Hangul Recognition Using Support Vector Machine

International Journal of Computer Applications Technology and Research
Volume 7–Issue 07, 281-285, 2018, ISSN:-2319–8656
www.ijcat.com 281
Hangul Recognition Using Support Vector Machine
Rahmatina Hidayati
Department of Electrical Engineering
University of Brawijaya
Malang, East Java, Indonesia
Moechammad Sarosa
State Polytechnic of Malang
Panca Mudjirahardjo
University of Brawijaya
Abstract: The recognition of Hangul Image is more difficult compared with that of Latin. It could be recognized from the structural
arrangement. Hangul is arranged from two dimensions while Latin is only from the left to the right. The current research creates a
system to convert Hangul image into Latin text in order to use it as a learning material on reading Hangul. In general, image
recognition system is divided into three steps. The first step is preprocessing, which includes binarization, segmentation through
connected component-labeling method, and thinning with Zhang Suen to decrease some pattern information. The second is receiving
the feature from every single image, whose identification process is done through chain code method. The third is recognizing the
process using Support Vector Machine (SVM) with some kernels. It works through letter image and Hangul word recognition. It
consists of 34 letters, each of which has 15 different patterns. The whole patterns are 510, divided into 3 data scenarios. The highest
result achieved is 94,7% using SVM kernel polynomial and radial basis function. The level of recognition result is influenced by many
trained data. Whilst the recognition process of Hangul word applies to the type 2 Hangul word with 6 different patterns. The difference
of these patterns appears from the change of the font type. The chosen fonts for data training are such as Batang, Dotum, Gaeul,
Gulim, Malgun Gothic. Arial Unicode MS is used to test the data. The lowest accuracy is achieved through the use of SVM kernel
radial basis function, which is 69%. The same result, 72 %, is given by the SVM kernel linear and polynomial.
Keywords: Support Vector Machine; SVM; Kernel Polynomial; Kernel Linear; Kernel Radial Basis Function; Hangul
1. INTRODUCTION
Optical Character Recognition (OCR) is a character
introduction system with images input. It contains texts that
would be converted to the edited versions[1]. The work of
OCR system depends on the kind of processed text.
Generally, the text is divided into three categories. They are
written, printed, and typed text[2].
Some researches on OCR System have been conducted.
One of methods ever used is Support Vector Machine (SVM).
A kind of character which had ever been searched by using
SVM is Hindi number, which is known as Numeral Kanada,
upper case and lower case alphabet A-Z [2,3,4]. The SVM
method is used with different data, Korean characters known
as Hangul.
Hangul recognition is more difficult compared with Latin
due to its complicated arrangement. Hangul is arranged from
2 dimensions (both left and right side), while Latin is arranged
from left to the right [5].
A Research on Hangul recognition has ever been
conducted, where the writer applies the Stochastic
Relationship Modeling to the recognition process of Hangul
syllable writing. The output of the research is Hangul syllable
texts[6].
So far, the research on Hangul recognition is conducted
with Hangul texts output. The current research will improve
the image conversion of Hangul with Latin text output. The
image of Hangul converted into Latin text can be used as a
learning material on how to read Hangul.
OCR system, in general, is divided into three steps. They
are preprocessing, feature extraction, and recognition. Pre-
process includes three stages: binarization to change the
grayscale image into black white; segmentation, which is
processed through connected component labeling, to separate
the input into individual word; and thinning to decrease some
information pattern (thin Line) in order to be easily analyzed
[7]. The research will employ algorithm Zhang Suen, which
works faster than the other thinning algorithms[8].
The next step, after pre-process, is feature extraction. It
has an important role on recognition process. It works through
generating the basic component of the image called
features[9]. The feature extraction used in the current research
is chain code. The last process is recognition, using the SVM
method with some kernels (linear, polynomial, and radial
basis function).
2. METHODOLOGY
Generally, OCR system is divided into three main steps.
They are preprocessing (binarization, segmentation, and
thinning), feature extraction in which in this research uses
chain code, and recognition by applying Support Vector
Machine (SVM) method. Figure 1 shows the general process
of Hangul recognition.
Hangul Image
Figure 1. Block diagram of Hangul recognition
The input image used in the current research is the letter
and Hangul word. The letter consists of 21 vowels and 13
consonants, shown in Figure 2 with Latin letter. Each letter
has 15 different forms. The whole data are amounted to 510.
Preprocessing
Feature Extraction
Hangul Recognition

