Off-line Handwritten Signature Recognition Using Wavelet Neural Network

Document Sample
Off-line Handwritten Signature Recognition Using Wavelet Neural Network Powered By Docstoc
					                                                       (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                 Vol. 8, No. 6, 2010



        Off-line Handwritten Signature Recognition
              Using Wavelet Neural Network
           Mayada Tarek1                               Taher Hamza                                  Elsayed Radwan
   Computer Science Department,               Computer Science Department,                 Computer Science Department,
Faculty of Computers and Information       Faculty of Computers and Information         Faculty of Computers and Information
             Sciences,                                  Sciences,                                    Sciences,
           Mansoura, Egypt                            Mansoura, Egypt                              Mansoura, Egypt



Abstract‫ــــ‬Automatic signature verification is a well-             in conjunction with a simple Euclidean distance
established and an active area for research with numerous           classifier; proposing a system for off-line signature
applications such as bank check verification, ATM access,           verification consists of four subsystems based on
etc. Most off-Line signature verification systems depend            geometric features, moment representations, envelope
on pixels intensity in feature extraction process which is          characteristics and wavelet features; applying wavelet
sensitive to noise and any scale or rotation process on             on signature verification [2,3,4,5].
signature image. This paper proposes an off-line
handwritten signature recognition system using Discrete             Although these methods achieved a good results, they
Wavelet Transform as feature extraction technique to                still suffer from the exchangeability of signature
extract wavelet energy values from signature image                  rotation and the distinguish-ability of person signature
without any dependency of image pixels intensity. Since             size. Most of these feature extraction methods depend
Discrete Wavelet Transform suffers from down-sample                 on signature shape or pixels intensity in specific region
process, Wavelet Neural Network is used as a classifier to          of signature. However, pixels' intensity are sensitive to
solve this problem. A comparative study will be illustrated         noise and also the signature shape may vary according
between the proposed combination system and pervious                to translation, rotation and scale variations of signature
off-line handwritten signature recognition systems.                 image [6].
Conclusions will be appeared and future work is proposed.
                                                                    Two types of feature can be extracted from signature
                                                                    image; first, global features which are extracted from
  Keywords-Discrete Wavelet Transform (DWT); Wavelet                the whole signature, including block codes [7]; second,
Energy; Wavelet Neural Network (WNN); Off-line                      local features which are calculated to describe the
Handwritten Signature.
                                                                    geometrical and topological characteristics of local
                                                                    segments [8]. Because of the absence of dynamic
                I.    INTRODUCTION                                  information in offline verification system, global
                                                                    features extraction are most appropriate [9]. One of the
In the field of personal identification, two types of               most appropriate global features extraction techniques is
                                                                    wavelet transform, since it extracts time-frequency
biometrics means can be considered; first, physiological
                                                                    wavelet coefficients from the signature image [8].
biometrics, which involves data derived from the direct             Wavelet Transform is especially suitable for processing
measurement of some part of the human body; for-                    an off-line signature image where most details could be
example fingerprint-, face-, palm print-, retina-based              hardly represented by functions, but could be matched
verification. Second, behavioural biometrics, which                 by the various versions of the mother wavelet with
involves data derived from an action taken by a person,             various translations and dilations [10]. Also, wavelet
or indirectly measures characteristics of the human                 transform is invariant to translation, rotation and scale
                                                                    of the image. Because of the advantage of wavelet
body; for-example: speech-, keystroke dynamics and                  transform, this paper uses it in feature extraction stage.
signature-based verification [1].
                                                                    Since one of problems that face wavelet is the huge size
In the last few decades, researchers have made great                of its coefficients, statistical model can be introduced to
efforts on off-line signature verification [1] for-                 represent them. This paper uses wavelet energy as
example; using the statistics of high grey-level pixels to          statistical model to represent all wavelet coefficients in
identify pseudo-dynamic characteristics of signatures;              efficient way. Another problem is down-sample process
developing technique based on global and grid features              which can lose some important extracted feature from
                                                                    signature image[11]. This paper proposes a Wavelet
   1                                                                Neural Network (WNN) technique for off-line signature
    Corresponding Author
   Mail: mayaatarek@yahoo.com
                                                                    recognition to overcome the disadvantages of Discrete
   Tel : 020108631688                                               Wavelet Transform (DWT) down-sample process.

