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

net/publication/258650160

Handwritten Signature Verification using Neural Network

Article · January 2012

DOI: 10.5120/ijais12-450114

CITATIONS READS

48 4,711

2 authors:

Ashwini Pansare Shalini Bhatia

Fr. Conceicao Rodrigues College of Engineering University of Canberra
5 PUBLICATIONS   66 CITATIONS    41 PUBLICATIONS   297 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Multimodal affective computing for automated depression analysis View project

Image Denoising View project

All content following this page was uploaded by Shalini Bhatia on 16 January 2015.

The user has requested enhancement of the downloaded file.

 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

Handwritten Signature Verification using

Neural Network

Ashwini Pansare Shalini Bhatia

Assistant Professor in Computer Engineering Associate Professor in Computer Engineering
Department, Mumbai University, India Department, Mumbai University, India

ABSTRACT Vigorous research has been pursued in handwriting analysis

A number of biometric techniques have been proposed for and pattern matching for a number of years. In the area of
personal identification in the past. Among the vision-based Handwritten Signature Verification (HSV), especially offline
ones are face recognition, fingerprint recognition, iris HSV, different technologies have been used and still the area
scanning and retina scanning. Voice recognition or signature is being explored. In this section we review some of the recent
verification are the most widely known among the non-vision papers on offline HSV. The approaches used by different
based ones. As signatures continue to play an important role researchers differ in the type of features extracted, the training
in financial, commercial and legal transactions, truly secured method, and the classification and verification model used.
authentication becomes more and more crucial. A signature
by an authorized person is considered to be the “seal of
1.1 Hidden Markov Models Approach
approval” and remains the most preferred means of Hidden Markov Model (HMM) is one of the most widely used
authentication. The method presented in this paper consists of models for sequence analysis in signature verification.
image prepossessing, geometric feature extraction, neural Handwritten signature is a sequence of vectors of values
network training with extracted features and verification. A related to each point of signature in its trajectory. Therefore, a
verification stage includes applying the extracted features of well chosen set of feature vectors for HMM could lead to the
test signature to a trained neural network which will classify it design of an efficient signature verification system. These
as a genuine or forged. Models are stochastic models which have the capacity to
absorb the variability between patterns and their similarities.
Keywords In HMM stochastic matching (model and the signature) is
Biometrics, error back propagation algorithm, center of mass, involved. This matching is done by steps of probability
neural network, and normalized area of signature. distribution of features involved in the signatures or the
probability of how the original signature is calculated. If the
results show a higher probability than the test signatures
1. INTRODUCTION probability, then the signatures is by the original person,
As signature is the primary mechanism both for authentication otherwise the signatures are rejected.
and authorization in legal transactions, the need for efficient In paper [6], a system is introduced that uses only global
auto-mated solutions for signature verification has increased features. A discrete random transform which is a sinograph is
[1].Unlike a password, PIN, PKI or key cards – identification calculated for each binary signature image at range of 0 − 360,
data that can be forgotten, lost, stolen or shared – the captured which is a function of total pixel in the image and the intensity
values of the handwritten signature are unique to an individual per given pixel calculated using non overlapping beams per
and virtually impossible to duplicate. Signature verification is angle for X number of angles. Due to this periodicity, it is
natural and intuitive. The technology is easy to explain and shift, rotation and scale invariant. A HMM is used to model
trust. The primary advantage that signature verification each writer signature. The method achieves an AER of 18.4%
systems have over other type’s technologies is that signatures for a set of 440 genuine signatures from 32 writers with 132
are already accepted as the common method of identity skilled forgeries.
verification [2].
A signature verification system and the techniques used to 1.2 Neural Networks Approach
solve this problem can be divided into two classes Online and The main reasons for the widespread usage of neural networks
Off-line [3].On-line approach uses an electronic tablet and a (NNs) in pattern recognition are their power and ease of use.
stylus connected to a computer to extract information about a A simple approach is to firstly extract a feature set
signature and takes dynamic information like pressure, representing the signature (details like length, height,
velocity, speed of writing etc. for verification purpose. Off- duration, etc.), with several samples from different signers.
line signature verification involves less electronic control and The second step is for the NN to learn the relationship
uses signature images captured by scanner or camera. An off- between a signature and its class (either “genuine” or
line signature verification system uses features extracted from “forgery”). Once this relationship has been learned, the
scanned signature image. The features used for offline network can be presented with test signatures that can be
signature verification are much simpler. In this only the pixel classified as belonging to a particular signer. NNs therefore
image needs to be evaluated. But, the off-line systems are are highly suited to modeling global aspects of handwritten
difficult to design as many desirable characteristics such as signatures.
the order of strokes, the velocity and other dynamic The proposed system in [7] uses structure features from the
information are not available in the off-line case [4, 5]. The signatures contour, modified direction feature and additional
verification process has to wholly rely on the features that can features like surface area, length skew and centroid feature in
be extracted from the trace of the static signature images only. which a signature is divided into two halves and for each half