www.ijcat.com 282
Figure 2. The Letters of Hangul and Latin[10]
Meanwhile for word, there are 6 ways on how to arrange
the letter of Hangul into word. The first type is shown in
Figure 3[10]. The discussion focuses on data type 2.
‘
Figure 3. 6 the ways to arrange the letter of Hangul[10]
Meanwhile, the example of type 2 Hangul word is shown
in Figure 4. Each word consists of 6 different forms. They are
achieved by changing the word font. The used fonts are Arial,
Batang, Dotum, Gaeul, Gulim, and Malgun Gothic.
Figure 4. The example of Hangul word
2.1 Preprocessing
The preprocessing includes 3 steps. First, the binarization
or thresholding, is implemented to change the grayscale image
become black white. The process of thresholding will produce
the binary image, the image which has two steps of grayish
(black and white). Generally, the process of floating the
grayscale image to produce the biner image are as follows[11]:
(1)
With g (x,y) is the binary image from grayscale image f
(x,y) and T assert the percentage of threshold.
Second, thinning is used to decrease some information to
a pattern becomes thin line in order to be easy to analyzed[7].
In this case, it will be applied the Zhang Suen algorithm
which has faster performance compare with the other thinning
algorithm[8]. To process thinning algorithm is shown in the
Figure 6. This algorithm uses the pixel 3x3 and 8 degrees as
in the Figure 5. P1 is a pixel that will be checked, if it fulfills
the fixed condition, so the pixel will be deleted. The
conditions are as follows[12]:
(a) 2 ≤ B(P1) ≤ 6 (2)
(b) A(P1) = 1 (3)
(c) P2 x P4 x P6 = 0, and (4)
(d) P4 x P6 x P8 = 0 (5)
(e) P2 x P4 x P8 = 0, and (6)
(f) P2 x P6 x P8 = 0 (7)
Figure 5. The pixel 3x3 with 8 degrees

www.ijcat.com 283
Start
P2 x P4 x P8 = 0 &
P2 x P6 x P8 = 0 ?
End
2 ≤ B(P1) ≤ 6 &
A(P1) = 1 ?
Delete pixel
Delete pixel
P2 x P4 x P6 = 0 &
P4 x P6 x P8 = 0 ?
Delete pixel
Matrix of
thinning result
Tidak
Tidak
Tidak
Ya
Ya
Ya
Scan pixel
Figure 6. The Diagram algorithm of Zhang Suen
2.2 Feature Extraction
The next step after preprocessing is feature extraction,
which has an important role on recognition. This process will
generate the necessary component of an image called
features[9]. The used feature extraction is chain code which
functions as the direction search. The direction usually uses
the following regulation Figure 7.
3 2 1
4 0
5 6 7
Figure 7. The direction of sequence code with 8 degrees[13]
The sequence code is made through checking the
direction of a pixel connected to the other pixels with 8
degrees. Each direction has different number. The pointed
sign shows the first pixel. It is a place where the next steps
will be fixed[13]. Figure 8 shows letter B with its sequence
code.
Figure 8. The direction of the sequence code in letter B
In order to be processed into SVM, the feature extraction
must have the same amount of features. Therefore,
normalization needs to be done. It also aims to decrease the
amount of feature which reoccurs. The normalization process
for chain code can be done with the following pattern[12]:
(8)
Note:
N = the amount of feature wanted
= the amount of feature normalized
= the amount of all letters normalized
2.3 Support Vector Machine (SVM)
The Support Vector Machine (SVM) is found in 1992 by
Boser, Guyon, and Vapnik. The basic concept is the
combination of computation theories introduced in previous
years, such as the margin hyperplane and kernel. The concept
of SVM can be seen as an effort to look for the best
hyperplane to separate two classes in the input grade.
Figure 9 shows some data of the two class member (+1
and -1), where (a) shows some alternative separated line
(Discrimination Boundaries), and (b) shows the best
hyperplane. The best hyperplane can be found by measuring
the distance between the hyperplane and the nearer data
known as margin [13]. The data in the margin is called
support vector[14].
Figure 9. (a) The alternative separated field, (b) The best
separated field with the biggest margin[14]

