Neural Networks based Off-line Signature Verifcation

A B. Tech Project Report Submittedin Partial Fulfillment of the Requirements

for the Degree of

Bachelor of Technology

by

M. Raghu Ram Reddy(05010221)

C. Sai Ravi Kiran(05010232)

under the guidance of

C. Mahantha

to the

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

INDIAN INSTITUTE OF TECHNOLOGY GUWAHATIGUWAHATI - 781039, ASSAM

CERTIFICATE

This is to certify that the work contained in this thesis entitled “Neural Networks based

Off-line Signature Verifcation” is a bonafide work of (Roll No. 05010221,05010232),

carried out in the Department of Electronincs and Communication Engineering, Indian In-

stitute of Technology Guwahati under my supervision and that it has not been submitted

elsewhere for a degree.

Supervisor: C. Mahantha

Professor,

May, 2009 Department of Electronincs & Communication Engineering,

Guwahati. Indian Institute of Technology Guwahati, Assam.

i

Acknowledgements

We take this opportunity to thank our BTP supervisor, Dr. C.Mahanta for her able

guidance, support and patience. The guidance and the feedback she provided were very

important in getting our project done. We would also like to thank our Department of

Electronics and Communications Engineering, Indian Institute of Technology Guwahati.

ii

Contents

List of Figures iv

List of Tables v

1 Introduction 4

1.1 Neural Network: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 PRIOR WORK: 8

3 Preprocessing and Features Extraction 11

3.1 Data Capturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.2 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2.1 Data Cropping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2.2 Noise reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2.3 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.2.4 Width normalization . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.2.5 Thinning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.3.1 Global Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.3.2 Grid Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Training and Testing 18

iii

4.1 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.1.1 Radial Basis Function (RBF) Neural Networks . . . . . . . . . . . . 19

4.2 testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

5 Conclusion and Future Work 21

5.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

References 23

iv

List of Figures

3.1 Image containing ’salt and pepper noise’ . . . . . . . . . . . . . . . . . . . 13

3.2 Output Image from Median Filter: . . . . . . . . . . . . . . . . . . . . . . 14

4.1 Architectural Graph of RBF . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1

Abstract

Signing documents is something that most every adult is familiar with. In our

personal lives we sign everything from personal checks to birthday cards. In the

business world we sign things such as expense accounts and other official docu-

ments. This lends itself well for signature recognition to be used as a means of

biometric verification in electronic commerce. This type of signature identifi-

cation is different however from the normal two-dimensional signature that one

would find on a form or document. Biometric signature recognition operates

in a three-dimensional environment where, not only is the height and width

of pen strokes measured, but also the amount of pressure applied in the pen

stroke to measure the depth that would occur as if the stroke was made in the

air. This helps to reduce the risk of forgery that can occur in two-dimensional

signatures.

One drawback to this form of encryption is that people do not always sign

documents in exactly the same manner. The angle at which they sign may be

different due to seating position or due to hand placement on the writing sur-

face. Therefore, even though it is three dimensional which adds to its ability to

discern impostors, it is not as accurate as other forms of biometric verification.

These types of systems are not as expensive as some of the higher end systems

such as iris scanners, and they are priced more in the range of voice and fin-

gerprint scanners which makes them quite affordable for network use.

2

Signature verification problem is concerned with determining whether a par-

ticular signature truly belongs to a person, so that forgeries can be detected.

In signature recognition problem, a signature database is searched to establish

identity of a given signature. As signatures are the pri-mary mechanism both

for authentication and authorization in legal transactions, the need for research

in efficient auto-mated solutions for signature recognition and verification has

increased in recent years.

In this paper we present an off-line signature recognition and verification sys-

tem using global and grid features of the signatures. An artificial neural net-

work based on Radial Basis Function (RBF) Neural Networks. Performance

measures like the learning rate, FAR and FRR are analyzed.

3

Chapter 1

Introduction

For centuries, handwritten signatures have been an integral part of validating business

transactions, contracts and agreements..The distinctiveness of a handwritten signature

helps to prove the identity of the signer, while the act of signing a document represents the

signer’s acceptance of its terms and also codifies the document’s contents as being official

and complete at the time it was signed.

Signature Recognition: Biometric identification by automatically scanning a person’s

signature and matching it electronically against a library of known signatures.