44
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

a position of the centre of gravity is calculated in reference to Contribution:

the horizontal axis. For classification and verification two In this paper we present a model in which neural network
approaches are compared the Resilient Back propagation classifier is used for verification. Signatures from database are
(RBP) neural network and Radial Basic Function(RBF) using pre-processed prior to feature extraction. Features are
a database of 2106 signatures containing 936 genuine and extracted from pre-processed signature image. These extracted
1170 forgeries. These two classifiers register 91.21% and 88 features are then used to train a neural network. In verification
% true verification respectively. stage, on test signatures pre-processing and feature extraction
is performed. These extracted features are then applied as
1.3 Template matching approach input to a trained neural network which will classify it as a
Fang et al. [8] proposed two methods for the detection of genuine or forged signature.
skilled forgeries using template matching. One method is
based on the optimal matching of the one-dimensional Organization of the paper: The rest of the paper is organized
projection profiles of the signature patterns and the other is as follows. In section 2, the signature verification model is
based on the elastic matching of the strokes in the two- described. In section 3, the algorithm is presented .Results
dimensional signature patterns. Given a test signature to be generated by the system is presented in section 4 and
verified, the positional variations are compared with the concluded in section 5.
statistics of the training set and a decision based on a distance 2. METHODOLOGY
measure is made. Both binary and grey-level signature images
are tested. The average verification error rate of 18.1% was In this section, block diagram of system is discussed. Fig. 1
achieved when the local peaks of the vertical projection gives the block diagram of proposed signature verification
profiles of grey-level signature images were used for system which verifies the authenticity of given signature of a
matching and with the full estimated covariance matrix person. The design of a system is divided into two stages:
incorporated.
1) Training stage 2) Testing stage

1.4 Statistical approach A training stage consist of four major steps 1) Retrieval of a
Using statistical knowledge, the relation, deviation, etc signature image from a database 2) Image pre-processing 3)
between two or more data items can easily be found out. To Feature extraction 4) Neural network training
find out the relation between some set of data items we A testing stage consists of five major steps 1) Retrieval of a
generally follow the concept of Correlation Coefficients. In signature to be tested from a database 2) Image pre-processing
general statistical usage refers to the departure of two 3) Feature extraction 4) Application of extracted features to a
variables from independence. To verify an entered signature trained neural network 5) Checking output generated from a
with the help of an average signature, which is obtained from neural network.
the set of, previously collected signatures, this approach
follows the concept of correlation to find out the amount of Training stage
divergence in between them.
A unique method is introduced in [9]. In this approach various Sample database
features are extracted which include global features like image
gradient, statistical features derived from distribution of pixels Testing Stage
of a signature and geometric and topographical descriptors
like local correspondence to trace of the signature. The Input signature Input signature
classification involves obtaining variations between the
signatures of the same writer and obtaining a distribution in
distance space. For any questioned signature the method
obtains a distribution which is compared with the available Pre processing Pre processing
known and a probability of similarity is obtained using a
statistical Kolmorogorv-Smirnov test. Using only 4 genuine
samples for learning, the method achieves 84% accuracy Feature extraction Feature extraction
which can be improved to 89% when the genuine signature
sample size is increased. This method does not use the set of
forgery signatures in the training/learning.
Neural network Applying extracted features
1.5 Support Vector Machine training to trained neural network
Support Vector Machines (SVMs) are machine learning
algorithms that uses a high dimensional feature space and training
estimate differences between classes of given data to
generalize unseen data. The system in [10] uses global, Verification
directional and grid features of the signature and SVM for
classification and verification. The database of 1320
signatures is used from 70 writers. 40 writers are used for
Verification Result
training with each signing 8 signatures thus a total of 320
signatures for training. For initial testing, the approach uses 8
original signatures and 8 forgeries and achieves FRR 2% and
FAR 11%. Fig 1: Block Diagram of Handwritten Signature
Verification System using NN.

45
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

Fig. 2 shows one of the original signature image taken from 2.1.3 Bounding box of the signature:
a database and all the subsequent figures show the resultant In the signature image, construct a rectangle encompassing
signature image obtained after performing the steps the signature. This reduces the area of the signature to be used
mentioned in an algorithm. for further processing and saves time. Fig. 5 shows signature
enclosed in a bounding box.

