Sie sind auf Seite 1von 8

International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 1

Volume 2, No. 1, January 2013

Signature Recognition & Verification System Using Back


Propagation Neural Network
Nilesh Y. Choudhary, GF’S GCOE, Jalgaon, India
Mrs. Rupal Patil, GF’S GCOE, Jalgaon, India
Dr. Umesh. Bhadade, GF’S GCOE, Jalgaon, India
Prof. Bhupendra M Chaudhari, Govt. Polytechnics Nadurbar, India

ABSTRACT pattern screening. It is easier for people to migrate from


using the popular pen-and-paper signature to one where
The fact that the signature is widely used as a means of the handwritten signature is captured and verified
personal identification tool for humans require that the electronically.
need for an automatic verification system. Verifwication
can be performed either Offline or Online based on the There are two main streams in the signature recognition
application. However human signatures can be handled task. First approach requires finding information and
as an image and recognized using computer vision and can recognize signature as the output of the system and
neural network techniques. With modern computers, it is seen that in a certain time interval, it is necessary to
there is need to develop fast algorithms for signature make the signature. This system models the signing
recognition. There are various approaches to signature person and other approach is to take a signature as a
recognition with a lot of scope of research. In this static two-dimensional image which does not contain
paper, off-line signature recognition & verification any time-related information [1].in short, signature
using back propagation neural network is proposed, recognition can be divided into two groups. Online and
where the signature is captured and presented to the offline.
user in an image format. Signatures are verified based
on features extracted from the signature using Invariant The online signature recognition, where signatures are
Central Moment and Modified Zernike moment for its acquired during the writing process with a special
invariant feature extraction because the signatures are instrument, such as pen tablet. In fact, there is always
Hampered by the large amount of variation in size, dynamic information available in case of online
translation and rotation and shearing parameter. signature recognition, such as velocity, acceleration and
Before extracting the features, preprocessing of a pen pressure. So far there have been many widely
scanned image is necessary to isolate the signature part employed methods developed for online signature
and to remove any spurious noise present. The system is recognition for example, Artificial Neural Networks
initially trained using a database of 56 persons (ANN)[2,3], dynamic time warping (DTW)[4,5], the
signatures obtained from those 56 individuals whose hidden Markov models (HMM)[6,7].
signatures have to be authenticated by the system. For
each subject a mean signature is obtained integrating The off-line recognition just deals with signature
the above features derived from a set of his/her genuine images acquired by a scanner or a digital camera. In
sample signatures .This signature recognition& general, offline signature recognition& verification is a
verification system is designed using MATLAB. This challenging problem. Unlike the on-line signature,
work has been tested and found suitable for its purpose. where dynamic aspects of the signing action are
captured directly as the handwriting trajectory, the
dynamic information contained in off-line signature is
INTRODUCTION
highly degraded. Handwriting features, such as the
handwriting order, writing-speed variation, and
Handwritten signature is one of the most widely skillfulness, need to be recovered from the grey-level
accepted personal attributes for identity verification of pixels.
the person. The written signature is regarded as the
primary means of identifying the signer of a written In the last few decades, many approaches have been
document based on the implicit assumption that a developed in the pattern recognition area, which
person’s normal signature changes slowly and is very approached the offline signature verification problem.
difficult to erase, alter or forge without detection. The Justino,[8] propose an off-line signature verification
handwritten signature is one of the ways to authorize system using Hidden Markov Model . Zhang, Fu and
transactions and authenticate the human identity Yan [9] proposed handwritten signature verification
compared with other electronic identification methods system based on Neural ‘Gas’ based Vector
such as fingerprints scanning and retinal vascular

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 2
Volume 2, No. 1, January 2013

Quantization. Vélez, Sánchez and Moreno [10] propose The input signature is captured from the scanner or
a robust off-line signature verification system using digital high pixel camera which provides the output
compression networks and positional cuttings. [11, 12, image in term of BMP Colour image. The
13] preprocessing algorithm provides the required data
suitable for the final processing. In the feature
The signature recognition & verification system shown extraction phase the invariant central moment and
in Fig 1 is broadly divided into three subparts Zernike moment are used to extract the feature for the
1) Preprocessing, 2) Feature extraction,3) Recognition classification purpose. In classification the Back
& Verification. propagation Neural Network is used to provide high
accuracy and less computational complexity in training
and testing phase of the system.

