Handwriting Thai Signature Recognition System...

Abstract— In this paper, we proposed to develop Handwriting

Thai Signature Recognition System (HTSRS) based on multilayer

perceptron and radial basis network . In order to achieve this

objective, a combination of back-propagation and generalized

regression neural networks is proposed. The first stage is

implemented by using two parallel back propagation and

generalized regression neural network in order to reduce training

time and make a system adaptable. The second stage is implemented

by using the generalized regression neural network for final decision.

The system is composed of two principal phases : signature feature

extraction and classification. For the extraction step, global and grid

features of all signature images are first extracted. Then this features

are feed into the first and second stage neural network classifier for

training. The second group is for testing and unknown signature.

The experimental results show that the accuracy of the proposed

system performs rate is 90.00%.

Keywords— Signature, Recognition, Neural Network, Multilayer

Perceptron, Back Propagation Algorithm, Generalized Regression

Network.

I. INTRODUCTION

OMPUTER systems have a role to society. Whether it is

bordered by the government. Or private broadly The

system has been applied to many applications on the computer

. It is bordered to the public , business, industry and finance .

Applications above Including the development of computer

systems to identify , manage transactions, also called biometric

system which bordered bringing mathematical methods . Or

statistical analysis used to identify an individual or verify a

person automatically. The physical attributes of different

individuals such as fingerprint, hand geometry, retina pattern,

iris pattern, facial detection. Or use behavioral characteristics

Naruemol Chumuang

is with the Department of Information Technology

Faculty of Information Technology, King Mongkut’s University of

Technology North Bangkok Bangkok, Thailand ([email protected]). Mahasak Ketcham is with Department of Information Technology,

Management Faculty of Information Technology, King Mongkut’s University

of Technology North Bangkok Bangkok, Thailand

([email protected]).

of an individual for example voice, gait recognition, signature

recognition [6],[8],[9],[10],[12],[13],[16],[17].

At present time the trends to use of tablet PCs are likely to

rise and operating one is easy to use. However, the language

input method for most tablet PCs is a virtual keyboard; a set of

virtual on screen buttons that allow typing by touching.

Because there is no actual button on virtual keyboard then

operating is difficult. On-line optical character recognition

(OCR) is a solution to make the input method more

convenient. OCR is applied to earn machine-understandable

text from our handwriting on tablet screen. For example,

WritePad [1] and PhatPad [2] are on-line OCR software

systems for iPad and android tablets. But this software still

does not support Thai

From the study of research on Thai handwriting such as that

of online handwriting signature recognition based on wavelet

energy feature matching [3] in which presents new algorithm

of classification. They do not use Thai signature in the

experiment. The formal language in Thai with its unique 44

alphabets 32 vowels and 4 tones. In general Thai words have 3

writing levels; body, upper vowels and lower vowels. More

difficult to classifies handwriting Thai alphabets with head and

without head [4],[5],[7],[15],[16].

For the above reasons mentioned already . Thus creating a

challenge for us. How to design of handwritten Thai signatures

. We have presented a new algorithm by describing the Feature

Extraction in two stages, the first stage is a Global feature to

split the alphabet and separate level of the signature. The

second stage , we use Grid feature to determine the density of

black pixel in the signature image. Next we proposed

methodology with MLP[11] and RBN [14] to recognition and

decision. The last we show experimental results.

II. PREPROCESSING

For this research We have used a sample signature image

format “.jpg” of individuals known signatures of 10 persons.

Given 60 signatures for one person see in Figure 1. Then

stored the image data signatures of 600 images into database.

Applying signature image to the system need to process the

data below.

Handwriting Thai Signature Recognition

System based on Multilayer Perceptron

and Radial Basis Network

Naruemol Chumuang, and Mahasak Ketcham

C

Int'l Conference on Advanced Computational Technologies & Creative Media (ICACTCM’2014) Aug. 14-15, 2014 Pattaya (Thailand)

http://dx.doi.org/10.15242/IIE.E0814539 39

1) Resize the image to a standard. In this research were used

in size 100x150 pixels.

2) Skeleton analysis of the signature. Due to sign sometimes

may used the different pen.

บริุนทร์

จรัญญา

กาญจนา

คุณาวฒิุ

เมตตา

นฤมล

พศิิษฐ ์

Fig. 1 Example handwriting Thai signature images

III. FEATURE FROM SIGNATURE IMAGES

Features of the images used in this research is a binary.

Which were divided into two groups according to the

characteristics of specific features.

A. Global Feature

1) Signature height. The aspect ratio Signature calculated

using height to the width of the image [2].

2) Image area. An area with the images. This means that the

number of these black pixels are the actual size of the

line that separates without background.

3) Pure width. Net width of the signature image.

4) Pure height. Net height of the signature image.

5) Base line shift. Amenable to the vertical center of the left.