                                                                                                                                     1

                                                              13                              http://sites.google.com/site/ijcsis/
                                                                                              ISSN 1947-5500
                                                     (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                               Vol. 8, No. 6, 2010

WNN takes full advantages of the partial-resolution               image. Dynamic systems use on-line acquisition devices
characteristic of the wavelet transform and the nonlinear         that generate electronic signals representative of the
mapping behaviour of Artificial Neural Networks                   signature during the writing process [1].
(ANN) [15].
                                                                  It is well known that no two genuine signatures of a
This paper proposes a combination model between                   person are precisely the same and some signature
DWT and WNN techniques for off-line handwritten                   experts note that if two signatures written on paper were
signature recognition system. DWT technique will                  same, then they could be considered as forgery by
analysis signature image to extract wavelet detail                tracing .Unfortunately, off-line signature verification is
coefficients. To reduce the huge number of these                  a difficult discrimination problem because of dynamic
coefficients with the same accuracy, a statistical model          information regarding the signing velocity, pressure and
is represented by wavelet energy. Because of the                  stroke order are not available also an off-line
problem of down sample, WNN technique will be used                handwritten signature is depend for instance on , the
as a suitable classifier technique to overcome this               angle at which people sign may be different due to
problem. Also, a modified back-propagation technique              seating position or due to support taken by hand on the
is used in learning WNN. A testing stage examines the             writing surface and all this information can’t be extract
unseen signature. Moreover, a comparative study will              from static image[12].
be illustrated between the proposed combination system
and pervious off-line handwritten signature recognition
systems. Conclusions will be appeared and future work             B. Wavelet Transform :
is suggested.
                                                                  Wavelet Transform (WT) [13] is become a powerful
The rest of this paper organized as; in Section 2,                alternative analysis tool to Fourier methods in many
Handwritten signature, wavelet transform (WT),                    signal processing applications. The main advantages of
Wavelet Neural Network (WNN) are mentioned.                       wavelets is that they have a varying window size, being
Methodology and applications using a combination                  wide for slow frequencies and narrow for the fast ones,
between DWT and WNN techniques is described in                    thus leading to an optimal time-frequency resolution in
Section 3. Section4, consists of the result of the                all the frequency ranges. Furthermore, owing to the fact
proposed combination system and a comparative study               that windows are adapted to the transients of each scale,
between three strategies (signature image pixels                  wavelets lack the requirement of stationary. There are
intensity value as input to ANN , signature wavelet               two types of Wavelet Transform; Continous Wavelet
energy values as input to ANN and signature wavelet               Transform(CWT), Discrete Wavelet Transform (DWT).
energy values as input to WNN). Finally section 5
concludes the paper.                                              The Continuous Wavelet Transform [14] of a 1-D signal
                                                                  x(t) is defined as in equation (1):
                                                                                          1                    t−b
                II. PRELIMINARIES                                         ����(a,b) (t)=              ����(����) ���� (       )   dt               (1)
                                                                                         √|����| ����                 a

A. Handwritten Signature                                          Where ψ(t) is the mother wavelet or the basis function
                                                                  which, in a form analogous to sins and cosines in
Handwritten signatures are widely accepted as a means             Fourier analysis. All the wavelet functions used in the
of document authentication, authorization and personal            transformation are derived from the mother wavelet
verification. For legality most documents like bank               through translation (shifting) b and scaling (dilation or
cheques, travel passports and academic certificates need          compression) a.
to have authorized handwritten signatures. In modern
society where fraud is rampant, there is the need for an          The Discrete Wavelet Transform [14], which is based
automatic Handwritten Signature Verification system               on sub-band coding is found to yield a fast computation
(HSV) [6]. Dependency on automation is due to the                 of wavelet transform. It is easy to implement and
difficulty faced in visual assessment for different types         reduces the computation time and resources required.
and different sizes of signatures. Simple, cursive,
graphical and not a connected curve pattern are some of           In CWT, the signals are analyzed using a set of basis
the different types of signatures and machines are far            functions which relate to each other by simple scaling
superior when it comes to processing speed and                    and translation. In the case of DWT, a time-scale
management of large data sets with consistency [12].              representation of the digital signal is obtained using
                                                                  digital filtering techniques. The signal to be analyzed is
                                                                  passed through filters with different cut off frequencies
Automatic HSV systems are classified into two types:              at different scales[14].
offline HSV and online HSV: static or off-line system
and dynamic or on-line system .Static off-line system             In DWT, the extension to 2-D is usually performed by
gain data after writing process has been completed .In            using a product of 1-D filters. The transform is
                                                                  computed by applying a filter bank as shown in
this case the signature is represented as a grey level
                                                                  Figure 1. L and H to denote the 1-D low pass and high
                                                                                                                                             2

                                                            14                                      http://sites.google.com/site/ijcsis/
                                                                                                    ISSN 1947-5500
                                                              (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                        Vol. 8, No. 6, 2010

pass filter, respectively. The rows and columns of image                   neuron parameters .The output of WNN is therefore a
are processed separately and down sampled by a factor                      linear combination of several multidimensional
of 2 in each direction which may cause losing important                    wavelets [15].
feature. Resulting in one low pass image LL and three
detail images HL, LH, and HH. Figure 2a shows the
one-level decomposition of Figure 1in the spatial
domain. The LH channel contains image information of
low horizontal frequency and high vertical frequency,
the HL channel contains high horizontal frequency and
low vertical frequency, and the HH channel contains
high horizontal and high vertical frequencies. Three-
level frequency decomposition is shown in Figure 2b.
Note that in multi-scale wavelet decomposition only the
LL sub-band is successively decomposed [13].