1. SIGNATURE DATABASE signed 4 forgery signatures in the training set the total
number of signatures is 675 (12 x 56) are used. In order
For training and testing of the signature recognition and to make the system robust, signers were asked to use as
verification system 675 signatures are used. The much as variation in their signature size and shape and
signatures were taken from 56 persons. The templates the signatures are collected at different times without
of the signature as shown in Fig 2 seeing other signatures they signed before.
For testing the system, another 112 genuine signatures
For training the system 56 person’s signatures are used. and 112 forgery signatures are taken from the same 56
Each of these persons signed 8 original signature and persons in the training set.

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 3
Volume 2, No. 1, January 2013

2.1 Converting Color Image to Gray Scale


Image
In today’s technology, almost all image capturing and
Scanning devices gives their output in color format. A
color image consists of a coordinate matrix and three
color matrices. Coordinate matrix contains X, Y
coordinate values of the image. The color matrices are
labeled as red (R), green (G), and blue (B). The
technique presented in this study are based on grey
scale images, therefore, scanned or captured color
images are initially converted to grey scale using the
following equation (1)

Gray color = 0.299*Red + 0.5876*Green


+0.114*Blue
(1)

Fig 3.Scanned Image

Fig 4.Colour to Gray Scale Image

2.2 Noise Reduction


Noise reduction (also called “smoothing” or “noise
filtering”) is one of the most important processes in
Fig 2.Signature Templates image processing. Images are often corrupted due to
positive and negative impulses stemming from
decoding errors or noisy channels. An image may also
2. PREPROCESSING be degraded because of the undesirable effects due to
illumination and other objects in the environment.
Preprocessing algorithms is nothing but data Median filter is widely used for smoothing and
conditioning algorithm which provide data for feature restoring images corrupted by noise. It is a non-linear
extraction process. It establishes the link between real process useful especially in reducing impulsive or salt-
world data and recognition & verification system. The and-pepper type noise. In a median filter, a window
preprocessing of the trajectory of input signature pattern slides over the image, and for each positioning of the
directly facilitates pattern description and affects the window, the median intensity of the pixels inside it
quality of description. Any image-processing determines the intensity of the pixel located in the
application suffers from noise like touching line middle of the window. Different from linear filters such
segments, isolated pixels and smeared images. This as the mean filter, median filter has attractive properties
noise may cause severe distortions in the digital image for suppressing impulse noise while preserving edges.
and hence result in ambiguous features and a Median Filter is used in this study due to its edge
correspondingly poor recognition and verification rate. preserving feature [14,15, 16, 17].
The preprocessing step is applied both in training and
testing phases. Background elimination, noise
reduction, width normalization and skeletonization are
the sub steps

Fig 5.Noise Removal

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 4
Volume 2, No. 1, January 2013

2.3 Background Elimination and Border In these equations:


Clearing , - pixel coordinates for the normalized signature,
In Many image processing algorithms require the , - pixel coordinates for the original signature,
separation of objects from the image background. M- one of the dimensions (width or height) for the
Thresholding is the most easily & sophistically normalized signature
applicable method for this purpose. It is widely used in
image segmentation [18, 19].