The four legal properties of a handwritten signature are briefly stated below:

1. Authentication of the signer: a handwritten signature allows positive verification

of the signer’s identity

2. Acceptance: the signature conveys willful intent and acceptance of the terms stated

in the document.

3. Integrity: the signature establishes the integrity of the signed document, indicating

that it has not been altered in any way.

4. Non-repudiation:the accumulated effect of the above three factors promises such a

high degree of purpose that the signer cannot deny he or she has signed.

4

Handwritten signatures are of different shapes and sizes and the variations in them are so

immense that it is difficult for a human being to distinguish a genuine signature from a

forged one by having a glance at the signature.There are different types of signatures used

in real life. Broadly speaking, signatures can be classified as ,

1. Simple Signatures: These are the ones where the person just writes his or her name

2. Cursive signatures:These are the ones that are written in a cursive way

3. Graphical signatures:The signatures can be classified as graphical when cursive

signatures depict geometric patterns.

Automated recognition of handwritten signatures became imperative when it was dif-

ficult to distinguish genuine signatures from simulated forgeries on the basis of visual as-

sessment. This led to computer recognition of handwritten signatures, which though a bit

slow, is more reliable and efficient.

There are 2 types of handwritten recognition schemes practised today:

1. Online recognition:

On-line handwriting recognition involves the automatic conversion of text as it is

written on a special digitizer, where a sensor picks up the pen-tip movements as well

as pen-up or pen-down switching. That kind of data is known as digital ink and

can be regarded as a dynamic representation of handwriting. The obtained signal

is converted into letter codes which are usable within computer and text-processing

applications.This method focusses on Dynamic systems produce signals varying with

time (including velocity,acceleration, pressure, position). The signer uses the optical

pen and starts writing on the paper.The sensor picks up the image and also the

physical characteristics of the handwriting like velocity of movement of hand and

accelaration between handstrokes and pressure exerted at the position and records

the data along with the dynamic image of the signature.When the signature system is

5

trained then when the signee signs the document the image is picked up dynamically

and compared with the data stored for the user.If the data matches then the user

is authenticated otherwise he isnt.This method is the most viable but expensive.The

forgery must be extremely perfect to get around this but the method cannot be

bypassed by casual copies of signatures.

2. Offline recognition:

Off-line handwriting recognition involves the automatic conversion of text in an im-

age into letter codes which are usable within computer and text-processing appli-

cations. The data obtained by this form is regarded as a static representation of

handwriting.The technology is successfully used by businesses which process lots of

handwritten documents, like insurance companies. The quality of recognition can

be substantially increased by structuring the document (by using forms). In Offline

recognition case the signature appears as a 2D (graylevel or binary) image. The

static signature verification is considered to be much moredifficult because timing

and dynamic information are highly degraded in that case.The off-line method uses

an optical scanner to obtain the handwriting data written on paper.In this mechanism

the user signs on a piece of paper which is read by a scanner or a camera.The image

is then fed to a computer.The computer stores the image as specific to the signer.It

is used to identify the user by the image.

In the present project our aim is to create a offline recognition scheme using neural

networks.

1.1 Neural Network:

Artificial neural networks are made up of interconnecting artificial neurons.The Perceptron

is a binary classifier that maps its input x to an output value f across the matrix.A multilayer

perceptron is a feedforward artificial neural network model that maps sets of input data onto

6

a set of appropriate output. It is a modification of the standard linear perceptron in that

it uses three or more layers of neurons with nonlinear activation functions.The Perceptron

is a single layer neural network whose weights and biases could be trained to produce a

correct target vector when presented with the corresponding input vector. The training

technique used is called the perceptron learning rule. The perceptron generated great

interest due to its ability to generalize from its training vectors and work with randomly

distributed connections. Perceptrons are especially suited for simple problems in pattern

classification.Our perceptron network consists of a single neuron connected to two inputs

through a set of 2 weights, with an additional bias input.The perceptron calculates its

output using the following equation:

f(x) = 1 if w.x+ b > 0 0 else (1.1)

In general the Neural Network can be trained to store a specific bit vector.Then the

vector which is the input is stored in the database of the neural network.The weights of

the neural network are trained through learning mechanisms.