                                                                                    Figure 3 : The structure of the Wavelet Neural Network


                                                                           In this WNN model, the hidden neurons have wavelet
                                                                           activation functions ψ and have two parameter at,bt
                                                                           which represent dilation and translation parameter of
                                                                           wavelet function and V is the weight connecting the
                                                                           input layer and hidden layer and U is the weight
                                                                           connecting the hidden layer and output layer.
        Figure 1: A one-level wavelet analysis filter bank.
                                                                           Let Xn ={ xi },i=1,......,L and n=1,......N be the WNN
                                                                           input to no. n sample ; Yn ={ yk },k=1,......,S represents
                                                                           the output of WNN ; D={ dk },k=1,......,S represents the
                                                                           expected output ; Vij represents the connection weight
                                                                           between no. i node (input layer) and . j node (hidden
                                                                           layer) ; Ujk represents the connection weight between
                                                                           no. j node (hidden layer) and k node (output layer) .
                                                                           Where N is the number of Sample ; S is the number of
                                                                           output node ; L is the number of input node ; M is the
                                                                           number of hidden layer.
          Figure 2 : Wavelet frequency decomposition.
                                                                                 III.    WAVELET NEURAL NETWORK FOR
                                                                                        OFF-LINE HANDWRITTEN SIGNATURE
C. Wavelet Neural Network :                                                                       RECOGNITION

WNN is a combination technique between neural                              According to the fact that there aren’t two genuine
network and wavelet decomposition .The advantages of                       signatures of one person are precisely the same, many
the WNN are a high-speed learning and a good                               efforts have been done in order to comprehend the
convergence to the global minimum [15].The reason for                      delicate nuances of person signatures [12]. Especially
the application of WNN in case of such a problem as                        off-line signature recognition needs more effort because
classification is that the feature extraction and                          of the absence of dynamic information that can’t be
representation properties of the wavelet transform are                     extracted from static image [12]. Also, the problems of
merged into the structure of the ANN to further extend                     translation, rotation and scale variation of signature
the ability to approximate complicated patterns [16].                      image are still found when dealing with signature image
                                                                           pixels’ intensity [6].
The WNN can be considered an expanded perceptron
[17]. The WNN is designed as a three-layer structure                       This paper presents an implementation for off-line
with an input layer, a wavelet layer, and an output layer.                 handwritten signature recognition system using DWT
The topological structure of the WNN is illustrated in                     technique in feature extraction phase and WNN in
Figure 3.                                                                  classification phase to overcome all the above problems
                                                                           with off-line handwritten signature recognition system.
In WNN, both the position and dilation of the wavelets                     DWT technique depends on analyzing all signature
as well as the weights are optimized. The basic neuron                     shapes (continuous case) instead of analyzing the pixels
of a WNN is a multidimensional wavelet in which the                        intensity or segmentation part of signature (discrete
dilation and translation coefficients are considered as                    case). Because of the problem of down-sample caused
                                                                                                                                             3

                                                                     15                               http://sites.google.com/site/ijcsis/
                                                                                                      ISSN 1947-5500
                                                         (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                   Vol. 8, No. 6, 2010