Fig 2: Signature image from the database

2.1 Pre-processing
The pre processing step is applied both in training and testing Fig 5: Signature image with a bounding box
phases. Signatures are scanned in gray. The purpose in this
phase is to make signature standard and ready for feature 2.2 Feature Extraction
extraction. The pre-processing stage improves quality of the
image and makes it suitable for feature extraction [11]. The The choice of a powerful set of features is crucial in signature
prepossessing stage includes verification systems. The features that are extracted in this
phase are used to create a feature vector. A feature vector of
dimension 24 has been used to uniquely characterize a
2.1.1 Converting image to binary candidate signature. These features are extracted as follows:
A gray scale signature image is converted to binary to make 1) Maximum horizontal and vertical histogram
feature extraction simpler. Horizontal histogram is calculated by going through each row
of the signature image and counting number of black pixels. A
2.1.2 Image resizing row with maximum number of black pixels is recorded as
The signatures obtained from signatory are in different sizes maximum horizontal histogram. Similarly, a vertical
so, to bring them in standard size, resizing is performed, histogram is calculated by going through each column of the
which will bring the signatures to standard size 256*256 as signature image and finding a column with maximum number
shown in Fig. 3. of black pixels.

2) Center of mass
Split the signature image in two equal parts and find center of
mass for individual parts.

3) Normalized area of signature

It is the ratio of area of signature image to the area of
signature enclosed in a bounding box. Area of a signature is
the number of pixels comprising it.
Fig. 3 Signature image after resizing
𝑆𝑖𝑔𝑛𝑎𝑡𝑢𝑟𝑒 𝐴𝑟𝑒𝑎
𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑎𝑟𝑒𝑎 = 𝐴𝑟𝑒𝑎 Eq. (1)
𝑒𝑛𝑐𝑙𝑜𝑠𝑒𝑑 𝑖𝑛 𝑎 𝑏𝑜𝑢𝑛𝑑𝑖𝑛𝑔 𝑏𝑜𝑥
2.1.3 Thinning
Thinning makes the extracted features invariant to image 4) Aspect Ratio
characteristics like quality of pen and paper. Thinning means It is the ratio of width of signature image to the height of the
reducing binary objects or shapes to strokes that are single image. This is done because width or height of person’s
pixel wide. Fig. 4 shows thinned signature image.
signature may vary but its ratio remains approximately equal.

𝑤𝑖𝑑𝑡 𝑕 𝑜𝑓 𝑠𝑖𝑔𝑛𝑎𝑡𝑢𝑟𝑒 𝑖𝑛 𝑎 𝑏𝑜𝑢 𝑛𝑑𝑖𝑛𝑔 𝑏𝑜𝑥

𝐴𝑠𝑝𝑒𝑐𝑡 𝑅𝑎𝑡𝑖𝑜 = 𝐻𝑒𝑖𝑔 𝑕𝑡 Eq. (2)
𝑜𝑓 𝑠𝑖𝑔𝑛𝑎𝑡𝑢𝑟𝑒 𝑖𝑛 𝑎 𝑏𝑜𝑢𝑛𝑑𝑖𝑛𝑔 𝑏𝑜𝑥

5) Tri surface feature

Two different signatures may have same area .so; to increase
the accuracy of the features three surface feature has been
used. In this, a signature is divided into three equal parts and
area for each part is calculated. Eq. (1) is then used to
Fig 4: Signature image after Thinning

46
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

calculate normalized area of each part. Figure (6) shows tri 3. ALGORITHM
surface feature. Table 1 gives algorithm for the handwritten signature
verification system in which neural network is used for
verifying the authenticity of signatures.