Thresholding is choosing a threshold value H and


assigning 0 to the pixels with values smaller than or
equal to H and 1 to those with values greater than H.
We used Thresholding technique for separating the
signature pixels from the background pixels. Clearly, in Fig 6. Normalized Image
this application, we are interested in dark objects on a
light background, and therefore, a threshold value H,
called the brightness threshold, is appropriately chosen 3. FEATURE EXTRACTION
and applied to image pixels f(x, y) as in the following
Equation (2) Feature extraction, as defined by Devijver and Kittle
[20] is “Extracting the information from the raw data
If f(x, y) ≥ H then which is most relevant for classification stage. This
f(x, y) = Background data can be minimized within-class pattern variation
else f(x, y) = Object and increases the inter-class variations.” Therefore,
(2) achieving a high recognition performance in signature
Signature image which is located by separating it from recognition system is highly influenced by the selection
complex background image is converted into binary of efficient feature extraction methods, taking into
image white background taking the pixel value of 1. consideration the domain of the application and the type
Vertical and horizontal (histogram) projections are used of classifier used [21]. An efficient feature extraction
for border clearing. For both direction, vertical and algorithm should require two characteristics: Invariance
horizontal, we counted every row zeros and the and reconstruct-ability Features [21] that are invariant
resulting histogram is plotted sideways. to certain transformations on the signature which would
be able to recognize many variations of these
2.4 Signature Normalization signatures. Such transformations include translation,
Signature dimensions may vary due to the irregularities scaling, rotation, stretching, skewing and mirroring.
in the image scanning and capturing process.
Furthermore, height and width of signatures vary from On the other hand, the ability to reconstruct signature
person to person and, sometimes, even the same person from their extracted features ensures that complete
may use different size signatures. First, we need to information about the signature shape is present in these
eliminate the size differences and obtain a standard features. In this feature extraction step, the well known
signature size for all signatures. After this normalization feature set in pattern recognition is used. one is depends
process, all signatures will have the same dimensions. on invariant central moment designed by Hu’s [22]
In this study, we used a normalized size of 50x50 pixels which is used for scale and translation normalization
for all signatures that will be processed further. During and other is modified Zernike moment[23] which is
the normalization process, the aspect ratio between used for rotation normalization.
width and height of a signature is kept intact.
3.1 Invariant Central Moment
Normalization process made use of the following The moments of order (u + v) of an image composed of
equation (3) & (4). binary pixels B(x, y) are proposed by [24], [25] as
shown in eq. (5).

(3)
(5)
The body’s area A and the image’s center of
mass i s found from eq. 6.
(4)
(6)

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 5
Volume 2, No. 1, January 2013

The central moments, which are translation Invariant,


are given by eq. 7.

(12)
Where
(7)
The orthogonality property of Zernike moments, as
expressed in the eq.8, allows easy image reconstruction
Finally, the normalized central moments, which are
from its Zernike moments by simply adding the
translation and scale invariant, are derived from the
information content of each individual order moment.
central moments as shown in eq. 8.
Moreover, Zernike moments have simple rotational
transformation properties interestingly enough the
Zernike moments of a rotated image, have identical
(8) magnitudes to those of the original one, where they
Where merely acquire a phase shift upon rotation.
K=1+ (u+ v)/2 for u+v≥2
Therefore, the magnitudes of the Zernike moments are
3.2 Zernike Moments rotation invariant features of the underlying image.
Zernike polynomials are a set of complex polynomials Translation and scale-invariance, on the other hand, are
which form a complete orthogonal set over the interior obtained by shifting and scaling the image into the unit
of the unit circle [26].The form of polynomial is shown circle.
by eq. 10.