by DWT technique, WNN technique will be used in                          According to the fact that there aren’t two genuine
classification stage to overcome this problem.                           signatures of one person are precisely the same, the
                                                                         differences in the same person signature may exist in
The proposed Off-line Handwritten Signature                              details. Because of the details of an image will access
recognition system as depicted in Figure 4 involves                      by high pass filter, DWT is used to access high pass
four stages:                                                             information of person’s signature images. This
                                                                         information is fused to obtain pattern of each person’s
        Scan and removing noise stage.                                  signatures that contains all details information of his/her
        Feature extraction stage.                                       signatures [21]. Details information extracted by DWT
        Classification stage.                                           technique must be extracted using suitable wavelet
        Test stage.                                                     function to off-line handwritten signature recognition
                                                                         application. According to the previous work in off-line
First stage, Scan and removing noise stage, each off-                    handwritten signature recognition have apply
line handwritten signature is scanned due to creating                    Daubechies 4, 12 and 20 wavelets functions as depicted
signature image. Because of the scanning process,                        in Figure 5 [5] as a mother wavelet function, which can
removing noise from signature image is an important                      preserve maximum details of the original image, reflect
task. In this paper, the median filter [18] is used to                   outline of the image objectively and decrease the FRR.
remove noise for two reasons. First, it preserves the
structural shape of the signature without removing small
strokes. Second, the absence of dealing with median
filter in wavelet transform technique, which work to
analysis image with low/high-pass filters corresponding
to its wavelet function.
                                                                                   db 4                db 12               db 20
The median filter is a nonlinear digital filtering
technique which is often used to remove noise. Noise                            Figure 5: Daubechies 4, 12 and 20 wavelets functions
reduction is a typical pre-processing step that improves
the results. The median filter considers each pixel in the               After DWT is applied on the image, wavelet
image in turn and looks at its nearby neighbours to                      coefficients from the approximation sub-band is discard
decide whether or not the pixel intensity value is                       and interested in wavelet coefficients from the details
representative of its surroundings. The median filter                    sub-bands of all the decomposition levels . This entire
replaces the pixel with the median of its neighbouring                   coefficient is very large to be used as feature extraction
pixel intensity values. The median is calculated by first                model from an image. These wavelet coefficients can be
sorting all the pixel intensity values from the                          represented as statistical features such as mean, median,
surrounding neighbourhood into numerical order and                       standard deviation, energy and entropy [22]. In this
then replacing the pixel being considered with the                       paper, wavelet energy values for details wavelet sub-
middle pixel intensity value [19].                                       band is the reduced vector that contain the main
                                                                         information that represent person signature from the
Second stage, Feature extraction stage is the most                       huge wavelet decomposition values.
important component for designing the intelligent
system based on pattern recognition. The pattern space                   While off-line handwritten signature image is sensitive
is usually of high dimensionality. The objective of the                  to translation, rotation and scale changes; the same
feature extraction is to characterize the object by                      images with different scale or rotational may have
reducing the dimensionality of the measurement space                     different wavelet coefficients. The main reason is that
(i.e., the original waveform). The best classifier will                  the efficient implementation of 2D-DWT requires
perform poorly if the features are not chosen well [20].                 applying a filter bank along the rows and columns of an
                                                                         image [23].




                               Feature
                           extraction using             Wavelet energy
                               Wavelet
                              Transform                    coefficient




                                     Figure 4: Proposed off-line Handwritten signature Recognition System
                                                                                                                                         4

                                                                 16                               http://sites.google.com/site/ijcsis/
                                                                                                  ISSN 1947-5500
                                                                                                   (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                                             Vol. 8, No. 6, 2010

Due to the separability of the filters, the separable 2D-                                                       The back-propagation algorithm seems to be superior in
DWT is strongly oriented in the horizontal and vertical                                                         this handwritten signature verification environment
directions. This makes it hardly possible to extract                                                            [25]. In a back-propagation neural network[26], the
translation, rotation and scale invariant features from                                                         learning algorithm has two phases. First, a training input
the wavelet coefficients. Wavelet energy can keep the                                                           pattern is presented to WNN input layer. The WNN
main characteristic of these wavelet coefficients and                                                           propagates the input pattern from layer to layer until the
make the same images with different translation,                                                                output pattern is generated by the output layer. If this
rotation and scale having the same wavelet energy                                                               pattern is different from the desired output, an error is
values[23]. Wavelet energy values can be computed                                                               calculated and then propagated backward through the
after analysis signature image to it’s wavelet sub-image                                                        WNN from the output layer to the input layer. The
coefficient at three level analysis (LLx, HLx, LHx,                                                             weights and both the position and dilation of the
HHx). The percentages of energy of these high                                                                   wavelets layer are modified as the error is propagated.
frequency sub-images at the k-level wavelet                                                                     The modified back-propagation training algorithm in
decomposition is defined in equation (2,3,4)[24]:                                                               WNN [27]as shown in Figure 6.