And the right to determine the direction of the image.

6) Horizontal and vertical center of the signature image are

calculated using the formulas in Eq.1 [4]

],[

],[

maxmax

maxmax

11

11

yxb

yxbxC

y

y

x

x

y

y

x

x

x

],[

],[

maxmax

maxmax

11

11

yxb

yxbyC

y

y

x

x

x

x

y

y

y

(1)

7) The maximum black pixels in the vertical by the image in

each column. Then find the largest value of the number

of black pixels of all columns 150 columns.

8) The maximum black pixels on the horizontal by the

image in each row. Then find the largest value of the

number of rows of black pixels all 100 rows.

9) The density histogram h(i) presents the density for each

row i The density distribution describes the probability of

having h foreground pixels.

10) The density histograms h(j) presents the density for each

column j. See in Figure 2.

Fig 2 Peaks determined by horizontal strokes are evident in the

density histogram (upper figure). In the density probability

distribution (lower graphic)

B. Grid Feature

In this stage we make block sizes brace onto the signature

image. To analyze the density of the foreground of the image

by the number of black pixel found in each block of the block,

for example blocks see in Figure 3. Which we used in this

experiment. Our analysis of the appropriateness of the size of

the image signature HTSRS in which the image is determined

standard size to 100X150 pixels and thus determine size 4x5,

5x10 and 10x10 blocks. For density of black pixel in each

block show in Figure 4.

Fig 3 Example block size 10x10 pixel

Calculate the number of black pixels in each block can be use

Eq.2

],[)(max

1

j

j

j

j yxbiblock

(2)

Body

Upper vowel

Lower vowel

0 0

1 1

0 0

0 0

0 0

0 0

0 0

0 0

0 0

0 0

0 1 1 0 0 0 0 0 0 0

0 0 0 1 1 0 0 0 0 0

0 0 0 0 1 1 0 0 0 0

0 0 0 0 0 0 1 1 0 0

0 0 0 0 0 0 1 0 0 0

0 0 1 1 0 0 0 0

1 1 1 1 1 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

0 0



where i is sequence block, b is black pixel, x and y are pair

of coordinates in each block and j is position of coordinates.

Fig 4 Example grid block size (a) 4x5, (b) 5x10 and (c) 10x10 blocks

IV. FEATURE EXTRACTION

For the analysis of image features. We divided into two

groups: a group of features Global Feature and Grid Feature.

Take the two groups into neural networks by a one-on-one

networking. The database are stored the common features of a

person can show in Figure 5. Which this storage need to

collected data of all 10 persons into the database. So, they had

feature of every persons are stored in Glb_Data 600 images. In

same storage for grid feature. Stored all of them in a vector

table. This feature storage grid must contain the information

about individuals all 10 signatures so Grd_Data each to have

all 600 vectors show in Figure 6.

Fig. 5. Data storage global feature of one person with 60

signatures

Fig. 6 Data storage grid feature of one person with 60 signatures

V. MULTILAYER PERCEPTRON AND RADIAL BASIS NETWORK

For HTSRS the training neural network to classify the

signature are divided into two phases, which are described in

more detail below.

A. First stage classification

First, we used of Multilayer perceptron by using Back

propagation Algorithm and Radial basis network using

Generalize regression network. Both networks are neural

networks with training then we must divided data into two

groups: the first group data used for training and the other data

used for testing, see as Figure 7.

Fig. 7 The average of the Global features using in training with 10

persons all 600 vectors

All of global features is stored in Glb_Data of 600 vectors

and then calculated for average of all individual signatures.

The average vector is store in a database which will be training

T1. For the same grid feature. Stored in the table Grd_Data

vector is used to calculate the average vector of all individual

as well. Then store the average vector calculated in the form of

100 –dimensional 10 vectors in T2 for training set.

When the two group of all feature to separate for training

(T1), testing (T2) then the next step into the neural network.

We can show framework in Figure 8.

Fig 8 HTSRS framework of the signature features into a neural

network

When importing data into a single feature MLP network

with one then results are achieved through the training and

testing the resulting in “A1” with 10-dimensional vector table

size of 600 vectors, and “Y1” is the result of 100-dimensional

N

Start

Global Feature

Glb_Data SigImg>=60

End

Y

N

Start

Grid Feature

Grd_Data SigImg>=60

End

Y

Glb_Data

1

60

120

.

.

.

.

.

600

Avg

Avg

Avg

Avg

Avg

Avg

.

.

.

.

.

T11

T12

T13

T110

Signature Image

Global Feature Grid Feature

GRN MLP

GRN

N

First stage Classification

Second stage Classification



vector 600 vector grid next two features put into the RBF

neural network again.

This procedures will achieve results through training and

testing in Table A2 is 10-dimensional vector 600 vector and

output Y2 is 100-dimensional vector table of 600 vectors.

