Static Digital Signature Recognition and

Verification using Neural Networks

Prachi Chauhan, Subhash Chandra and Sushila Maheshkar
Department of Computer Science and Engineering
Indian School of Mines, Dhanbad
Email: [email protected], [email protected], [email protected]

Abstract—Biometrics play a crucial role in establishing an weight. The sum of weighted inputs represents the
individuals identity. A signature is one of the most widely strength of input signal and a step function is applied on this
recognized way to authorize transactions and authenticate the sum to determine its output. This output is further
human identity as compared to other electronic identification transmitted to other perceptions’. Artificial Neural
methods such as fingerprint and retina scans. Due to a huge Networks are found very useful in determining patterns in
demand for authentication, fast algorithms need to be data, which the human brain is naturally good at. Be it
assimilated for signature recognition and verification. Human Character Recognition, Image Compression, Stock Market
signatures can be treated as an image and the techniques of Prediction, Travelling Salesman Problem or even ideas as
neural networks can be applied to them for recognition and astonishing as an electronic nose, ANNs have found their
verification. This paper exploits a database of samples of offline utility to a great extent.
signatures that are captured in an image format and this
database is used to train the neural network. In this paper, the pattern detection and memory retention
properties of ANNs have been utilized to verify and
Keywordsí Offline-Signature, Neural Network, Biometric, recognize authentic offline signatures. Input to the network
Training, Testing, Perceptron. being digitized images of the signatures, hence making it a
mere pattern recognition task. The verification process is
carried out by comparing each individual test signatures
I. INTRODUCTION characteristic parameters with those of the signatures
The handwritten signatures have been used for decades already present in the database. The process results in a
to authenticate human identities. A signature is unique for Boolean value which is true for a correct match and false for
each individual pertaining to pressure, intensity, speed, area an incorrect match.
and such other properties, hence they often become difficult
to verify with the naked eye [1]. Therefore it becomes II. OVERVIEW OF BIOMETRICS AND SIGNATURE
essential to verify signature as an image rather than a VERIFICATION
collection of letters and symbols. As we know that the Personal verification and recognition is an actively
Nervous system in a living organism is one of the most vital growing area of research which includes tremendous
biological systems for the sustenance of life. Neurons or the exploitation of biometrics of an individual [2]. Biometrics
nerve cells are the functional units of nervous system which are characterized by personal traits such as fingerprint,
carry impulses or messages to each and every corner of our retina, iris, odor and gait; they are reliable because of their
body to generate various responses to various stimuli. natural uniqueness for every person [3]. However,
Analogous to this biological neural network system in the fingerprint or retina scans do not certify a clear intent of the
brain, scientists have developed the concept of a person as a handwritten signature does. Signature
mathematical model of a neuron, perceptron, which forms verification are classified as static and dynamic. Static
the artificial neural network which can be used to perform signature recognition operates by obtaining handwritten
deductions and pattern observations like the human brain. In signatures on paper and then digitizing them through an
the brain, if the sum of the input electrochemical signals optical scanner or a camera. Signatures are then recognized
received is sufficiently powerful to activate the neuron, the by their shape, angle of inclination, curvature, etc. Creating
signal is transmitted to any other connecting neurons. a system with the ability to recognize hand written signature
Similarly, in a perceptron, input signals are numerical values and verify its authenticity is a challenging issue to deal with
which can be modulated by multiplying by a value called a [4]. The task is narrowed down to drawing the threshold of
the range of genuine variation whereas in case of dynamic
signature verification signatures are taken on pressure IV. 4. PROPOSED ALGORITHM
sensitive tablets or stylus operated personal digital- In this paper, the recognition and verification of offline
assistants. It uses the spatial coordinate information, signature samples using ANNs is presented as it follows a
inclination, pen speed, pressure data, etc. to recognize the paradigm which models human learning patterns [9]. The
signatures. Forgery is an act of deceit designed to falsify an proposed technique as shown in Figure 1. is explained in
individual’s identity or counterfeit a document [5]. Although detail in the next subsection:
technology has given a lot to the world in positive, it has
indeed given efficient tools to criminals to commit crimes Step 1 Setting up of a database of scanned signatures.
like forgery [6].
TABLE 1. SAMPLES OF SELF CREATED DATABASE
Random forgery Signature is forged to be a close match
in shape to the original one. In this, the forger tries to imitate S No. G1 G2 G3 F1 F2
the movement by a quick scribble to make a similar looking
signature. Simple forgery The signer accesses the original
signature and practices it to perfection for misuse. The
forger creates a more precise imitation of the original P1
signature through a slow and calculated movement of the
hand. Skilled forgery In this, the forger creates a perfectly
P2
designed copy of the original signature by tracing an outline
of the signature in the similar location and angle and then
filling it with ink, or by other such techniques. In this paper, P3
we focus on the geometric features based on gray scale
information from image containing handwritten signatures. P4
The organization of the paper is as follows. In Section 2 we
discuss about the biometrics, Section 3 describes
P5
preliminaries of the artificial neural network, Section 4
explains proposed algorithm (preprocessing, feature
extraction), Section 5 training and testing . Section 6 Step 2 Normalization of images through the preprocessing
concludes the paper. steps, conversion to grey scale, resizing and conversion to
binary images.
III. PRELIMINARIES Step 3 Training the neural network with sample images from
Artificial Neural Networks ANNs): Creating a system the database.
with the ability to recognize hand written signature and
verify its authenticity is a challenging issue to deal with, it Step 4 Testing the ANN with a genuine and forged signature
generally goes out of the conventional practice of writing not present in the database.
algorithms [7]. The only feasible solution has its motivation
from the concept of neurons in the human brain. A neuron Step 5 Recognition and Verification of the signature.
receives excitatory input sufficiently large compared with its
inhibitory input and then sends a spike of electrical activity
down its axon [8]. Learning occurs when the influence of
one neuron on another changes by impulses firing through
the synapse. A perceptron is designed to simulate the brain
neuron under a smaller processing power than the human
brain [5]. It can also be optimized to have similar properties
like the human brain, which are: 1. Training It is able to
accumulate several inputs and create an overall solution. 2.
Utilization It uses the previous knowledge from an earlier
created solution to solve a fresh problem. Mathematically, a
neuron’s network function g(k) is defined as a composition
of other functions x(k), which can further be defined as a
composition of other functions. This can be conveniently
represented as a network structure, with arrows depicting the Fig 1. Block diagram of Proposed Algorithm
dependencies between variables.