                                                                                                                In this work, The input layer represents wavelet energy
       (����)
                  100 ∗   (�������� ���������������������������������������������������� ������������������������ �������� �������������������� ����)2                    values feature vector to neural network. The output
 ������������       =                                                                              (2)
                               (���������������������������������������������������� ������������������������)2                                       layer represents the ability to recognize the human
                                                                                                                signature. The middle layer determined the ability to
                  100 ∗   (�������� ���������������������������������������������������� ������������������������ �������� �������������������� ����)2
  ������������ (����) =                                                                                   (3)           learn the person signature recognition. Because of the
                               (���������������������������������������������������� ������������������������)2
                                                                                                                ability of Morlet function to deal with big input domain
                  100 ∗     �������� ���������������������������������������������������� ������������������������ �������� �������������������� ����   2                 [28] and represents its wave form in equation, Morlet
 ������������ ���� =                                                                                      (4)           function will be the suitable wavelet activation
                                 ���������������������������������������������������� ������������������������ 2
                                                                                                                functions ψ in WNN to recognize offline handwritten
                                                                                                                signature application. Morlet function equation and it’s
Third stage, Classification stage, after we get the                                                             derivation in equation (5,6)[27]:
suitable wavelet energy values that represent signature
image, we take this values as input to WNN and train                                                                                                          ���� 2
                                                                                                                                 ���� ���� = cos 1.75���� exp −                          (5)
this network with a modified Back-propagation (BP)                                                                                                             2
training algorithm to get efficient off-line signature
recognition. Using WNN for two reasons; first,                                                                  Then.
traditional ANN has many trade-off because of complex
                                                                                                                      �������� ����                                             ���� 2
computations, huge iterations and learning algorithms                                                                         = − ���������������� 1.75���� + 1.75 sin 1.75���� exp −          (6)
                                                                                                                         ��������                                              2
are responsible for slowing down the recognition rate
using ANN; second ,recover losing important
information from signature image in DWT technique
because of down-sample process as depicted in
Figure1.




                                                                                                                                                                                    5

                                                                                                          17                                http://sites.google.com/site/ijcsis/
                                                                                                                                            ISSN 1947-5500
                                                                                  (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                            Vol. 8, No. 6, 2010

Input : wavelet energy values extracted from signature image
Output : Class of recognized Signature

 Step1: Initialize weights and offsets.
              Set all weights and node offsets to small random values.
        Initialize position and dilation parameter for each wavelet neuron in wavelet layer.
              To choose centre point (p)between interval [z 1,z2] (input domain),then
                                   b1=p , a1=0.5(z1-z2)
              Interval [z1,z2] is divided into two parts by point p.
              In each sub-interval, we recursively repeat the same procedure which will
                       initialize b2, a2 and b3, a3 and so on, until all the wavelet are initialize.

 Step2: Present input and desired outputs
          Present a continuous valued input vector X1, X2…..XL
           and specify the desired output D1,D2,….DS.

              If the net is used as a classifier then all desired outputs are typically set to zero except for that corresponding
             to the class the input is from. That desired output is1. The input could be new on each trial or samples from a
             training set could be presented cyclically until stabilize.

 Step 3: Calculate Actual Output
                                                                  ����                               ����             ����
                                                                                                   ����=1 ������������ ��������    − ��������
                                                      ������������ =           ������������ ����                                                                                                                                   7
                                                                                                            ��������
                                                                 ���� =1


 Step 4: Calculate Error function
                                                                                 ����          ����
                                                                    1
                                                          ���� =                                                 ����
                                                                                                   (������������ − �������� )2                                                                                                 8
                                                                   2����
                                                                             ����=1 ����=1
 Step 5: Propagate error to weights and position and dilation parameter
                                                                             ����                                           ����             ����
                                                        ��������     1                                                        ����=1 ������������ ��������    − ��������
                                                               =                                   ����
                                                                                       (������������ − �������� ) ����                                                                                                           9
                                                       ���������������� ����                                                                 ��������
                                                                            ����=1

                                                                            ����          ����
                                                       ��������     1                                                               ��������(����) ������������
                                                              =                                            ����
                                                                                              [(������������ − �������� ) ������������                          ]                                                                    10
                                                      ���������������� ����                                                                  �������� ��������
                                                                         ����=1 ����=1


              ����              ����
              ����=1 ���� �������� �������� −���� ����
Where ���� =
                      ���� ����

                                                                       ����         ����                                                             ����             ����
                                                     ��������    1                                                            ��������(����)               ����=1 ������������ ��������    − ��������
                                                           =                                        ����
                                                                                        (������������ − �������� ) ������������                    −                         2                                                       11
                                                     ������������ ����                                                               ��������                        ��������
                                                                    ����=1 ����=1




                                                                             ����         ����
                                                        ��������    1                                                               �������� ����              1
                                                              =                                               ����
                                                                                                  (������������ − �������� ) ������������                       −                                                                    12
                                                        ������������ ����                                                                  ��������             ��������
                                                                            ����=1 ����=1




    Step 6: Update weights and position and dilation parameter
                                                                                                        ��������
                                                             ��������        ��������
                                                          ������������ +1 = ������������ − ����                                       ��������     ��������
                                                                                                                − ���� ������������ − ������������ −1                                                                              13
                                                                                                       ����������������
  where :α is learning rate
        ���� is momentum factor
 Step 7: Repeat by going to step 2

                                         Figure 6: Back-propagation training algorithm in Wavelet Neural Network




                                                                                                                                                                                                                          6

                                                                                                        18                                                                    http://sites.google.com/site/ijcsis/
                                                                                                                                                                              ISSN 1947-5500
                                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                           Vol. 8, No. 6, 2010




