IMPROVEMENT ON THE INDIVIDUAL RECOGNITION

SYSTEM WITH WRITING PRESSURE BASED ON RBF

Lina Mi, Fumiaki Takeda

Department of Intelligent Mechanical System, Kochi University of Technology

185 Miyanokuchi, Tosayamada-cyo, Kami-gun, Kochi, Japan

Keywords: Writing pressure, RBF, Gaussian function, Neuro-template

Abstract: In our previous research work, an individual re

cognition system with writing pressure employing neuro-

template of multiplayer feedforward network with sigmoid function has been developed. Although this

system was effective on recognition for known registrant, its rejection capability for counterfeit signature

was not good enough for commercial application. In this paper, a new activation function was proposed to

improve the rejection performance of the system for counterfeit signature on the premise of ensuring the

recognition performance for known signature. The experiment results showed that compared with original

system the proposed activation function was seemed to be effective to improve the rejection capability of

the system for counterfeit signature with keeping the recognition capability for known signature satisfied.

1 INTRODUCTION

In recently years, biometrics information such as

fingerprint, iris is gradually applied into the

individual recognition systems replacing the security

material such as licence or security information such

as personal identification number (PIN), password.

Compared with conventional recognition systems,

high security is one of superiority of recognition

system with biometrics information because of

difficulty of imitation or copy for biometrics

information.

Every person’s signature has unique character in

ter

ms of rhythm and force employed during

signature. Writing pressure data detected

dynamically from signature procedure can

characterize individual signature if being properly

processed. Being friendly biometrics information,

writing pressure has little spirit resistance compared

with other biometrics information such as fingerprint

and iris. Furthermore, unlike handwriting, writing

pressure data is dynamic signal and invisible to users,

that contributes to more difficulties of imitation and

copy, therefore writing pressure has higher security

as personal information for individual recognition

than handwriting. In our research work, writing

pressure has been successfully employed to develop

individual recognition system.

In our system, neuro-template of multiplayer

feed

forward neural network (MFNN) with sigmoid

as activation function of hidden and output layer was

used to construct kernel part of the system. The

experiments showed that the system with sigmoid

function was effective on recognition for the patterns

having been learned (known registrant), but the

rejection capability for counterfeit writing which is

from imitating legitimate registrant’s signature by

intruders was not good. To solve the previous

problem, Gaussian function, which is one of radial

basic function (RBF), was proposed as activation

function of the NN in this paper. In the experiment,

both recognition and rejection performance of the

system with Gaussian function were compared with

that of original system with sigmoid function, and

the results showed that the proposed method has

effectively improved the rejection performance of

the system for counterfeit signature with keeping

recognition capability for known pattern satisfied.

2 BASIC STRUCTURE OF THE

INDIVIDUAL RECOGNITION

SYSTEM

The individual recognition system with writing

pressure is composed of hardware section and

software section.

157

Mi L. and Takeda F. (2005).

IMPROVEMENT ON THE INDIVIDUAL RECOGNITION SYSTEM WITH WRITING PRESSURE BASED ON RBF.

In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 157-162

DOI: 10.5220/0002511401570162

 SciTePress

2.1 Hardware Construction of the

System

The hardware section of system, which is

constructed by the electronic pen, data collection

box and personal computer, is in charge of data

detection and transmission. Figure 1 illustrated the

hardware structure of the system. In inner of

electronic pen, there is a pressure sensor with 0.1g

writing pressure resolution and 4ms time resolution,

that makes high resolution of the extraction of

writing pressure data possible. In the hardware

section, writing pressure data are detected by the

electronic pen, and then loaded to PC via the data

collection box. After that, the registration and

recognition procedure are executed on these data

based on the Neuro-template.

2.2 Software Construction of the

System

This section completes most function of the system

such as data preprocess, new registration, registrant

recognition and result display etc, It is composed of

registration part and recognition part which are

independent each other. Both of two parts are based

on neural template of MFNNs. Registration

subsystem is in charge of recruiting new legitimate

registrant: generating and training one neuro-

template for him (or her). Recognition subsystem

has role of recognizing the identity of the user who

entries the system from all the candidate registrants

who have registered on the system legitimately by

matching this user’s signature with all of existing

neuro-templates and then evaluating output of each

template.

