important research


area which deals with recognition and classification of
input data to one of the given set of classes. Automatic
handwritten signature verification is an application of
pattern recognition were the user authentication is done
using handwritten signatures which is a widespread
behavioral biometric as well as eliminate signature
forgery. The paper aims to study the working of two
offline signature system that uses pseudo-dynamic feature
extraction. The performance of these two systems are also
compared and analyzed.
Keywords biometry ,signature
extraction,LBP, HOG, GLCM

classification,

Offline signature verification system captures


signature images made using either scanner or camera. Other
steps involved are preprocessing, feature extraction and
classification. Preprocessing operations helps in incrementing
the quality of input image and recognition ratio of the system.
Feature extraction produces reduced representation of the
input data. Then, the extracted feature is used to train and
classify the signature into appropriate classes. The phases of
an offline signature verification system are as shown in fig.1
[1] :

feature

I. INTRODUCTION
Authentication is the process of verifying an
individual identity. One authentication method is to use
biometrics that refers to the inherent factors which describes
what a user is. Biometrics refers to physiological or behavioral
features which are hard to be mimicked or copied by another
individual. Hence, biometrics authentication is a well-known
method for an individuals identity verification. It verifies
user's identity by comparing an encoded value with a stored
value in the database of the concerned biometric characteristic.
Handwritten signatures are the most widespread
behavioral biometric used as means of identity verification. It
is used for a wide range of application which varies from
business and financial transactions, bank applications,
property and confidential documents validation, passport
verification and other fields that require identification. Thus,
automatic signature verification systems are essential to
authenticate a person suitably and also to eliminate signature
forgery.
Automatic verification using signatures can be either
online or offline depending on the data capturing method.
Signature data in former systems are acquired using devices
such as electronic tablet whereas data are acquired in offline
systems by cameras or scanner. Online verification systems
are accurate and dependable which uses dynamic features that
make forgery attempts difficult .However, offline systems are
still used since they are user-interactive, less invasive and
doesnt require any special acquisition instruments.
IAETSD 2015: ALL RIGHTS RESERVED

Fig.1: Phases of signature verification system

II. FEATURE EXTRACTION


Feature extraction is an essential step in pattern
recognition which extracts significant information about an
object and uses these features to classify that object. In offline
signature verification, feature extraction computes or
quantifies important characteristics of the signature input.
Feature extracted can be either global or local. If the features
are calculated over the entire image then, the extracted
features are global whereas local features describe signature
image properties over the set of its partitioned regions. Here,
three texture based pseudo-dynamic feature are been
compared. They are:
A. Local binary pattern (LBP)
Local binary pattern (LBP) [2] operator is a powerful
method used for texture classification. The basic LBP operator
compares each pixel in a image with its 8 neighbors. If the
value of center pixel's is greater than its neighbor's value, then
it is encoded as 1 otherwise it is encoded as 0. Thus, an 8-digit
binary number is generated for each pixel in the image. Then,
a histogram is computed which calculate the number of
occurrence of each binary code. The histogram calculated
provides information about the similar texture patterns.

www.iaetsd.in

13

ISBN: 978-15-08772460-24
Date: 8.3.2015
PROCEEDINGS OF INTERNATIONAL CONFERENCE ON RECENT INNOVATIONS IN ENGINEERING AND TECHNOLOGY

Feature vector is formed by combining the histogram


computed for each binary code.
Rotation invariant LBP is used in [3] which is
defined as:

P 1
P 1 P 1

C n 2 G (i, j ), i j n
n0
i 0 j 0

Texture entropy (E): Entropy measures the loss of


information or message in a transmitted signal and
also measures the image information. Entropy is
obtained using the formula:
P 1 P 1