                                    Feature
                                extraction using              Wavelet energy
                                    Wavelet
                                   Transform                    coefficient

Second Person Signature




                                         Figure 7: Testing stage of off-line Handwritten signature Recognition System


Finally, Test stage, after learning WNN, we can                                 Three WNN will be found to compare the recognition
examine the ability of WNN to verify the signature of                           rate between tree wavelet function. Modified BP
any person as shown in Figure 7.In this stage our goal                          training algorithm as in Figure 6 is used to train WNN.
is to input signature image and recognize the person                            Finally, testing WNN with trained signature . Figure 9
signature. After scanning and removing noise from                               shows the recognition rate to (db4,db12,db20) wavelet
person signature image, wavelet coefficients produce                            detail coefficients using WNN as mention above. As a
after analysis image with DWT technique and then                                result from Figure 9, Db20 is recognizing to be the
compute wavelet energy value from wavelet detail                                suitable wavelet function which have high recognition
coefficients, finally, this wavelet energy values are                           rate in our database to offline handwritten application.
taken as input to test WNN classifier to find result in
output layer with only 1 value in only one neuron.
Number of neurons in the output layer represents the
number of person that system recognize.


                          IV.   RESLUT :

This section summarizes the results of using DWT
technique (wavelet energy values) as feature extraction
technique and WNN as classifier to off-line handwritten
signature recognition system. This paper uses nine
person handwritten signatures as show in Figure 8,
each person has twenty image of his handwritten                                     Figure 9: Offline handwritten signature recognition rate using
                                                                                             (db4,db12,db20) wavelet detail coefficients
signature ,ten for train stage and ten for test stage .
                                                                                After determine the suitable extracting wavelet
                                                                                function, wavelet energy from each signature image is
                                                                                computed using equation 2,3,4 with db20 as wavelet
                                                                                function at three level analysis. Nine wavelet energy
                                                                                coefficients are represented each signature image.


                                                                                        Table 1: WNN architecture and training parameters
                                                                                     The number of layers                         3
                                                                                  The number of neuron on the                    Input:9
                                                                                           layers                               Hidden:18
                                                                                                                                Output:9
              Figure 8: Sample Signature images
                                                                                 The initial weights and biases                  Random

In feature extraction stage, wavelet detail coefficients                          Wavelet Activation functions               Morlet function
are extracted from signature image using (db4 or db 12                                   Learning rule                      Back-Propagation
or db20) wavelet function. To determine the suitable                                         MSE                                 0.0001
wavelet function to our database, WNN is used as a                                       Learning rate                             0.1
classifier to evaluate the suitable one. Wavelet detail
coefficients (at one level analysis) of signature image                                Momentum factor                            0.009
according to one wavelet function is taken as trained
data to WNN.


                                                                                                                                                     7

                                                                         19                                 http://sites.google.com/site/ijcsis/
                                                                                                            ISSN 1947-5500
                                                               (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                         Vol. 8, No. 6, 2010

In classification stage, WNN is used with parameters
shown in Table1. These parameters are selected for
WNN structure after several different experiments. In
these experiments, the WNN is employed with different
parameters such as the number of hidden layers, the size
of the hidden layers, value of the moment constant and
learning rate, and type of the activation functions.
Wavelet energy values for each signature image are the
input features to WNN input layer. Each neuron in
WNN output layer represent a person.

In the test stage , Appling the test wavelet energy                             Figure 11: Comparative Study between signature image pixels
values of the test signature to trained WNN. Evaluating                      intensity value as input to ANN (ANN)and signature wavelet energy
the proposed off-line handwritten signature recognition                        values as input to ANN (WE+ANN)and signature wavelet energy
system by recognition rate to each person as shown in                               values as input to WNN (WE+WNN)with training data.
Figure 10.