7

Chapter 2

PRIOR WORK:

A great deal of work has been done in the area of offline signature verification for the detec-

tion of random forgeries. Earlier work on offline signature verification deals primarily with

casual and random forgeries. Many searchers therefore found it sufficient to consider only

the global features of a signature.As signature databases became larger and archers moved

toward more difficult skilled forgery detection tasks,we saw a progression not only to more

elaborate classifiers,but also to the increased use of local features and matchingtechniques.

1. Baltzakis developed a neural network-based system forthe detection of random forg-

eries. The system uses globalfeatures, grid features , and texture features to represent

each signature. Foreach one of these feature sets, a special two-stage perceptron one

class one network classification structureis implemented. In the first stage, the clas-

sifier combines thedecision results of the neural networks and the Euclidean distance

obtained using the three feature sets.

2. Kaewkongka uses the Hough transform to extract the parameterised Hough spacefrom

a signature skeleton as a unique characteristic featureof a signature. A backpropaga-

tion neural network is used toevaluate the performance of the method.

3. Quek used global baseline features ( the vertical and horizontal position in the signa-

8

ture image which corresponds to the peak in the frequency histogram of the vertical

and horizontal projection of the binary image, respectively), pressure features (that

correspond to high pressure regions in the signature), and slant features (which are

found by examining the neighbours of each pixel of the thinned signature). He then

conducts two types of experiments. The first group of experiments use genuine sig-

natures and forgeries as training data, while the second group of experiments use

only genuine signatures as training data.These experimentsare conducted on the sig-

natures of 15 different writers, thatis, 5 writers from 3 different ethnic groups. For

each writer,5 genuine signatures and 5 skilled forgeries are submitted.When genuine

signatures and forgeries are used as trainingdata, the average of the individual EERs

is 22.4%.

4. El-Yacoubi uses HMMs and the cross-validation principle for random forgery detec-

tion. A grid is superimposed oneach signature image, egmenting it into local square

cells.From each cell, the pixel density is computed so that each pixel density repre-

sents a local feature. Each signature image is therefore represented by a sequence

of feature vectors,where each feature vector represents the pixel densities associated

with a column of cells. The cross-validation principleinvolves the use of a subset (vali-

dation set) of each writerstraining set for validation purposes. Since this system aims

to detect only random forgeries, subsets of other writers training sets are used for

impostor validation. Two experiment sare conducted on two independent data sets,

where each data set contains the signatures of 40 and 60 writers, respectively.Both

experiments use 20 genuine signatures for training and 10 for validation

5. Justino uses a discrete observation HMM to detectrandom, casual, and skilled forg-

eries. A grid segmentation scheme is used to extract three features: a pixel density

feature, a pixel distribution feature (extended-shadow-code),and an axial slant fea-

ture. A cross-validation procedure isused to dynamically define the optimal number

9

of states foreach model (writer). Two data sets are used. The first data setcontains

the signatures of 40 writers with 40 genuine signa tures per writer. This data set is

used to determine the optimal codebook size for detecting random forgeries. This op-

timised system is then used to detect random, casual, andskilled forgeries in a second

data set. The second data set contains the signatures of 60 writers with 40 training

signatures, 10 genuine test signatures, 10 casual forgeries, and 10 skilled forgeries per

writer rate).An FRR of 2.83 % and an FAR of1.44%, 2.50%, and 22.67% are reported

for random, casual,and skilled forgeries, respectively

In this project we propose an off-line signature recognition system, which is based

on neural network structure in combinations with two powerfully feature sets. We use two

different sets of features, each describing a different aspect of the signature: global features,

grid information features. The Radial Basis Function algorithm is used to train the neural

network.

10

Chapter 3

Preprocessing and Features

Extraction

A signature verification system has five components: Data capture:the process of con-

verting the signature into digital form.

Preprocessing:the data transformation in a standard format.

Feature extraction:the process of extracting key information from the digital represen-

tation of the signature.

Comparison process:matches extracted features with templates stored in a database.

Usually, the output is a fit ratio.

Performance evaluation:the decision step typically made by thresholding the fit value.

Before extracting features from the signature, the signature has to be preprocessed to

give accurate results.

3.1 Data Capturing

In offline signature verification, first the signature is scanned. If the scanned image is not a

gray scale image then it is converted into gray image using the matlab function rgb2gray(),

11

then its read using matlab command imread() to give image data. This binary data is used

in preprocessing of the signature. The commands imshow() and imtool() can be used to

see the gray image.