E G (i, j ) log G (i, j )


i 0 j 0

Where,

P 1 P 1

P represents number of neighbors


R is the radius of the neighborhood.
g c denotes gray level value of center pixel
gp is the gray level of pth neighbor

i 0 j 0

B. Histogram of oriented gradients (HOG)


Histogram of oriented gradients (HOG) [2] is
another feature descriptor used for object identification. The
HOG descriptor technique computes the gradient orientation
information of each pixel in localized portions of an image.
For each localized portion, the gradient histogram is
calculated. Then, normalization is done across each block
which provides better invariance to illumination, shadowing,
and edge contrast .Finally, the HOG descriptors from all the
blocks are collected and is combined to form feature vector.
R-HOG (Rectangular-HOG) descriptor is used in [3].
Firstly, the first-order gradient of input signature image is
computed and the amplitude and direction of every pixel of
the gradient image is obtained .Then, the signature image is
divided into fixed number of rectangular zones. For each zone,
the histogram of oriented gradient is calculated which is
distributed between 0 to 180 degree. Then, normalization is
done for each histogram using L1-norm algorithm. Combining
the histogram from all the rectangular grids helps in achieving
feature vector.
C. Gray-Level Co-occurrence Matrix (GLCM)
Gray-level co-occurrence matrix (GLCM) [3] is a
statistical method used to analyze texture of an image. It is
also known as the gray-level spatial dependence matrix.
GLCM calculates how often a pixel with gray-level value i
occurs either horizontally, vertically or diagonally to adjacent
pixel with value j. After creating GLCM, the following
texture features are extracted:
Texture homogeneity (H): It is the measure of the
closeness of the elements distributed in the GLCM to
the GLCM diagonal. It is defined using the formula:
P 1 P 1

H G (i, j )
i 0 j 0

Texture correlation (O): It defines measures of the


joint probability of the occurrence of the specified
pixel pairs in GLCM and is defined as:

Texture Contrast (C): It is the measures of the local


variations in the gray-level co-occurrence matrix
which is defined as:

IAETSD 2015: ALL RIGHTS RESERVED

ijG (i, j ) mi m j

i j

Here, G is a square matrix of size P were P denotes the gray


level. Gray level value of a pixel is represented y i and j.
mi and i describes mean and standard deviation of G( i , j) rows
and mj and j describes mean and standard deviation of G( i, j)
columns.

III. SIGNATURE CLASSIFICATION APPROACHES


The last phase of offline signature verification system
is its verification or classification stage .In this stage, input
signature is classified to any of the two classes i.e. the positive
(genuine) or negative (forgery) class. Correctness of
verification stage depends upon the training provided to the
classifier using the extracted features.
Classification in [3] is performed using two different
approaches i.e. Writer-dependent Support Vector Machine
(WSVM) classifiers and Global Real Adaboost method. The
training of these classifiers are done using three pseudodynamic features namely rotation invariant LBP, GLCM and
Rectangular HOG (R-HOG). SVM is a statistical learning
technique designed for class two problems that creates a hyper
plane between data sets to indicate which class it belongs to.
In WSVM approach, SVM model is trained for each signer in
the database. To avoid over-fitting caused during the training
phase, linear kernel is used in WSVM. Global Real Adaboost
is a writer-independent classifier. To model write-independent
classifier, a sub-feature vector called dissimilarity feature
vector is computed based on original features. If signature
samples are genuine, then all the element of dissimilarity
feature vector will be close to 0 is known as positive vector .If
samples are forgery, the element in feature vector will be far
from 0and is known negative vector. Once the positive and
negative training samples are obtained, classifier is trained.
The combined basic LBP and HOG features are used to train
Global SVM (GSVM) and User-dependent SVM (USVM) in
[2]. GSVM is a user-independent classifier which helps in
distinguishing between genuine and skilled forgery signatures
while USVM is trained for each user which differentiates
genuine and random forgeries. In GSVM, the difference
vectors are computed by comparing the features of
genuine/skilled forgery signature with the features of reference

www.iaetsd.in

14

ISBN: 978-15-08772460-24
Date: 8.3.2015
PROCEEDINGS OF INTERNATIONAL CONFERENCE ON RECENT INNOVATIONS IN ENGINEERING AND TECHNOLOGY

signature. The resultant vector is normalized such that each


element of the vector denotes the standard deviation between
query feature and the reference feature. In USVM, feature
vector derived from the reference and random forgery
signature for each user is used for training.