3 PREPROCESSING

The writing pressure data detected by hardware

section can not be fed into neural network directly

because of too much redundant data. Therefore the

preprocessing of writing pressure data is

indispensable and crucial for system. There are three

steps for the data preprocessing, and in each step

different treatment is implemented and different data

are obtained.

Registration Based on NN

Recognition Based on NN

First, after three times of signature given by

registrant, the writing pressure data set with more

than 1000 data are transformed into normal data

with number of about 300 by Moving Average

method. During this process, the individual feature

of pressure data is extracted and data scale is

compressed greatly.

Second, validity check is implemented on

normal data obtained in first step basing on two

factors: correlation coefficient

and statistical

distance

The correlation coefficient

between Data1 and

Data2 is calculated as following equation:

)2(

)(1

))1((

(Data1(s))

)2()1(

))(2)(1(

dAveS

sData

dAveS

dAvedAve

sDatasData

∗−

∑

∗−

∑

∗

⎯ (1)

Here,

and are average of

Data1 set and Data2 set respectively with assuming

that there are S elements in each data set. The value

)1(dAve

)2(dAve

varies between –1 and 1. More near 1 R is,

more closely two data correlate.

Statistical distance (Euclidean distance) between

Data1 and Data2 is calculated as following equation:

∑

−=

sDatasDataD

))(2)(1(

⎯(2)

The deviation between each two of three writing

pressure data are calculated according to their

correlation coefficient and statistical distance, if

deviation is more than upper limit, the

corresponding signature data are treated as abnormal

and registrant is asked for re-signature.

In this system, only three writing pressure data

set as enforce data are not enough for purposive

pattern learning of Neuro-template, furthermore the

inhibit data for non-purposive pattern are also

necessary. So not only purposive simulative data

(enforce data) but also non-purposive simulative

data (inhibit data) are constructed based on the

detected writing pressure data and the corresponding

normal data are then constructed.

Last, In order to obtain 50 slab data for Neuro-

template from 300 normal data, mask data are made

by comparison on distribution of standard error and

average value between enforce normal data and

inhibit data, then 50 slab data for the Neuro-template

are obtained by filtering middle data with the mask

data.

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

158

hidden laye

4 ALGORITHM

purposive pattern

For traditional MFNN, when recruiting new pattern,

the whole neural network will have to be

restructured and retrained, that would lead to

expensive cost of computation and time. In order to

eliminate the restriction on the number of registrants

in the recognition system and simplify the

recruitment of new pattern, neural template

matching method showed as Figure 2 was proposed

and applied into this individual recognition system.

In this method, one neural template corresponds to

one pattern (registrant). When new registration is

completed, a new neural template is constructed and

trained for this new registrant, other neuro-templates

generated previously will remain untouched unless

mis-outputs are caused by signature of new

registration in these templates. That is to say instead

of training all of templates, only the training for new

template and the retraining for the existing templates

in which the mis-recognition occurred are involved

in the procedure of new registration. Therefore the

cost of calculation and time for new registration is

greatly decreased and registration procedure is

simplified.

non-purposive pattern

In the neural template matching method, each

neuro-template is constructed by three layers

feedforward neural network with structure of 50×35

×2 which is shown in Figure 3, one of two output

layer units corresponds to purposive pattern

(purposive registrant) and the other to non-purposive

one (non-purposive registrant). The function of each

neuro-template is evaluating whether the input data

is the purposive pattern of this template or not rather

than deciding which pattern the input data is from all

of templates. Recognizing the pattern of input

signature from existing templates is completed by

template matching procedure. The experiment

results showed that this strategy was effective for

individual recognition system.