3.2 Preprocessing

The preprocessing stage has four different parts: noise reduction, data area cropping, width

normalization and skeletonization. It is noted that the proposed method is appl- ied to

binary images only.

3.2.1 Data Cropping

Data cropping is a process implied to seperate the signature from the background. This is

done by using p-tile thresholding. p-tile Thresholding: In this process a threshold value

is selected and in the image data the pixel values below the threshold are set as ’0’ and

above the threshold are set as ’1’. Using this method the signature is seperated from the

dark background.

3.2.2 Noise reduction

Dirt on camera or scanner lens, imperfections in the scanner lighting, etc introduces noises

in the scanned signature images. A filtering function is used to removal the noises in the

image. Filtering function works like a majority function that replaces each pixel by its

majority function. A noise reduction filter is applied to the binary scanned image. The

goal is to eliminate single white pixels on black background and single black pixels on white

back ground. In order to accomplish this, we apply a 3 x 3 mask to the image with a simple

decision rule: if the number of the 8-neighbors of a pixel that have the same color with the

central pixel is less than two, we reverse the color of the central pixel.

Noise reduction can be done using several matlab functions depending on the noise and

type of feature extraction. Some of those functions are medfilt2, wiener2, filter2(fspecial(’average’,3),*)

12

etc.

To remove the ’salt and pepper noise’ we can use the median filter. This median filter

passes a window on the image and keeps shifting the image. Noise elimination is done by

using the matlab function medfilt2(). Figure(3.1) shows the image with noise. Figure(3.2)

shows the image after removal of noise.

Fig. 3.1 Image containing ’salt and pepper noise’

3.2.3 Segmentation

For segmentation a projection technique is used. Specifically, vertical and horizontal pro-

jections are calculated and using of two threshold values we identify the area of the image

that contains the signature (we discard the white space surrounding the sig- nature).

13

Fig. 3.2 Output Image from Median Filter:

3.2.4 Width normalization

Height and width of signatures vary from the person to person or sometimes same person

may use different size signatures. Therefore these size differences need to be elimin- ated.

During the normalization process The image size is adjusted so that the width reaches a

default value while the heighttowidth ratio remains unchanged. The size normalization in

offline signature verification is important because it establishes a common ground for image

comparison. A low spatial resolution makes all signatures look like the same while a very

high spatial resolution may high lighten the variability.

3.2.5 Thinning

Thinning is a morphological operation that is used to remove selected foreground pixels

from binary image, somewhat like opening. It can be used for several applications, but

is particularly useful for skeletonization. In this mode it is commonly used to tidy up

14

the output of edge detectors by reducing all lines to single pixel thickness. Thinning is

normally only applied to binary image, and produces another binary image as output.To

reduce thickness into single pixel thickness and have base shape of the signature, Hilditch’s

skeletoniziation algorithm is applied to the binary image. The skeletonization is done using

the matlab function ”bwmorph(bw,’thin’,Inf)”.

3.3 Feature Extraction

The choice of the features that will be provided to the classifiers of the system is very im-

portant. In this work we use two different sets of features. global features, grid information

features . Both global features and grid information features are classic in pattern recogni-

tion problems. The global features provide information about specific cases concerning the

structure of the signature. The grid features provide information about the area of blocks

of pixels.

3.3.1 Global Features

1. Height to Width ratio:The height and width of a signature were directly measured

as the size of the image after blank edges of the image were removed. The blank edge

removal was accomplished by sequentially eliminating rows (or columns) on the edge

of the image if the total number of pixels in the row (or column) was less than 2 (an

arbitrary threshold). The ratio of the height and width is calculated.

2. Image area: The number of black pixels of the image. When it is about skeletonized

images, it represents a measure of the lines density in a signature image.

3. Pure width: The width of the image with horizontal blank spaces removed.

4. Pure height: The height of the image with vertical blank spaces removed.

5. Baseline shift: The defcrence between the vertical centers of gravity of the left and

15

the right part of the image. It was taken as a measure for the orient- ation of the

signature.