(10)
Where
is the Length of the vector from the origin to the point
(x, y), θ is the angle between this vector and the x axis
in the Counterclockwise direction and the radial
polynomial is Fig 7. Rotation Normalization

4. BACK PROPAGATION ARTIFICIAL NEURAL


NETWORK
(10) There are several algorithms that can be used to create
Zernike moments are the projections of the image an artificial neural network, but the Back propagation
function onto these orthogonal basis functions. The [27] was chosen because it is probably the easiest to
Zernike moment of order n with repetition m for a implement, while preserving efficiency of the network.
digital image is given by Backward Propagation Artificial Neural Network
(ANN) use more than one input layers (usually 3). Each
of these layers must be either of the following:
• Input Layer – This layer holds the input for the
network
(11) • Output Layer – This layer holds the output data,
Where, * is the complex conjugate operator and usually an identifier for the input.
x2+y2≤1. • Hidden Layer – This layer comes between the input
layer and the output layer. They serve as a
To calculate the Zernike moments for a given image, its propagation point for sending data from the
pixels are mapped to the unit circle x2+y2≤1. This is previous layer to the next layer.
done by taking the geometrical center of the image as
the origin and then scaling its bounding rectangle into A typical Back Propagation ANN is as depicted in Fig 8
the unit circle, as shown in Figure 7. Due to the The black nodes (on the extreme left) are the initial
orthogonality of the Zernike basis, the part of the inputs. Training such a network involves two phases. In
original image inside the unit circle can be the first phase, the inputs are propagated forward to
approximated using its Zernike moments Anm up to a compute the outputs for each output node. Then, each
given order nmax using of these outputs is subtracted from its desired output,
causing an error [an error for each output node].

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 6
Volume 2, No. 1, January 2013

In the second phase, each of these output errors is


passed backward and the weights are fixed. These two
phases are continued until the sum of square of output
errors reaches an acceptable value. Each neuron is
composed of two units. The First unit adds products of
weights coefficients and input signals while the second
unit realizes nonlinear function, called neuron
activation function. Signal is adder output signal and
= is output signal of nonlinear element. Signal y is
also output signal of neuron. To teach the neural
network, we need data set. The training data set consists
of input signals 1 2 assigned with corresponding
target (desired output). The network training is an
iterative process. In each iteration weights coefficients
of nodes are modified using new data from training data
set. Each teaching step starts with forcing both input Fig 8. A 3-layer neural network using back propagation
signals from training set. After this stage we can
determine output signals values for each neuron in each When the application launches, it waits for the user to
network layer determine whether he wishes to train or verify a set of
signatures. At the training stage, based on the back
Symbols represent weights of connections propagation neural network algorithm, the user gives
between output of neuron and input of neuron in eight 12 different images as input, of which the real
the next layer. input to the network, are the individual pixels of the
images. When input is confirmed and accepted, it
In the next algorithm step, the output signal of the passes through the back propagation neural network
network is compared with the desired output value algorithm to generate an output which contains the
(the target), which is found in training data set. The network data of the trained images. The back
difference is called error signal of output layer propagation artificial neural network simply calculates
neuron. It is impossible to compute error signal for the gradient of error of the network regarding the
internal neurons directly, because output values of these networks modifiable weights. In this paper we a multi-
neurons are unknown. For many years the effective layer neural network designed by O.C Abikoye [28],
method for training multilayer networks has been
unknown. Only in the middle eighties the back 5. TRAINING AND TESTING
propagation algorithm has been worked out. The idea is
to propagate error signal (computed in single teaching
The recognition phase consists of two parts, training
step) back to all neurons, which output signals were
and testing respectively which is accomplished by back
input for discussed neuron.
propagation neural network.
The weights' coefficient used to propagate errors
As explained in Section 1. 672 images in our database
back are equal to this used during computing output
belonging to 56 people are used for both training and
value. Only the direction of data flow is changed
testing. Since 8 (out of 12) input vectors for each image
(signals are propagated from output to inputs one after
were used for training purposes, there are only 224
the other). This technique is used for all network layers.
(56*4) input vectors (data sets) left to be used for the
If propagated errors came from few neurons they are
test set. Under normal (correct) operation of the back
added. The illustration is below
propagation neural network, only one output is expected
to take a value of “1” indicating the recognition of a
signature represented by that particular output. The
other output values must remain zero. The output layer
used a logic decoder which mapped neuron outputs
between 0.5-1 to a binary value of 1. If the real value of
an output is less than 0.5, it is represented by a “0”
value. The back propagation neural network program
recognized all of the 56 signatures correctly. This result
translates into a 100% recognition rate. We also tested
the system with 15 random signatures which are not
contained in the original database.