www.ijcat.com 284
Practically, the real data does not only consist of two
classes and they cannot be separated linearly. To solve this, an
approach needs to be conducted so that SVM is able to
classify the data input with N-Dimension or multi class, one
of which is one against one[14]. Besides, to arrange the data
input into higher dimension feature space, a function called
kernel is needed. Table 1 shows some kernels used in SVM
[15], and Figure 10 shows the process of recognition with
SVM.
Table 1. Kernel function in SVM[16]
Kernel Fungsi
Linear K(x,y) = x,y
Polynomial K(x,y) = (ᵞx.y+r)d
Radial Basis Function K(x,y) = exp(-ᵞ||x-y||2
, ᵞ>0
ϒ, r, and d are kernel parameter
Start
Testing data
image
Multiclass SVM recognition
Recognition result
Kernel accuracy
End
Chain code features
Figure 10. The diagram of Hangul recognition line with
Multiclass Support Vector Machine
3. EXPERIMENT RESULT
Algorithm line Zhang Suen in Figure 6 shows that slight
result is produced such as in Figure 11 (b).
Figure 11. (a) Binary image, (b) Letter CH after thinning
After achieving the slight feature, the system takes it in
accordance with the direction search in Figure 7. It generates
the features of CH letter as follow:
0;0;0;0;0;0;0;0;0;0;0;5;6;6;6;6;6;6;6;6;6;6;6;6;6;6;6;6;6;5;4;4;
4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;4;0;0;0;0;0;0;
0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;5;6;6;6;6;6;
6;6;6;6;6;6;6;6;6;6;6;6;5;. . . . ;CH;
If the feature is normalized according to the pattern (8), it
has the same amount as the minimal feature from input image.
Then, it will generate the following normalization feature:
0;0;0;0;0;0;0;0;5;6;6;6;6;6;6;6;6;6;6;6;6;6;5;4;4;4;4;4;4;4;4;4;
4;4;4;4;4;4;4;4;4;4;4;4;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;
0;0;0;0;5;6;6;6;6;6;6;6;6;6;6;6;6;6;5;. . . . ;CH;
After getting the feature, system will do the recognition
process. The letter recognition consists of 34 letters; each of
them formulates 15 forms. So, they are 540 in total. The data
scenario is shown in Table 2.
Table 2. Data scenario of Hangul letter recognition
Data Scenario Training Data Testing Data
DS1 170 (data A) 170 (data C)
DS2 340 (data A+B) 170 (data C)
DS3 510 (data A+B+C) 170 (data C)
Data testing process for DS1 and DS2 uses new data which
have not been tested before while that for DS3 uses the tested
ones.
Accuracy result of Hangul letter recognition for every
kernel is shown in Table 3.
Table 3. Accuracy result of Hangul letter recognition
Kernel SVM DS1 DS2 DS3
Linear 88,82% 94,11% 100%
Polynomial 89,41% 94,7% 100%
RBF 89,41% 94,7% 100%
Recognition mistake often occurs to the letter which has
the same form, such as 우 (U) with 유(YU) or the other way
around. Another letter which looks similar is 의(UI) with 위
(WI). The system output of Hangul letter is shown in Figure
12.
Figure 12. The result of Hangul letter recognition
In Hangul word recognition, researcher uses the form 2,
where the training data are 195 and the testing data are 39.
The accuracy for every kernel is shown in Table 4.(a) (b)