6. Slant Angle:Though slant angle represents a prominent, idiosyncratic feature of a

signature, it is mainly a psychophysical concept. A well-established alg- orithm for

measuring slant angle is not available. A way to find slant angle is find global and

local slant angles. Global slant angle: Nagel and Rosenfeld [8] define the slant

angles from local properties only and do not capture the global slant property of a

signature. In the proposed system, the image is first rotated by 30degree. in the clock-

wise direction accommodating signatures that have a heavy slant. The image is then

rotated in the anti-clockwise direction in steps of 2degree. and horizontal projection

is calculated. The rotation is stopped when the horizontal projection reaches the

maximum and the slant angle is the difference of 30.. and the angle at which the

horizontal projection is the maximum.

7. COM:Vertical center of mass & Horizonal centre of mass.

cy =

∑maxyy=1 y

∑maxxx=1 b[x, y]∑maxx

x=1

∑maxyy=1 b[x, y]

(3.1)

cx =

∑maxxx=1 x

∑maxyy=1 b[x, y]∑maxx

x=1

∑maxyy=1 b[x, y]

(3.2)

8. Max. Projections:Maximum Vertical projection & Maximum Horizontal projec-

tion.

9. Number of edge points: The pixels, which have only one immediate neighbor, are

the edge points. For example, the letter h has 3 edge points.

10. Number of cross points: The pixels, which have more than or equal to three

neighbors is considered to be a cross point. For example, the letter t has 1 cross

point.

16

11. Number of closed loops: The number of closed loops can be defined as CL = 1

+ (EL EP)/2 with EP denoting the number of edge points and EL the number of

extra departures, defined as sum of number of 8-neighbors over all cross points minus

2. This describes the amount of complexity that the signature lines involve.

3.3.2 Grid Features

The image is divided to 96 rectangular rcgions (12x8) and for each region, the area is

calculated.

17

Chapter 4

Training and Testing

4.1 Training

Neural networks are among the most commonly used classifiers for pattern recognition

problems.Training a neural network model essentially means selecting one model from the

set of allowed models that minimises the cost criterion. There are numerous algorithms

available for training neural network models; most of them can be viewed as a straightfor-

ward application of optimization theory and statistical estimation.

In signature verification once the signature features are extracted, we use this database

of signatures to train the neural network. The neural network algorithm used to train the

network is Radial Basis Function neural network.The network was trained by modifying

the weights between the layers. For each training iteration (epoch), the error is calculated

by taking the difference between the output and the expected output. System performance

is reported in terms of false rejection rate and false acceptance rate. False Rejection Rate

(FRR) is the percentage of genuine signatures rejected as false and False Acceptance Rate

(FAR) is the percentage of false signatures accepted as genuine.

18

4.1.1 Radial Basis Function (RBF) Neural Networks

This approach is taken by viewing the design of a neural network as a curve-fitting ap-

proximation problem in a high-dimensional space. According to this view point, learning

is equivalent to finding a surface in a multidimensional space that provides a best fit to

the training data. The input layer is made up of source nodes. The second layer is a

hidden layer of high enough dimension, which serves a different purpose from than in a

multilayer perceptron. The output layer supplies the response of the network to the acti-

vation patterns applied to the input layer. The transformation from the input space to the

hidden-unit space is nonlinear, whereas the transformation from the hidden-unit space to

the output space is linear. The mathematical expression of an RBF is given as

y =∑

i

Wig1(x) =∑

i

Wigi(||x− ci||) (4.1)

where wi is the weight from the ith neuron of the hidden layer to the output layer.

gi(x) is an activation function and in general Gaussian Function. In Gaussian Function, x

denotes input vector, ci denotes center, where ——Ux-ci—— standard Euclidean distance,

and i as spread. Gaussian Function’s mathematical expression is

gi(||x− ci||) = exp(−||x− ci||2

2σi2

) (4.2)

Figure:1 shows the architectural graph of a RBF.

4.2 testing

In the recognition phase, features from the verification samples were given as input to the

neural network and the value obtained from the output layer of the neural network was

observed. If the output value is close to 1 in the correct position, then this implies that

the corresponding person is recognized. Test signature samples include samples from the

19

Fig. 4.1 Architectural Graph of RBF

training set and 5 genuine signatures samples. These test samples were used for recognition.

For verification, 5 simple forged signatures were used for forgery detection.