                                                                                Figure 12: Comparative Study between signature image pixels
                                                                             intensity value as input to ANN (ANN)and signature wavelet energy
                                                                               values as input to ANN (WE+ANN)and signature wavelet energy
  Figure 10: Proposed off-line handwritten signature recognition                       values as input to WNN (WE+WNN) testing data.
                          system result
                                                                                 V.    CONCLUSIONS AND FUTURE WORK
All system evaluation is made by two concept False
Acceptance Rate(FAR) which indicates how many
                                                                            Handwritten signature recognition plays an important
forgeries were incorrectly classified as genuine
signatures ,and False Rejected Rate(FRR) which                              role in our daily life especially in any bank and any
indicates how many genuine signatures were incorrectly                      ATM system. Off-Line Handwritten Signature
rejected by the system. To the training signatures FAR                      recognition is a difficult task than On-line one because
and FRR is 0.01% and to the testing signatures FAR                          of absence of dynamic information in off-Line signature
and FRR is 0.07% .                                                          image such as angle of written style of written and so
                                                                            on. This paper proposed an off-Line handwritten
To evaluate our proposed system a comparative study
                                                                            recognition system with Four stages. First stage is
between three off-line handwritten signature systems is
                                                                            scanning signature image and removing noise using
made:
                                                                            median filter. Second stage, extract feature from each
1-signature image pixels intensity value as input to                        signature image using DWT technique with the
ANN(ANN)                                                                    advantage of multi-scale and with respect the
2- signature wavelet energy values as input to ANN
                                                                            translation, rotation and scale variations of signature
(WE+ANN)
                                                                            image. Computing wavelet energy values from DWT
3- Our proposed system signature wavelet energy
                                                                            details sub-bands coefficient to all person signature
values as input to WNN (WE+WNN).                                            images using the suitable wavelet function to our
                                                                            database. Daubechies 20 (db20) is recognize as a
 Figure 11 represent the recognition rate to each
training person data and Figure 12 represent the                            suitable wavelet function with three levels analysis after
                                                                            a comparative study with other wavelet function. Third
recognition rate to each testing person data. Figure 11
                                                                            stage, taking the wavelet energy values as input to
and Figure 12 concluded that our proposed system has
                                                                            WNN with Morlet function as activation function in
the highest recognition rate.
                                                                            hidden layer. Finally, testing trained WNN with
                                                                            seen/unseen signature to evaluate our proposed system
                                                                                                                                              8

                                                                      20                               http://sites.google.com/site/ijcsis/
                                                                                                       ISSN 1947-5500
                                                                   (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                             Vol. 8, No. 6, 2010

recognition rate. A comparative study between three                             [12] K R Radhika, M K Venkatesha and G N Sekhar,” Pattern
                                                                                    Recognition Techniques in Off-line hand written signature
off-line handwritten signature systems is made
                                                                                    verification - A Survey”, PROCEEDINGS OF WORLD
( signature image pixels intensity as input to ANN ,                                ACADEMY        OF    SCIENCE,     ENGINEERING       AND
signature wavelet energy values as input to ANN and                                 TECHNOLOGY ,vol. 36, ISSN 2070-3740 , 2008.
signature wavelet energy values as input to WNN). The
                                                                                [13] Engin Avci , Abdulkadir Sengur, Davut Hanbay, " An optimum
conclusion will found that our proposed system
                                                                                    feature extraction method for texture classification“, Expert
(wavelet energy values as input to WNN) has high                                    Systems with Applications: An International Journal, Published
recognition rate.                                                                   by Elsevier Ltd, Volume 36 , Issue 3,2009,p.p 6036-6043.

To improve our system recognition rate, each person                             [14]     http://www.dtic.upf.edu/~xserra/cursos/TDP/referencies/Park-
                                                                                       DWT.pdf , 17-8-2010.
signature should have its own wavelet function in
feature extraction stage. Genetic algorithm will be used                        [15]S.Sitharama Lyengar,E.C.Cho,Vir V.Phoh ,”Foundations of
as a searching strategy in the future work to found the                             Wavelet         Networks        and           Application”,
suitable wavelet function to each person signature.                                 chapman&Hall/CRC Press LLC , chapter 4, 2002.

                                                                                [16] Xian-Bin Wen, Hua Zhang, and Fa-Yu Wang,” A Wavelet
                 ACKNOWLEDGEMENTS:                                                  Neural Network for SAR Image Segmentation”, Sensors ,
                                                                                    Vol.9,No.9,2009,p.p 7509-7515 .
The authors would like to thank .Prof. Albert Swart for
                                                                                [17] Zhang Q. and Benveniste A,“Wavelet networks“, IEEE Trans.
making his signature database available to us.
                                                                                    On Neural Networks ,Vol.3, ,1992,p.p 889-898.

The first author would like to thank Sarah El.metwally                          [18] ri Gross and Longin Jan Latecki,”Digital geometric methods in
and Eslam Foud for their encouragements.                                            document image analysis”,Pattern Recognition,Vol. 32,No.3,
                                                                                    1999,pp. 407.
                         REFERENCES:                                            [19] http://homepages.inf.ed.ac.uk/rbf/HIPR2/median.htm, 24-6-2010.