TABLE 1 Algorithm for Handwritten Signature

Verification using Neural Network

Input: signature from a database

Output: verified signature classified as genuine or forged.
1. Retrieval of signature image from a database.
Fig 6: Tri surface feature 2. Preprocessing the signatures.
2.1 Converting image to binary.
6) The six fold surface feature 2.2 Image resizing.
Divide a signature in three equal parts and find bounding box 2.3 Thinning.
for each part. Then calculate centre of mass for each part. 2.4 Finding bounding box of the signature.
Draw a horizontal line passing through centre of mass of each 3. Feature extraction
part and calculate area of signature above and below centre of 3.1 Maximum horizontal and vertical histogram
mass within a bounding box. This provides six features. 3.2 Centre of mass
3.3 Normalized area of signature
3.4 Aspect ratio
3.5 The tri surface feature
3.6 The six fold surface feature
3.7 Transition feature
4. Creation of feature vector by combining extracted features.
5. Normalizing a feature vector.
6. Training a neural network with a normalized feature vector.
7. Steps 1 to 5 are repeated for testing signatures.
8. Applying normalized feature vector of test signature to
trained neural network.
Fig 7: Six fold surface feature
9. Using a result generated by the output neuron of the neural
7) Transition feature network declaring a signature as a genuine or forged.
Traverse a signature image in left to right direction and each
time there is a transition from 1 to 0 or 0 to 1, calculate a ratio
between the position of transition and the width of image 4. RESULTS AND DISCUSSION
For training and testing of the system many signatures are
traversed and record it as a feature. Repeat a same process in
used. The results given in this paper are obtained using the
right to left, top to bottom and bottom to top direction. Also
“Grupo de Procesado Digital de Senales” (GPDS) signature
calculate total number of 0 to 1 and 1 to 0 transitions. This
database [12]. The results provided in this research used a
provides ten features.
total of 1440 signatures. Those 1440 signatures are
comprised of 30 sets (i.e. from 30 different people) and, for
2.2.4 Creation of feature vector each person there are 24 samples of genuine signatures and 24
A feature vector of size 24 is formed by combining all the samples of forgeries. Figure 6 shows some of the signatures in
extracted features as discussed in section 2.2.
the GPDS database. To train the system, a subset of this
2.2.5 Training a neural network database was taken comprising of 19 genuine samples taken
Extracted 24 feature points are normalized to bring them in
from each of the 30 different individuals and 19 forgeries
the range of 0 to 1.These normalized features are applied as
input to the neural network. made by different person for one signature. The features
extracted from 19 genuine signatures and 19 forged signatures
2.3 Verification for each person were used to train a neural network. The
In the verification stage, a signature to be tested is pre- architecture of neural network used has input layer, hidden
processed and feature extraction is performed on pre layer and output layer [13]. Number of neurons in the input
processed test signature image as explained in section 2.2 to layer are 24, 24 neurons in the hidden layer and one neuron in
obtain feature vector of size 24. After normalizing a feature the output layer. After applying a feature vector of test
vector it is fed to the trained neural network which will signature if the output neuron generates value close to +1 test
classify a signature as a genuine or forged.
signature is declared as genuine or if it generates value close
to -1 it is declared as forged.Fig.8 shows performance graph
of the training a two layer feed forward neural network using
Error Back Propagation Algorithm (EBPTA).

47
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

Performance is 0.0999998, Goal is 0.1

When the neural network was presented with 570 genuine
1
10 signatures from 30 different persons, it classified all 570
genuine signatures as genuine and when 570 forged signatures
from 30 different persons were applied it recognized all 570
signatures as forgeries. Thus FAR and FRR of the system is
0%.Hence, Correct Classification Rate (CCR) is 100% for Re-
call.
Training-Blue Goal-Black

0
10

4.3 Result of testing neural network with

new signature samples from database
When the signatures samples not used for training neural
-1
10 network are applied as test signatures to the trained neural
network, it is called Generalization. The result of
generalization is shown in Table 3.

TABLE 3 Result of Testing Neural Network with New

-2 Signatures Samples from Database.
10
0 1 2 3 4 5 6 7 8
80786 Epochs 4
x 10 Samples Genuine Forged FAR FRR CCR In
presented Generaliz
Fig. 8 Performance graph of handwritten ation
signature verification using neural network.
150 120 30 - 20%
4.1 Performance Analysis genuine
False Acceptance Rate (FAR), False Rejection Rate (FRR)
and Correct Classification Rate (CCR) are the three 150 22 128 14.66 -
parameters used for measuring performance of system. FAR forged % 82.66%
is calculated by eq. (3), FRR is calculated by eq. (4) and CCR
is calculated by eq. (5).