In the previous research work, sigmoid function

was employed as activation function of neuro-

template in the individual recognition system. The

simulation experiments have shown that the system

was effective on classifying the signatures that

belong to known patterns (the average recognition

rate was over 95%), however rejection performance

for counterfeit signature (always unknown pattern)

was not satisfying. This problem caused by the

limitation of the MFNNs: the pattern space is

divided up into several areas corresponding to the

patterns that have been learned in a specific case and

the networks may be trained to have high accuracy

in classifying patterns for a set of known categories,

so it can successfully recognize the signature of

registrant whose pattern have been learned. But for

any pattern which is out of known categories, it is

also very likely to be classified as one of known

categories by MFNNs. That leads to the poor

rejection capability for counterfeit signatures.

To improve the rejection performance of the

system for counterfeit signature, Gaussian function,

which is one of RBF, was proposed as activation

function of neural unites in this paper. Instead of

dividing up the pattern space as MFNNs do, neuro-

template with Gaussian function learns the pattern

probability density. Therefore, when an out-of-

category pattern is evaluated, it is likely to be

recognized as an unknown category by neuro-

template with Gaussian function. Then the rejection

performance of the system for counterfeit signature

is expected to be improved. The expression of

Gaussian activation function of hidden and output

layers neuron is described as following equation:

)2exp()(

xxxf −−= ⎯ (3)

Where

and

are the centre vector and the

width parameter of Gaussian function respectively.

Both of them have direct effect on the convergence

and recognition capability of the individual

recognition system.

In the neuro-template training procedure,

improved Back Propagation (BP) is employed to

modify the weights between neighbour layer

neurons and corresponding expression was described

as following:

Figure 3: Structure of neuro-template

output layer Input layer

urpos

attern A

Template A

Template B

Template N

Check

value

Slab

value o

pen

pressure

non-purpos

attern A

Figure 2: Construction of Neuro-template Matching Mechod

IMPROVEMENT ON THE INDIVIDUAL RECOGNITION SYSTEM WITH WRITING PRESSURE BASED ON RBF

159

)2()1(

)(

)( −∆⋅+−∆⋅+

∂

⋅−=∆ tWtW

βαη

⎯ (4)

Where

are the learning rate, momentum

coefficient, and oscillation coefficient respectively.

The momentum item contributes to accelerating

convergence and the oscillation item has function of

escaping local minimum

. The neuro-templates are

trained with learning data set till the cost functions

meet the requirements of minimal error.

Figure 4: Registrant Signature and Corresponding

Writing Pressure data

5 EXPERIMENTS

In the research work of this paper, a series of

simulation experiments were made on both of

individual recognition system with Gaussian

function and that with sigmoid function for the

purpose of comparison. The conditions for neuro-

template learning are listed in table 1.

Table 1: Conditions of neuro-template learning

Generally the registration and recognition of the

system are operated on-line, but in order to ensure

the identical writing pressure data for neuro-

templates with different activation function, the

registrant signs on the signature sheet under the

same condition and off-line registration and

evaluation is implemented with data file in the

experiment. One of registrant signature and

corresponding writing pressure data are shown as

Figure 4.

To demonstrate average performance of the

system for any template, three registrants (labelled

as A, B, C) were selected randomly, 53 signature

samples were extracted from each purposive

registrant as purposive signatures and any three of

these samples were employed for neuro-template

learning and the left for evaluation. To evaluate

counterfeit rejection performance of the system, 90

samples were extracted as counterfeit signatures

from any five people (except A, B, C) with 18

samples per person for each registrants and the three

groups of five people who give the counterfeit

signatures were different each other.

The evaluation items in the experiments are

recognition ratio and rejection ratio. Recognition

ratio presents the percentage of signatures being

successfully recognized to all purposive signatures

(50). Rejection ratio for counterfeit signatures

presents the percentage of signatures being

successfully rejected as unknown pattern to all

counterfeit signatures (90). Assuming having three

registrants (A, B and C) in the system, taking

registrant A as purposive registrant for example,

signature (purposive signatures of A or counterfeit

signatures of A) is fed to each template and

evaluated by respective template. All possible results

of the system were illustrated in Table 2. Note that

output of the system will be one of these three

registrants or none of them (briefly N).