www.ijcat.com 285
Table 4. Accuracy of Hangul word recognition
Kernel SVM Accuracy
Linear 72%
Polynomial 72%
RBF 69%
Output from the recognition of Hangul word type 2 is
shown in Figure 13. Recognition mistake occurs in the word
which has almost similar form, such as 쵸 (CHYO) with 초
(CHO).
Figure 13. The recognition result of Hangul word type 2
4. CONCLUSIONS
The research concludes that the more the trained data, the
higher the degree of accuracy. However, it needs to be
reexamined until the fixed number of data which give the
highest accuracy with SVM method is found.
In letter recognition process, kernel polynomial and RBF
achieve the highest accuracy of 94,7% in data scenario 2
(DS2). On the other hand, linear process gives the lowest
accuracy, 88,82%, in letter recognition, and RBF in Hangul
word with 69%.
5. FUTURE WORK
The future research might employ the feature from the
image with another method, while in the recognition process,
the researcher can use SVM method with different kernel.
Hangul recognition into Latin form may also be improved by
adding the meaning of the trained word.
6. REFERENCES
[1] Seethalakshmi R., Sreeranjani T.R., & Balachandar T.
2005. Optical Character Recognition for Printed Tamil
Text Using Unicode. Journal of Zhejiang University
SCIENCE (2005), 1297-1305.
[2] Singh, D., Aamir Khan, M. & Bansal, A. 2015. An
Application of SVM in Character Recognition with
Chain Code. International Conference on
Communication, Control and Intelligent Systems (CCIS).
IEEE (2015), 167-171.
[3] Rajashekararadhya, S. V. & Ranjan, P. V. 2009. Support
Vector Machine based Handwritten Numeral
Recognition of Kannada Script. IEEE International
Advance Computing Conference, 381-386.
[4] Tran, D. C., Franco, P. & Orgier, J.M. 2010. Accented
Handwritten Character Recognition Using SVM –
Application to French. IEEE International Conference on
Frontiers in Handwriting Recognition, 65-71.
[5] Kyung-Won, K. & Jin H., Kim. 2003. Handwritten Hangul
Character Recognition with Hierarchical Stochastic
Character Representation. Proceedings of the Seventh
International Conference on Document Analysis and
Recognition.
[6] Kyung-Won, K. & Jin H., Kim. 2003. Handwritten Hangul
Character Recognition with Hierarchical Stochastic
Character Representation. IEEE Transactions On
Pattern Analysis And Machine Intelligence, vol. 25, no.
9, 1185-1196.
[7] Lam, L., Seong-whan, L., & Suen, C. Y. 1992. Thinning
Methodologies A Comprehensive Survey. IEEE
Transactions on Pattern Analysis and Machine
Intelligence, vol. 14, no. 9, 869-885.
[8] Haseena, M. H. F & Clara, A. R. 2017. A Review on an
Efficient Iterative Thinning Algoritm. International
Journal Research in Sciencee, Engineering and
Technology, vol. 6, no. 11, 541-548.
[9] Nasien, D., Haron, H. & Yuhaniz, S. S. 2010. Support
Vector Machine (SVM) For English Handwritten
Character Recognation. IEEE Second International
Conference on Computer Engineering and Applications,
249-252.
[10] Ju, S. & Shin, J. 2013. Cursive Style Korean Handwriting
Synthesis based on Shape Analysis with Minimized
Input Data. IEEE International Conference on High
Performance Computing and Communications &
International Conference on Embedded and Ubiquitous
Computing, 2231-2236.
[11] Putra, D. 2010. Pengolahan Citra Digital. Yogyakarta:
Penerbit Andi.
[12] Zhang, T. Y. & Suen, C. Y. 1984. A Fast Parallel
Algorithm for Thinning Digital Patterns.
Communication of the ACM, vol. 27, no. 3, 236-239.
[13] Sutoyo. Teori Pengolahan Citra Digital. Penerbit Andi.
2009.
[14] Vijaykumar, S. & Wu, S. 1999. Sequential Support
Vector Classifier and Regression. SOCO’99.
http://homepages.inf.ed.ac.uk.
[15] Chih-Wei H, Chih-Chung C, Chih-Jen L. A Practical
Guide to Support Vector Machine. Department of
Computer Science National Taiwan University, Taipei.
May 2016.
[16] Fadel, S., Ghoniemy, S., Abdallah, M., Sorra, H. A.,
Shour, A., Ansary, A. 2016. Investigating the Effect of
Different Kernel Functions on the Performance of SVM
for Recocnizing Arabic Characters. International
Journal Computer Science and Aplications, vol. 7, no.
1, 446-450.

Hangul Recognition Using Support Vector Machine

More Related Content

What's hot (19)

Similar to Hangul Recognition Using Support Vector Machine (20)

More from Editor IJCATR (20)

Recently uploaded (20)

Hangul Recognition Using Support Vector Machine