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑜𝑟𝑔𝑒𝑟𝑖𝑒𝑠 𝑎𝑐𝑐𝑒𝑝𝑡𝑒𝑑

𝐹𝐴𝑅 = ∗ 100 Eq. (3) The neural network when presented with 150 genuine
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑜𝑟𝑔𝑒𝑟𝑖𝑒𝑠 𝑡𝑒𝑠𝑡𝑒𝑑
signatures from 30 different persons classified 120 signatures
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙𝑠 𝑟𝑒𝑗𝑒𝑐𝑡𝑒𝑑
𝐹𝑅𝑅 = ∗ 100 Eq. (4) out of 150 as genuine and 30 signatures as forgeries. Thus
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙𝑠 𝑡𝑒𝑠𝑡𝑒𝑑
FRR of the system is 20% .When 150 forged signatures were
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒𝑠 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑟𝑒𝑐𝑜𝑔𝑛𝑖𝑧𝑒𝑑 given as input to neural network, it classified 22 signatures as
𝐶𝐶𝑅 = ∗ 100 Eq. (5)
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒𝑠 𝑡𝑒𝑠𝑡𝑒𝑑 genuine and 128 as forgeries. Thus FAR of the system is
14.66%. And hence the Correct Classification Rate is 82.66%
4.2 Results of Testing Neural network with for generalization.
trained samples
5. CONCLUSION
The genuine and forged signature samples used for training
neural network is applied in the testing phase to check This paper presents a method of handwritten signature
whether neural network classifies it correctly as genuine or verification using neural network approach. The method uses
forged. This is called Recall. The result of recall is as shown features extracted from preprocessed signature images. The
in Table 2. extracted features are used to train a neural network using
error back propagation training algorithm. As shown in Table
TABLE 2 Result of Testing Neural Network with Trained 2 CCR in recall is 100%. The network could classify all
Signature Samples. genuine and forged signatures correctly. When the network
was presented with signature samples from database different
Samples Genuine Forged CCR in Recall than the ones used in training phase, out of 300 such
signatures (150 genuine and 150 forged) it could recognize
presented 248 signatures correctly. Hence, the correct classification rate
of the system is 82.66% in generalization as shown in Table 3.
570 genuine 570 0 100%

570 forged 0 570 100%

48
 International Journal of Applied Information Systems (IJAIS) – ISSN : 2249-0868
Foundation of Computer Science FCS, New York, USA
Volume 1– No.2, January 2012 – www.ijais.org

6. REFERENCES Classification,” International Joint Conference on

Neural Networks, 2006.
[1] Prasad A.G. Amaresh V.M. “An offline signature
verification system” [8] S.Srihari. K. M. Kalera. and A. XU, “Offline Signature
Verification and Identification Using Distance
[2] Prashanth CR,KB Raja,KR Venugopal, LM
Statistics,” International Journal of Pattern Recognition
Patnaik,”Standard Scores Correlation based Offline
And Artificial Intelligence ,vol. 18, no. 7, pp. 1339–1360,
signature verification system”, International
2004.
Conference on advances in computing, control and
telecommunication Technologies 2009 [9] H. S. Srihari and M. Beall, “Signature Verifcation Using
Kolmogrov Smirnov Statistic,”Proceedings of
[3] R. Plamondon and S.N. Srihari, "Online and Offline
International Graphonomics Society,Salemo Italy , pp.
Handwriting Recognition: A Comprehensive Survey",
152–156, june,2005.
IEEE Tran. on Pattern Analysis and Machine
Intelligence, vol.22 no.1, pp.63-84, Jan.2000. [10] T.S. enturk. E. O¨ zgunduz. and E. Karshgil, “
Handwritten Signature Verification Using Image
[4] J Edson, R. Justino, F. Bortolozzi and R. Sabourin, "An
Invariants and Dynamic Features,” Proceedings of the
off-line signature verification using HMM for
13th European Signal Processing Conference EUSIPCO
Random,Simple and Skilled Forgeries", Sixth
2005,Antalya Turkey, 4th-8th September, 2005.
International Conference on Document Analysis and
Recognition, pp.1031-1034, Sept.2001. 211-222, [11] Ramachandra A. C ,Jyoti shrinivas Rao”Robust Offline
Dec.2000. signature verification based on global features” IEEE
International Advance Computing Conference ,2009.
[5] J Edson, R. Justino, A. El Yacoubi, F. Bortolozzi and R.
Sabourin, "An off-line Signature Verification System [12] Martinez, L.E., Travieso, C.M, Alonso, J.B., and Ferrer,
Using HMM and Graphometric features", DAS 2000 M. Parameterization of a forgery Handwritten Signature
Verification using SVM. IEEE 38thAnnual 2004
[6] B. Herbst. J. Coetzer. and J. Preez, “Online Signature
International Carnahan Conference on Security
Verification Using the Discrete Radon Transform and a
Technology ,2004 PP.193-196
Hidden Markov Model,” EURASIP.Journal on Applied
Signal Processing, vol. 4, pp. 559–571, 2004. [13] “An Introduction to Artificial Neural Systems” by Jacek
M. Zurada, West Publishing Company 1992.
[7] M. Blumenstein. S. Armand. and Muthukkumarasamy,
“Off-line Signature Verification using the Enhanced
Modified Direction Feature and Neural based

49

View publication stats