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 7
Volume 2, No. 1, January 2013

Only two of these signatures which are very similar to percentage error with the dimension of the feature
at least one of the 56 stored images resulted in “false vector is calculated.
positives” (output > 0.5) while the remaining 8 are
recognized correctly as not belonging to the original set The recognition system gives the 98% success rate by
(the output value was <= 0.5). Since recognition step is recognizing the all signature pattern correctly for all
always followed by the verification step, these kinds of that signature which is used in training. It gives the poor
false positives can be easily caught by our verification performance for signature that is not in the training
system. In other words, the verification step serves as a phase. Generally the failure to recognize/verify a
safeguard against “false positives” as well as “false signature was due to poor image quality and high
negatives”. similarity between 2 signatures. Recognition and
verification ability of the system can be increased by
6. RESULT AND CONCLUSION using additional features in the input data set. This
study aims to reduce to a minimum the cases of forgery
In this study, we presented Off-Line Signature in business transactions.
Recognition and Verification System using the back
propagation neural network which is based on steps of REFERENCES
image processing, invariant central moment invariants,
Zernike moment & some global properties and back [1] A. Pacut, A Czajka, “Recognition of Human
propagation neural networks. Signatures”, pp. 1560-1564, 2001.
[2] Ronny Martens, Luc Claesen, “On- Line
Both systems used a three-step process; in the first step, Signature Verification by Dynamic Time-
the signature is separated from its image background. Warping”, IEEE Proceedings of ICPR'96
Second step performs normalization and digitization of 1996.
the original signature. Invariant central moment [3] Quen-Zong Wu, I-Chang Jou, and Suh-Yin
invariants, Zernike moment and global properties which Lee, “On-Line Signature Verification Using
are used as input features for the back propagation LPC Cepstrumand Neural Networks”, IEEE
neural network are obtained in the third step. Transactions on Systems, Man, and
Cybernetics—Part B: Cybernetics, 27(1):148-
153, 1997.
[4] Pave1 Mautner, OndrejRohlik, Vaclav
Matousek, JuergenKempp, “Signature
Verification Using ART-2 Neural Network”,
Proceedings of the 9th International
Conference on Neural information Processing
(ICONIP'OZ) ,2: 636-639, 2002
[5] A. Jain, F. Griess, S. Connell, “On-line
signature Verification”, Pattern Recognition,
Vol. 35, No. 12, 2002
[6] W. Nelson, W. Turin, T. Hastie, “Statistical
methods for on-line signature verification”,
International Journal of Pattern
Recognition and Artificial Intellingence, 8,
1994
Fig. 9 Online Result for Proposed Method [7] R. Kashi, J. Hu, W.L. Nelson, W.Turin, “A
hidden markov model approch to online
As Shown in Fig 9 the proposed system is evaluated on handwritten signature verification”,
two performance criteria. The feature extraction stage International Journal on Document Analysis
and the overall recognition rate for achieving high and Recognition, Vol. 1, No.1, 1998.
recognition performance in signature recognition [8] E. J. R. Justino, F. Bortolozzi and R.
system is highly influenced by the selection of efficient Sabourin,( 2001) “Offline Signature
feature vector. In this paper, we compute the Verification Using HMM for Random, Simple
comparison between the previously implemented Hu‘s and Skilled Forgeries”, ICDAR 2001,
moment [28] Zernike moment [23]. Original feature International Conference on Document
vectors produced from the different moment invariant Analysis and Recognition, vol. 1, pp. 105--
techniques are applied for signature features extraction 110.
from the binary images of the signature and absolute [9] B. Zhang, M. Fu and H. Yan (1998 ),
“Handwritten Signature Verification based on

i-Xplore International Research Journal Consortium www.irjcjournals.org


International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) ISSN: 2319-4413 8
Volume 2, No. 1, January 2013