When the results of all four networks in first stage then we

calculated the Euclidean distance between the results obtained

with the input data. These are pair of them for minimum

distance.

Euclidean distance between A1 and Glb_Data

Euclidean distance between Y1 and Grd_Data

Euclidean distance between A2 and Glb_Data

Euclidean distance between Y2 and Grd_Data

Fig. 9 Minimum of Euclidean distance between A1 and Glb_Data

B. Second stage classification

After completing the training, the first step, we store the

position of the Minimum distance of each feature compared to

the results through training (A1, Y1, A2 and Y2), which makes

the Minimum distance 1, Minimum distance 2, Minimum

distance 3 and Minimum. distance 4 from the reference

position of the distance into second stage classification. For

the training to use it as well. We bring vectors of average value

from the individual person by two feature vectors are the same

as for training and decision in this stage. We can show

framework of this stage in Figure 10.

Fig. 10 Process in second stage classific

When the results came out we will bring it to reference

check whether correct or not.

VI. EXPERIMENTAL AND RESULT

Signature image in this research import from scanner 600

image. The experimental procedure can be summarized as

follows.

1) Convert signature images to binary and determine to

standard size in 100x150 pixel.

2) Skeletonization on images.

3) Analysis on images with two groups of feature are Global

feature and Grid feature.

4) The results from the analysis of the two group features

are classified in the first step with Back propagation and

Generalize regression network in one-on-one network.

5) Gather data from the classification in the first stage to

make that decision signature Image is true or false with

Radial Basis Network.

We adopt a Global feature by measuring the distance with

Euclidean Distance it’s shown graph distance’s signature

image that is as close to "กาญจนา" shown in Figure11. This

results is correct. However, this approach may give inaccurate

results. When the distance is measured over a similar

signature. We need more accuracy with MLP and GRN learn

and decide.

Fig 11 Distance between target and unknown image with

Euclidian distance

After global features and grid features are analysis. We have

vectors of data in database, then bring two features to first

stage classification with MLP and GRN. The results learning

with MLP quite well can shown in Figure 12. In the next stage

we use the Euclidean distance for measure between the target

Glb_Data Output

A1

Euclidean

distance

Minimun Distance

1

Start

Unknown Image

Load Data T1,T2

for Target in Training

RBN

Output NN_Out

from Network

End



with output from the database of all 10 persons see in Figure

13.

Fig. 12 Euclidean distance between target and output with global

feature of "บุรินทร์" (a), “สมเจตน์” (b) and “นฤมล” (c)

Fig. 13 Euclidean distance between target and output with grid

feature of "บุรินทร์" (d), “นาวิน” (e) and “คุณาวุฒิ” (f)

Result in First stage classification is the vector of the two

features to Generalize Regression Network (GRN). Results

showed graph that appears consistent are in rather up to 90%.

In experimental we tested 30 times by dividing the training set

66% and the test set 33%. The result in second stage

classification is final decision which experimental results

shown in Table 1.

(c)

(b)

(a)

(c)

(a)

(b)

(c)



TABLE I

RESULT IN SECOND STAGE CLASSIFICATION

NN No. Ranking Result Target Accuracy

2 23 บริุนทร์ บริุนทร์ 98.23

1 78 คณุาวฒิุ คณุาวฒิุ 77.76

1 314 นาวิน เมตตา 87.06

2 239 นฤมล นฤมล 90.89

1 279 จรัญญา จรัญญา 93.04

3 309 นาวิน นาวิน 91.05

1 387 กาญจนา กาญจนา 94.27

2 429 สมเจตน์ สมเจตน์ 91.25

2 523 พิสิษฐ์ พิสิษฐ์ 90.05

4 564 สรุศกัดิ์ สรุศกัดิ์ 91.02

The minimum Euclidean distance is selected out from 4

neural network in first stage. We will determine the order of

the network. The results came out with an accuracy of 90%

VII. CONCLUSION

This paper proposed a novel HTSRS algorithm, we used

two groups of features. Global feature such as image area, pure

height, pure width, center of Thai signature in the vertical,

center of signature, horizontal, etc. and Grid feature with three

block size then take all the features into a neural network for

training for signature images from 10 persons with each 60

images. By calculating the average of the signature of each

person from all database for represent.

When through the process of learning with Multilayer

perceptron (MLP) trained using back propagation algorithm

and then fed to the neural network is another layer Radial basis

Network (RBF), which select the Generalize Regression

Network (GRN) for use in decision.

VIII. FUTURE RESEARCH

In the second stage of HTSRS. We used data from a

database into a neural network. If the results of the

classification in the first stage into the second stage directly.

In another we will use other algorithm for training may be one

of the ways to make them better . It is best to develop and

bring to trial in the next study.