Maximum iteration number 1000

Learning rate 0.05

Momentum coefficient 0.95

Oscillation coefficient -0.1

Error threshold 0.0001

Value of (only for Gaussian)

0.3~1.0

Table 2: All Possible Results of Experiments

Evaluated signature

Purposive

signature

Counterfeit

signature

Source of signature A

Anyone except

A, B and C

Ideal Results A N

Success A N

N A

Possible

results

system

Failure

B or C B or C

In the previous table, the failure cases in which

purposive signature of A is mis-recognized as B or C

and the counterfeit signature for A is mis-recognized

as B or C can easily be overcome by our system

because of significant difference among the

signatures of the respective registrant.

Based on the previous condition and the sampled

data (including purposive and counterfeit), neuro-

templates with Gaussian Function and that with

sigmoid function are trained with the sequence of A,

B, C and evaluated respectively, and the experiment

results are presented in Table 3 and Table 4.

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

160

Table 3: Recognition ratio for purposive signatures

(ideal value: 100%)

Registrant Sigmoid Gaussian Difference

A 96% 96% -0%

B 98% 96% -2%

C 96% 92% -4%

Average 96.67% 94.67% -2%

Table 4: Rejection ratio for counterfeit signatures

(ideal value: 100%)

Registrant Sigmoid Gaussian Difference

A 64.44% 82.22% +17.78%

B 85.56% 90.0% +4.44%

C 71.33% 88.89 % +17.56%

Average 73.77% 87.04% +13.27%

The item of ‘Difference’ in above tables is for

easy discrimination on the performance change

between two systems. The minus sign (-) indicated

deterioration of the performance of system with

Gaussian function compared with that with sigmoid

function and the positive sign meant the

improvement. According to table 3 and table 4,

though the recognition capability of system with

Gaussian function decreased slightly (average –2%),

its rejection capability for counterfeit signatures was

improved greatly with average value of 13.27%.

This result suggested that the neuro-template with

Gaussian function proposed in this paper tended to

be effective on improving the rejection capability of

system for the counterfeit signatures and at the same

time keeping the recognition capability for

purposive registrant signatures satisfied.

As mentioned in section four, the recruitment of

new registrant in the system will lead to retraining

on the exiting templates in which mis-outputs are

caused by signatures for new register. That means

the neuro-template in the system is likely to be

influenced by the templates registered later. To

investigate the mutual influence of different neuro-

templates, more experiments were conducted on the

system with different number and different sequence

of templates. Registrant A and B in previous

experiments were selected for the purpose of

convenience. In these experiments, when registrant

B registered after A, retraining on template A was

induced by recruitment of B, however no retraining

on template B in the case of registrant A registering

after B. Note that any templates except A and B was

not included in this system.

First, the performance of template A with

sigmoid function and that with Gaussian function

under different situation were investigated and the

corresponding results were demonstrated in Table 5

and Table 6 respectively.

Table 5: Performance of template A with sigmoid function

under different situation (ideal value: 100%)

Performance

Order

of registrants

Recognition Rejection

A only 96% 57.78%

A, B 96% 64.44%

B, A 96% 57.58 %

Table 6: Performance of template A with Gaussian

function under different situation (ideal value: 100%)

Performance

Order

of registrants

Recognition Rejection

A only 96% 68.89%

A, B 96% 82.22%

B, A 96% 68.89 %

Table 7 and Table 8 showed the performance of

template B with sigmoid function and that with

Gaussian function under different cases respectively.

Table 7: Performance of template B with sigmoid function

under different situation (ideal value: 100%)

Performance

Order

of registrants

Recognition Rejection

B only 98% 85.56%

B, A 98% 85.56%

A, B 98% 85.56 %

Table 8: Performance of Template B with Gaussian

Function under Different Situation (Ideal Value: 100%)

Performance

Order

of registrant

Recognition Rejection

B only 96% 90.0%

B, A 96% 90.0%

A, B 96% 90.0 %

As can be seen from above tables, the

recognition performance of both template A and B

with different activation function were not

IMPROVEMENT ON THE INDIVIDUAL RECOGNITION SYSTEM WITH WRITING PRESSURE BASED ON RBF

161

influenced by the register sequence, that indicated

that the recognition performance of the system was

not affected by increase of templates. While for

counterfeit rejection capability, there were two cases:

one case was that new register led to retraining on

the templates registered previously (Table 5 and

Table 6); the other case was that no retraining was

resulted by new register (Table 7 and Table 8).

