Beruflich Dokumente
Kultur Dokumente
Abstract- Handwriting recognition ranks among the issue, Arabic handwritten script recognition has been deeply
highest and the most triumphant applications in the pattern investigated for years in the handwriting recognition
recognition domain. Despite being a developed field, many domain. Different algorithms like Support Vector Machine
enquiries are still needed and still represent a defiance mainly (SVM), Multi-Layer Perceptron (MLP), Artificial Neural
for the Arabic Handwritten Script (AHS). Recently, more Networks (ANN), Hidden Markov Model (HMM) and
regard has been given to Support Vector Machines (SVM) Convolutional Neural Networks (CNN), etc., have been
classifier for script recognition. Nevertheless, it has not been
exploited by researchers who have attained a lot of
put in application yet to the handwritten Arabic field if
favorable results. Performance and accuracy have been
compared with the other methods like ANN, CNN, RNN and
demonstrated by the systems in a large field of applications.
HMM. SVMs for AHS recognition is examined in this paper.
Alike, these systems have acquired a great triumph in
Handcrafted feature is handled as input by the suggested
method and gets going with a supervised learning algorithm.
Optical Character Recognition (OCR) in Latin and Asian
We chose the Multi-class Support Vector Machine with an
languages [5, 6]. A lot of researchers have utilized Hidden
RBF kernel and we tested it on Handwritten Arabic
Markov Models for offline Arabic handwritten and printed
Characters Database (HACDB) as well. It was proven that the recognition. It is noticeable that these models have proved
proposed method was effective thanks to the simulation results. their forcefulness [7, 8, 9, 10]. Y et, some elastic models like
We compared the well-functioning of this method with HMM are still considered a great hinder in Arabic
character recognition reliabilities coming from state-of-the-art handwriting recognition because of due to the huge
Arabic OCR which resulted in commendatory outcomes. variability and distortions of patterns.
Keywords-SVM; arabic handwritten recognition; handcraft CNN is made up of a number of convolutional and sub
feature; HACDB sampling layers which are non-mandatory accompanied by
Fully Connected Layers. The Fully Connected Layers are
the same as to the layers in a standard Multi-Layer
I. INTRODUCTION
Perceptron. Yet, MLP shows two restrictions in
During the two last decades and within the context of classification tasks: to begin with, there is not any
signal processing and pattern recognition, offline and online theoretical relationship between both the classification task
data classification has been granted great attention. It has and the MLP structure. Next, MLP derive hyper-planes
been widely applied to a variety of research fields naming as separation surfaces, in feature representation space, which
examples vision recognition task [I, 2], Automatic Speech are not optimal in terms of margin between the examples of
Recognition (ASR) [3] and analysis of two different classes.
Electroencephalograms (EEGs) signals [4] and
classification. It plays a progressively important perspective On the other hand, being considered as one of the
strongest and most vigorous algorithm in machine learning
in the Human-Computer Interaction field. Besides,
computer feigning and machine simulations of human invented by Vapnik [ I I] Support vector machines (SVM),
reading/writing/gesture have been given great importance have turned into being a popular approach used in various
for the last years. domains [12, 13, 14], like pattern classification, recognition,
and image processing.
The writing recognition belongs to the wider field that is
the pattern recognition. Mostly, because of being a detying Using SVM, Byun and Lee [IS] who handed out survey
applications of pattern recognition reviewed seven
The organization of the rest of the paper is the Determining an optimal separating hyper-plane
following. Tn Section IT, we describe the basic concepts (equation 2) or deciding a surface by adopting an
behind Support Vector Machine (SVM). Our experimental unconventional technique founded on mapping the sample
study using this model is next presented in the Section m. points into a high-dimensional feature space and categorized
Finally, Section TV discusses the overall results of the model utilizing a nonlinear transformation <D, even with non
and presents some concluding remarks. linearly separable data is the fundamental concept of SVM .
