View
6
Download
0
Category
Preview:
Citation preview
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 (lecho203@gmail.com). Mahasak Ketcham is with Department of Information Technology,
Management Faculty of Information Technology, King Mongkut’s University
of Technology North Bangkok Bangkok, Thailand
(maoquee@hotmail.com).
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
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 40
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
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 41
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
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 42
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)
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 43
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
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 44
Recommended