According to Table 5 and Table 6, in both two

systems with different activation function, the

rejection capability of template A was enhanced by

new register of B and this improvement was

especially remarkable with employment of Gaussian

function. From Table 7 and Table 8, the rejection

performance of template B kept untouched because

no retraining was caused by recruitment of template

A. these experiment results showed that recruitment

of new templates was helpful to decreasing the

possibility of mis-recognition on counterfeit

signatures in one template.

The results in Table 3 and 4 involved the

influence among templates. Excluding the influence

among templates, the rejection performance of

templates with different activation function was

investigated and corresponding results were listed in

Table 9. Note that in each case, only one registrant

was involved in the system.

Table 9: Rejection performance of template with

different activation function (ideal value: 100%)

Function

Registrant

Sigmoid Gaussian Difference

A only 57.78% 68.89% +11.11%

B only 85.56% 90.0% +4.44%

C only 73.77% 87.04 % +13.27%

Average 72.37% 81.97% +9.60%

It can be seen that even without the help of

favourable influence of templates, Gaussian function

was still effective on improving the rejection

capability of template.

During experiments, the width parameter σ was

found to have great influence on both recognition

capability and counterfeit rejection capability of the

system with Gaussian function. We are engaging in

developing automatic optimisation method of σ for

each registrant’s neuro-template.

6 CONCLUSION

In this paper, both of the construction and algorithm

of the individual recognition system with writing

pressure were firstly described. Then, Gaussian

function, which is one of RBF, was proposed as

activation function of neuro-template to improve the

rejection capability of the system for counterfeit

signatures. Furthermore the influence among neuro-

templates was investigated in this paper. The

experiments results suggested that the influence

among templates was favourable for rejection

capability of the system, and more importantly the

experiment results shown that Gaussian function

combined with neuro-template was seemed to be

very effective in improving rejection performance of

the system for counterfeit signatures on premise of

ensuring the recognition performance satisfied.

REFERENCES

Takeda, F. and Omatu, S., 1995. High speed paper

currency recognition by neural networks. IEEE

Transaction on Neural Networks, Vol.6, No.1,pp.73-77.

Takeda, F. and Omatu, S., 1996. A neuro-system

technology for bank note recognition, Proceedings of

the Japan/USA Symposium on Flexible Automation,

Boston,USA, Vol.2, pp.1511-1516.

Takeda, F. and Omatu S., 1998, Neural network systems

Technique and applications in paper currency

recognition neural networks systems, Technique and

Applications, Vo1.5, Chapter 4, pp.133-160,

ACADEMIC Press.

Takeda, F. Omatu, S., and Nishikege, 1996, A neuro-

recognition technology for paper currency using

optimized masks by GA and its hardware, Proceedings

of the International Conference on Information System

Analysis and Synthesis, Orlando, USA, pp147-152.

Takeda, F., Nishikage, T. and Matsumoto, 1998, Y.,

Characteristic extraction of paper currency using

symmetrical masks optimized by GA and neuro-

recognition of multi-national paper currency,

Proceedings of IEEE World Congress on

Computational Intelligence, Alaska, USA, Vol.1,

pp.634-639.

Fnakubo, M., 1991, Pattern Recognition, Japan: Kyouritsu

Press,pp15-25.

Abhijit S. Pandya, and Robert B. Macy, 1996, Pattern

Recognition with Neural Networks in C++, Chapter 5,

pp147-151, IEEE Press.

Kikuti, T., 1989, Introduction to neuro-computer, pp.35-

67, Japan: Omu Press.

Nagao, M., 1993, Pattern Recognition, Japan: CORONA

publishing CO. LTD., pp.92-161.

ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

162