(1) 4.1 Feature Extraction

where g is a predefined activation function, xi(k) is the input Region properties of the digitized signature images are
value in discrete time in k, wi (k) is weight value in discrete
time k, b is bias, g(k) is output value in discrete time k.
analyzed for 5 specific parameters namely, eccentricity, where p is the histogram count of the signature image.
convex area, standard deviation, entropy and orientation
[10]. A feature vector with 5 dimensions is constructed
corresponding to each image. Following are the region
properties of the image region in consideration. The image Orientation: It measures the angle (ranging from -90 to 90
degrees) between the x-axis and the major axis of the ellipse
region contains a specified pixel area which the scanned that has the same second-moments as the region. This value
signature occupies[11]. is converted to radians and used in our GUI.

Ecentricity: It is the ratio of the distance between the foci of

the ellipse and its major axis length. The value is between 0 4.2 Classification
and 1. An ellipse with an eccentricity 0 is a circle and the
one with its eccentricity 1 is a straight line.
The GUI launches on running the script in MATLAB
and waits for the user inputs. At the training stage, the user
(2) gives different images as input, of which the real input to the
network are the individual image pixels. When the input is
Convex Area: A Scalar that specifies the number of pixels accepted, it passes through the back propagation algorithm
in a Convex Image. It is a simple pixel count. and generates an output which contains the neural network
data of the trained images. The Back propagation artificial
Standard Deviation (ı): It measures the amount of variation neural network calculates the gradient of error
or dispersion on a set of mean data values. If deviation is corresponding to the networks modifiable weights. This was
closed to the mean data value then the variation is less chosen to implement due to its simplicity of implementation
otherwise spread over a wider range of values. and ability to maintain the efficiency of the network as
shown in Table 1. Backpropagation is an example of
supervised learning and is used to minimize the error
) between each layers response and

Table.1 Number of genuine/forgery signature recognized in each

database

the actual data. Neural network input-output transformation

equation is
Y = f(x,W) (6)

Here, W is the matrix of weight vectors and x represents the

input to the network.
Back propagation equations for training of hidden layers
are:
Entropy: This is a standard statistical measure of
randomness which is used to characterize the texture of the Nw = Nw + į1w (7)
input image.