Table 2: Performance Evaluation of WSVM and


Adaboost classifier
Pseudo-dynamic
features

Classifier

EER(%)

Rotation-invariant
LBP,
GLCM & R-HOG

WSVM

9.94%

Global Real
Adaboost

7.66%

IV. PERFORMANCE ANALYSIS


Performance of a Static/Offline signature verification
system is analyzed in terms of error rates. The common error
rates used are False Rejection Ratio (FRR), False Acceptance
Ratio (FAR), Average Error Rate (AER) and Equal error rate
(EER).
FRR and FAR are also known as Type I and Type II
error rates. The False Rejection Ratio (FRR) relates to the
genuine signatures that are rejected by the system whereas
False Acceptance Ratio (FAR) is related to the forgeries that
are classified as genuine signatures by the system.
Another error rate is Average Error Rate (AER)
which relates to the total error of the system that is the
occurrence of type 1 and type 2 error rates together. If the
system is set to have percentage of false acceptances and false
rejections to be approximately equal, then equal error rate
(EER) is calculated. The formulae to calculate the error rates
are:
1. FAR=Number of forgeries accepted x100
Number of forgeries tested
2.

FRR=Number of genuines rejected


Number of genuines tested

3.

AER=FAR + FRR

x100

In an automatic signature verification system if the


error rate obtained is lesser than the system provides better
recognition. While comparing the performance analysis of
methods discussed in [2] and [3], the lesser error rate was
provided by WSVM and Adaboost, thus providing more
accurateness than that of GSVM and USVM.
V. CONCLUSION
The working of three pseudo-dynamic feature
extraction techniques namely LBP, GLCM and HOG was
studied. The different classification approaches used for
classifying signature images were also studied.
After
analyzing the performance of these two systems using pseudodynamic features , it can be concluded that, WSVM and Real
Adaboost classifiers provides is the better recognition ratio
when tested using GPDS dataset.

VI. REFERENCES

The GSVM and USVM were tested in two phases. In


first phase, both classifiers are trained using 12 reference
signatures whereas in second phase training was done using 5
reference signatures. Analysis of these testing shows that more
performance is provided by USVM than that of GSVM since
it is modeled for each user. But when both classifiers and
features are combined, it provides better performance. The
performance analysis was done using GPDS dataset and the
result obtained is shown in the table 1.
Table 1: Performance Evaluation of combined
GSVM and USVM classifier
Pseudo-dynamic
features

EER(%)
(12 references)

EER(%)
(5 references)

HOG and LBP

15.41%

17.65%

[1]

[2]

[3]

[4]
[5]

Vitthal K. Bhosale, Dr. Anil R. Karwankar, Automatic Static Signature


Verification Systems: A Review International Journal of
Computational Engineering Research (ijceronline.com) Vol. 3 Issue. 2,
ISSN 2250-3005, February 2013
Yilmaz, M.B. and Yanikoglu, B. and Tirkaz, C. and Kholmatov, A,
Offline signature verification using classifier combination of HOG and
LBP features, IJCB 2011, pg. 1-7.
Juan Hu, Youbin Chen, Offline Signature Verification Using Real
Adaboost Classifier Combination of Pseudo-dynamic Features 12th
International Conference on Document Analysis and Recognition,2013
Juan Hu, Youbin Chen, Fusion of Features and Classifiers for Off-line
Handwritten Signature Verification, IEEE 2011
N. Dalal and B. Triggs. Histograms of oriented gradients for human
detection. CVPR'05, Volume 1, p.p. 886-893, Washington DC, 2005.

The WSVM and Real Adaboost method is tested


using GPDS dataset and the performance evaluation is shown
in table 2. In WSVM, training was done using 10 reference
signatures and testing was performed using genuine and
skilled forgery. Whereas in Real Adaboost method, training
set was selected with 10 reference and genuine signatures of
each user and dissimilarity vectors of 100 writers is computed
for training the Real Adaboost classifier.

IAETSD 2015: ALL RIGHTS RESERVED

www.iaetsd.in

15

Anda mungkin juga menyukai