Neural ‘Gas’ Based Vector Quantization”, [18] Erdem, U.M., “2D Object Recognition In
IEEE International Joint Conference on Neural Manufacturing Environment Using Implicit
Net-works, pp. 1862-186 Polynomials and Algebraic Invariants”, Master
[10] J. F. Vélez, Á. Sánchez, and A. B. Moreno ( Thesis, Bogazici University, 1997.
2003 ) , “Robust Off-Line Signature [19] Fu, K.S., Mui, J.K., “A survey On Image
Verification Using Compression Networks Segmentation”, Pattern Recognition, Vol. 13,
And Positional Cuttings”, Proc. 2003 IEEE pp.3-16, Pergoman Press, 1981.
Workshop on Neural Networks for Signal [20] Devijver, P.A. and J. Kittler, 1982, Pattern
Processing, vol. 1, pp. 627-636. Recognition: A Statistical Approach. Prentice-
[11] Q. Yingyong, B. R. Hunt, "Signature Hall, London, ISBN: 10: 0136542360.
Verification Using Global and Grid Features", [21] Trier, O.D., A.K. Jain and T. Taxt, Feature
Pattern Recognition, vol. 22, no.12, Great extraction methods for character recognition-a
Britain (1994), 1621--1629. survey. Patt.Recog., 29: 641-662.
[12] Drouhard, J.P., R. Sabourin, and M. Godbout, [22] Hu M, “Visual pattern recognition by moment
“A neural network approach to off-line invariants”, IRE Trans. Inf. Theory, IT-8, PP:
signature verification using directional 179–187.
PDF”, Pattern Recognition, vol. 29, no. 3, [23] Khotanzad, A. and Y.H. Hong, “Invariant
(1996), 415-424. image recognition by Zernike moments”, IEEE
[13] G. Rigoll, A. Kosmala, "A Systematic Trans. Patt. Anal. Mach. Intell., 12: 489-
Comparison Between On-Line and Off-Line 497. DOI: 10.1109/34.55109.
Methods for Signature Verification with [24] Theodoridis, S. and K. Koutroumbas, 2006,
Hidden Markov Models", 14th International “Pattern Recognition”, 3rd Edn., Academic
Conference on Pattern Recognition - vol. Press, ISBN: 10: 0123695317, pp: 856.
II, Australia (1998), 1755. [25] Reiss, T.H, “The revised fundamental theorem
[14] Lim, J.S., “Two-Dimensional and Image of moment invariants”, IEEE Trans. Patt.Anal.
Processing”, Prentice-Hall, 1990. Mach. Intell., 13: 830-834. DOI:
[15] Yang, X., and Toh, P.S., “Adaptive Fuzzy 10.1109/34.85675, 1991.
Multilevel Median Filter”, IEEE Transaction [26] Khotanzad, A, Y. H. Hong, “Invariant image
on Image Processing, Vol. 4, No. 5, pp.680- recognition by Zernike moments”, IEEE Trans.
682, may 1995. Patt. Anal.Intell, pp489-497, March 1990.
[16] Hwang, H., and Haddad, R.A. “Adaptive [27] Golda, A. 2005. Principles of Training multi-
Median Filters: new Algorithm and Results”, layer neural network using back propagation.
Transactions on Image processing, Vol. 4, No. [28] O.C Abikoye, M.A MabayojeR. Ajibade
4 pp.449-505, April 1995. “Offline Signature Recognition & Verification
[17] Rosenfeld, A., “Digital Picture Processing”, using Neural Network” International Journal of
Academic Press Inc., 1982. Computer Applications (0975 – 8887) Volume
35– No.2, December 2011

i-Xplore International Research Journal Consortium www.irjcjournals.org