E í™p ‫ כ‬logp (5) į1w = Nout ‫ כ‬íNout) ‫ כ‬Nerrorfactor (8)

Nerrorfactor = Nexpectedoutput í Nactualoutput (9)

 Table 2. FAR and FRR of each database

Fig 3 Signature verified as genuine

VI. CONCLUSION
The work showcased in this paper strictly concentrates on
offline digital signatures which are static in nature. GUIDE
function in the MATLAB toolbox was used to prepare the
script. The parameters to be tested were coded and the codes
were embedded into the script. The Artificial Neural
Network Technology employed for the desired pattern
matching task successfully verified all the input signatures
which were previously not present in the database. The task
of recognition and verification of every new signature of an
individual is achieved by matching its feature vector with
the images in the database. The reliability of this approach is
determined to increase even more if the neural network is
trained using a larger database. A systematic approach
resulted in verifying signatures with better efficiency as
Fig 4. Not genuine image detected shown in Table 3.

V. TRAINING AND TESTING

The Neural Network is constructed and trained using
these feature vectors and a Target matrix which teaches the
classifier of the authenticity of the signature image as shown
in Figure 5 and 6
Graphic User Interface (GUI): MATLAB R2011a version
has the GUIDE functionality which helps with the creation
of the GUI. The corresponding script is automatically
generated, which can be modified to add explicit functions
pertaining to each of the buttons and text boxes on the GUI
as shown in Figure 2. Images are verified as genuine or not
genuine and the result is displayed accordingly as shown in
Table 3 Shows accuracy rate of each saples
Figure 3 which recognized the signature as genuine while in
Figure 4 the signature is not genuine and FAR/FRR is
shown in Table 2.
 Fig. 5 Training neural network

Fig 6. Pattern Recognition (NN tool)

REFERENCES [6] D. Bertolini, L. S. Oliveira, E. Justino, and R. Sabourin,

[1] F. Leclerc and R. Plamondon, “Automatic signature verification: “Reducing forgeries in writer-independent off-line signature
The state of the art1989–1993,” International Journal of Pattern verification through ensemble of classifiers,” Pattern Recognition,
Recognition and Artificial Intelligence, vol. 8, no. 03, pp. 643–660, vol. 43, no. 1, pp. 387– 396, 2010.
1994. [7] A. Karouni, B. Daya, and S. Bahlak, “Offline signature
[2] D. Impedovo and G. Pirlo, “Automatic signature verification: recognition using neural networks approach,” Procedia Computer
the state of the art,” Systems, Man, and Cybernetics, Part C: Science, vol. 3, pp. 155–161, 2011.
Applications and Reviews, IEEE Transactions on, vol. 38, no. 5, pp. [8] H. H. Wai and S. L. Aung, “Feature extraction for offline
609–635, 2008. signature verification system,” IJCCER, vol. 1, no. 3, pp. 84–87,
[3] B. Herbst and H. Coetzer, “On an offline signature verification 2013.
system,” in 9th Annual South African Workshop on Pattern [9] K. Huang and H. Yan, “Off-line signature verification based on
Recognition, 1998, pp. 39–43. geometric feature extraction and neural network classification,”
[4] A. Pansare and S. Bhatia, “Offline signature verification using Pattern Recognition, vol. 30, no. 1, pp. 9–17, 1997.
neural network,” International Journal of Scientific & Engineering [10] R. Bajaj and S. Chaudhury, “Signature verification using
Research, Volume3, no. 2, 2012. multiple neural classifiers,” Pattern recognition, vol. 30, no. 1, pp. 17,
[5] D. Uppalapati, “Integration of offline and online signature 1997
verification Figure 6. Pattern Recognition of Neural network systems,
Department of Computer Science and Engineering, IIT, Kanpur,
2007.