REFERENCES

[1] WritePad Software, Retrieved Feb 21,2014, from

https://itunes.apple.com/us/app/writepad-for- ipad/id363618389?mt=8

[2] PhatPad software, Retrive Feb 21,2014, from

http ://www.phatware.com/index.php?q=page/phatpad/ipad#

[3] Chayut Wiwatcharakoses and Karn Patanukhom. “Two-Stage

Recognition for Printed Thai and English Characters Using Nearest

Neighbor and Support Vector Machine,” IEEE International

Conference on Signal Image Technology & Internet-Based Systems.,

pp. 71-78. 2013.

[4] Credit Karnchanapusakij, Phattharasuda Suwannakat and Waroonorn

Rakprasertsuk, “Online Handwriting Thai Character Recognition,”

IEEE International Conference on Computer Graphics : Imaging and

Visualization., pp. 323-328. 2009.

[5] Ekawat Chaowicharat, Nick Cercone and Kanlaya Naruedomkul,

“Novel Curve Signature and a Combination Method for Thai On-Line

Handwriting Character Recognition” Proceeding of the Twenty-Sixth

International Florida Artificial Intelligence Research Society

Conference, pp. 196201. 2012.

[6] H.Baltzakis and N. Papamarkos, “A new signature verification

technique based on a two-stage neural network classifier,” Pergamon,

pp 95-103. 2001.C. J. Kaufman, Rocky Mountain Research Lab.,

Boulder, CO, private communication, May 1995.

[7] Jianbin Zheng and Guangxi Zhu. “On-Line Handwriting Signature

Recognition Based on Wavelet Energy Feature Matching.” Intelligent

Control and Automation, pp. 9885-9888. 2006.M. Young, The

Techincal Writers Handbook. Mill Valley, CA: University Science,

1989.

[8] J.-P. Drouhard, R. Sabourin and M. Godbout. “A Neural Network

Approach to Off-Line Signature Verification Using Directional PDF,”

Pattern Recognition, pp. 415-424. September 1996.

http://dx.doi.org/10.1016/0031-3203(95)00092-5

[9] Kai Huang and Hong Yan. “Off-Line Signature Based on Geometric

Feature Extraction and Neural Network Classification,” Pattern

Recognition, pp. 9-17. January 1997.

http://dx.doi.org/10.1016/S0031-3203(96)00063-5

[10] Luan L. Lee, Toby Berger and Erez Aviezer, “Reliable On-Line Human

Signature Verification System,” IEEE Transaction on Pattern Analysis

and Machine Intelligence., pp. 643-647. June 1996.

http://dx.doi.org/10.1109/34.506415

[11] Martin T. Hagan., Howard B. Demuth. And Mark Beale. Neural

Network Design. Boston : PWS Publishing Company, c1995.G. R.

Faulhaber, “Design of service systems with priority reservation,” in

Conf. Rec. 1995 IEEE Int. Conf. Communications, pp. 3–8.

[12] Radmilo M. Bozinovic and Sargur N. Srihari. “Off-Line Cursive Script

Word Recognition,” IEEE Transaction on Pattern Analysis and

Machine Intelligence, pp. 68-83. January 1989.

http://dx.doi.org/10.1109/34.23114

[13] Reena Bejaj and Santanu Chaudhury. “Signature Verification using

Multiple Neural classifiers,” Pattern Recognition, pp.1-7. January 1997.

http://dx.doi.org/10.1016/S0031-3203(96)00059-3

[14] Simon Haykin. Neural Networks A Comprehensive Foundation. New

York : Macmillan College Publishing Company, c1994.

[15] Thanaruk Theeramunkong and Chainat Wongtapan. “Off-line isolated

handwritten Thai OCR using island-based projection with n-gram model

and hidden Markov models,” Elesvier Information Processing and

Management, pp. 139-160. 2005.

[16] V.E. Ramesh, M. Narasimha Murty. “Off-Line Signature Verification

Using Genetically Optimized Weighted Feature,” Pattern Recognition,

pp. 217-233. February 1999.

http://dx.doi.org/10.1016/S0031-3203(98)00141-1

[17] Yingyong Qi and Bobby R.Hunt, “Signature Verification Using Global

and Grid Feature,” Pattern Recognition, pp. 1621-1629. December

1994.

http://dx.doi.org/10.1016/0031-3203(94)90081-7



http://dx.doi.org/10.1016/0031-3203(95)00092-5

http://dx.doi.org/10.1016/S0031-3203(96)00063-5

http://dx.doi.org/10.1109/34.506415

http://dx.doi.org/10.1109/34.23114

http://dx.doi.org/10.1016/S0031-3203(96)00059-3

http://dx.doi.org/10.1016/S0031-3203(98)00141-1

http://dx.doi.org/10.1016/0031-3203(94)90081-7

Handwriting Thai Signature Recognition System...

Documents

Transcript of Handwriting Thai Signature Recognition System...