[1] S. Impedovo, G. Pirlo,” Verification of Handwritten Signatures: an          [20]Avci, E., Turkoglu, I., & Poyraz, M.,”Intelligent target
     Overview”, 14th International Conference on Image Analysis and                 recognition based on wavelet packet neural network”, Experts
     Processing , 2007, pp. 191-196 .                                               Systems with Applications: An International Journal,vol 29,Issue
                                                                                    1 ,2005,p.p 175-182.
[2] M. Ammar, Y. Yoshida, T. Fulumura, “ A New Effective
    Approach    for Off-line Verification of Signatures by Using                [21] Samaneh Ghandali, Mohsen Ebrahimi Moghaddam, “Off-Line
    Pressure Features”,. Proceedings of the 8th International                       Persian Signature Identification and Verification Based on Image
    Conference on Pattern Recognition, 1986, p.p 566–569.                           Registration and Fusion”, JOURNAL OF MULTIMEDIA, VOL.
                                                                                    4, NO. 3, 2009.
[3] Y. Qi , B.R. Hunt, “Signature Verification Using Global and Grid
    Features,” Pattern Recognition, vol.27, Issue.12, 1994, p.p 1621-           [22] A. Wahi, E. Thirumurugan “Recognition of Objects by
    1629.                                                                           Supervised Neural Network using Wavelet Features”, First
                                                                                    International Conference on Emerging Trends in Engineering and
[4] V.E. Ramesh, M.N. Murty, “Off-line Signature Verification Using                 Technology,2008.
    Genetically Optimized Weighted Features”, Pattern Recognition,
    vol.32, Issue.2, 1999, p.p 217-233.                                         [23] Chi-Man Pun and Moon-Chuen Lee ,”Log-Polar Wavelet Energy
                                                                                    Signatures for Rotation and Scale Invariant Texture
[5] P.S. Deng, H.-Y. M. Liao, C.W. Ho, and H.-R. Tyan,:“Wavelet-                    Classification”, IEEE TRANSACTIONS ON PATTERN
    Based Off-line Handwritten Signature Verification”, Computer                    ANALYSIS AND MACHINE INTELLIGENCE, VOL. 25, NO.
    Vision and Image Understanding, vol.76, no.3, 1999, p.p 173 -                   5, 2003
    190 .
                                                                                [24]http://www.mathworks.com/access/helpdesk/help/toolbox/wavele
[6]http://dspace.mak.ac.ug/bitstream/123456789/599/3/karanja-                       t/wenergy2.html, 17-8-2010.
    evanson-mwangi-cit-masters-report.pdf,17-8-2010
                                                                                [25] Alan McCabe, Jarrod Trevathan and Wayne Read,” Neural
[7] M. Kalera, S. Srihari, and A. Xu. “Offline signature verification             Network-based Handwritten Signature Verification”, JOURNAL
    and identification using distance statistics”, International Journal          OF COMPUTERS, VOL. 3, NO. 8, 2008.
    of Pattern Recognition and Artificial Intelligence. vol. 18, no.7,
    2004, pp. 1339-1360.                                                        [26] Insung Jung, and Gi-Nam Wang,” Pattern Classification of Back-
                                                                                    Propagation Algorithm Using Exclusive Connecting Network”,
[8] V. Nalwa.” Automatic on-line signature verification”, Lecture                   World Academy of Science, Engineering and Technology,VOL.
    Notes In Computer Science, Proceedings of the Third Asian                       36 ,2007.
    Conference on Computer Vision,1998, p.p 10 - 15 .
                                                                                [27] Ming Meng , Wei Sun,” Short-term Load Forecasting Based on
[9]http://research.microsoft.com/pubs/69437/handwritingregistration_                Rough Set and Wavelet Neural Network ”, International
    cvpr07.pdf, 17-8-2010.                                                          Conference on Computational Intelligence and Security,2008.

[10]Sing-Tze      Bow,“Pattern     recognition      and          image
    preprocessing”,Marcel Dekker,Inc, chapter 15,2002.                          [28] Mohd Fazril, Zaki Ahmad, Hj. Kamaruzaman, “The Performance
                                                                                    of Two Mothers Wavelets in Function Approximation”, Joural of
[11] G.Y. Chen, T.D. Bui, A. Krzyzak,” Contour-based handwritten                    Mathematical Research,Vol.1,No.2,2009.
    numeral recognition using multi-wavelets and neural networks”,
    Pattern Recognition ,Vol.36 ,2003,p.p 1597 – 1604.
                                                                                                                                                     9

                                                                           21                                http://sites.google.com/site/ijcsis/
                                                                                                             ISSN 1947-5500

				
DOCUMENT INFO
Description: IJCSIS is an open access publishing venue for research in general computer science and information security. Target Audience: IT academics, university IT faculties; industry IT departments; government departments; the mobile industry and computing industry. Coverage includes: security infrastructures, network security: Internet security, content protection, cryptography, steganography and formal methods in information security; computer science, computer applications, multimedia systems, software, information systems, intelligent systems, web services, data mining, wireless communication, networking and technologies, innovation technology and management. The average paper acceptance rate for IJCSIS issues is kept at 25-30% with an aim to provide selective research work of quality in the areas of computer science and engineering. Thanks for your contributions in September 2010 issue and we are grateful to the experienced team of reviewers for providing valuable comments.