20

Chapter 5

Conclusion and Future Work

5.1 Results

In the signature verification process while testing the neural network with training samples

the system correctly classified all the signatures, but when new testing samples are used

only95.83 percent of the samples are classified correctly.

i.e. the RBF function classified all the training samples correctly but some forgery

signatures were accepted as original signatures.

5.2 Conclusion

The main objective of the project is to verify a signature offline using neural networks.An

off-line signature recognition system using RBF(Radial Basis Function) neural network is

implemented in this project. The results show that the RBF gives a good accuracy in terms

of FRR and FAR rates.

5.3 Future Work

The future work of this project is to verify the signature database using other efficient

verification methods like Conic Section Function Neural Network (CSFNN), Multil Layer

21

Perceptron(MLP) Neural Network, Multi Layer Backpropogation method etc and compare

the reults of RBF Neural Network with these results. The signature verification can also

be changed by changing the features that can be extracted from a signature.So, the future

work of the verification of signature can be done with the same Neural Network methods

but using different signature features and compare the results with results of the present

project.

22

References

[1] Bajaj R., Chaudhary S., ”Signature Verification using multiple neural classi-

fiers”,Pattern recognition society 30,(1996),1-7

[2] Baltzakis H., Papamarkos N., ”A signature verification technique based on two stage

classifier,” Engineering applications of artificial science 14, (1998), 95-103.

[3] Brault J., Plamondon R., ”Segmenting handwritten signatures at their perceptually

important points,” IEEE sanctions on pattern analysis and machine intelligence (1993),

Vol 15.,pp. 953-957.

[4] Hanmandlu M., Madasu V.K.,Madasu S. ”NeuroFuzzy approaches to signature verifi-

cation,” Second national analysis on Document analysis and recognition (2003)

[5] Haykin S., ”Neural Networks A comprehensive foundation,” Macmillan college publish-

ing (1994)

[6] Huang K.,Yan H. ”Online signature verification based on Geometric Feature Extraction

and Neural Network Classification,” Pattern Recognition society , (/1996),9-17.

[7] Kiet T.H., Palaniappan R.,Raveendran P.,Takeda F. ”Signature verification system

using Pen Pressure for internet and E-commerce applications,” ISSRE (2001), Interna-

tional symposium on software reliability enginerring, Hong Kong,(2001)

23

[8] Murshed N.A.,Bortolozzi F.,Sabourin R. ”Offline signature verification us-

ing fuzzy ARTMAP neural network,” International Conference on neural net-

works,(1995),Perth,Australia,(1995)

[9] Oz C.,Ercal F.,Demir Z. ”Signature recognition and verification with an ANN,” Inter-

national Conference on Electrical and Electronics Enmgineering ,(2003), Bursa

[10] Parizeu M., Plamondon R., ”A comparative analysis of regional correlating dynamic

time warping and skeletal tree matching for signature verification,” IEEE transactions

on pattern analysis and machine intelligence (1990), , pp. 710-717.

[11] Yildirim T.,Ozyilmaz L., ”Dimensionality reduction in conic section Function neural

network ,” Sadhana academy proceedings in Engineering sciences, 2 (2002),675-683.

[12] Yildirim T., ”Development of conic section function Neural network in software and

analogue hardware,” Ph.D.Thesis,(1997)

[13] Qui Y.,Hunt B., ”Signature verification using global and grid features”, Pattern recog-

nition 27, 1 (1994), 1621-1629.

[14] Han K.,Sethi I. ”H?andwritten signature retrieval and identification,” Pattern recog-

nition letters (1996), pp. 83-90.

[15] Akban M., Lim Y., ”Off line signature verification using thickened templates,” COM-

CON, 3 (1995).

[16] Droughard j.,Sabourin R.,Godbout M. ”A neural network approach to offline Signature

verification using directional pdf,” Pattern Recognition , 11 (1996), 415-424.

[17] Lee L., Berger T. ”Reliable online signature verification system,” IEEE transactions

on Pattern Analysis and Machine learning 18 (1996), 643-647.

[18] Pandya A., Robert B. M., ”Pattern recognition in neural networks,” CRC Press and

IEEE press (1995),

24

[19] Lam L., Sucn C., ”A dynamic shape preserving thinning algorithm.,” Signal Processing

(1991), pp. 199-208.

[20] Bow S., ”Pattern Recognition and Image Preprocessing” Marcel Deccar (1992),

25