The most favorable hyper-plane is acquired by solving a
quadratic programming problem subject to regularization
parameters (equation 4). This transformation has been
! executed by kernel functions stating as examples linear,
Pre-processing Radial Basis Function, Sigmoid and Polynomial kernel
types which are defined as;
L' 1 1 1 1 1 y= 1 / ()2
[I 1 1 1 I I
I I 1 ..... 1 I I With d, o, 1' and yare parameters that will be determinate
empirically.
SVM
1 L-. f(x) = W T <D(x) + b (2)
I Classifier I -
Where W E: R" , b E: R and <D(x) is a feature map.
.1=1
1 1 1
552 2015 15th International Conference on Intelligent Systems DeSign and Applications (ISDA)
by these approaches once the number of classes exceeds 2 training set of 5.280 images and a test set of l.320 images.
classes, called multi-class SVM. Work in [19, 20] provides Details of the class for each shape are presented in Table 1.
more detailed information.
Optimal choice of kernels is crucial for determining how
<Y"
to carry out the projection into higher-dimensional space. In saad
this paper, because of the linearly inseparable of the feature
'J
space, an alteration was put in an application by mapping
Lam Alif
the input data (Xi, Yi) into a higher dimensional feature space
utilizing a nonlinear operator <!lex). Hence we can define,
the optimal hyper-plane (Ho) as the following:
Lam mim ha
(5)
..,jl J,\. J).. j.) Jl .J\ j Jj j,t
Where K(Xi' x) = exp(-Yllxi - x112) is the kernel function Alif lam ha
based on a Radial Basis Function (RBF), and sgn(.) is the
sign function. This classifier model named RBF kernel
SVM is used to do classification for Arabic Handwritten
Text. Fig. 4. Samples from the HACDB database written by 10 different
writers.
y
Arabic
Shape class Arabic Script Shape class
Script
Aeen (t)
t- I
J- 34
Class 2 .J.. 2
L 35
0
3 Lam_Alif (':!) 36
(a)
x
t: 4 )f 37
5 38
l::J. Class 3 Lam .Teem
y l::J. / .....Q.... 6 -.d 39
(..."J)
l::J. l::J.
l::J. , ...... """
\
l::J. f13(X}
fd x } l::J. Alif (l) 7 Lam_Mem (...J) ...(, 40
l::J../
"- /
l 8
Lam Mem .Te
e m () 41
c....N 12 ----0- 45
x
(b) ----"- Mem .Teem
13
( ) 46
Fig. 3. Approach's of Support Vector Machine; (a) one-versus-all
method,(b) one-versus-one method Daal(o) 14 Noon (0)
U 47
...I...- 15 U 48
m. EXPERIMENTAL RESULTS AND DISCUSSION Faa (w) C9 16 Raa (.)) J 49
We performed our experimental studies using SVM
model for recognizing offline Arabic character. We
-..9 17 Y 50
2015 15th International Conference on Intelligent Systems DeSign and Applications (ISDA) 553
27 60 <I> denotes the orientation of a Gabor filter,
o"x, O" represent the standard deviation in x- and y
t- 28 Taa (.b)
61 y
direction, respectively.
../:.--- 29 62
Kaaf (.)
30 Waao (J)
63
ll(u,,; A,", "" "
,
J c exp - 2, '
{ ( ( J ,, : c; - + ";(F,)' )}
i:,
31
)- 64
Laam (J)
J 32 Yaa (c.
05 65 (7)
J 33
0- 66 Where FJ , Fz and C are:
F; = ucosl/l+vsinl/l
B. Experiments setting
[n order to evaluate the efficiency of the suggested F2 = -usinrjJ + vcosrjJ
system based on SVM classifier, we investigated its
And C is a constant.
performance for training and recognizing characters of
HACDB database. To effectively train the model on more
data so as to perfectly handle the variability of handwriting, C. Experiments using SVM model
the size of the training set is extended ten times by the Tn our experiments, we investigated the performance of
elastic deformation technique suggested by Simard et al the SVM model for training and recognizing Arabic
[22]. Technical execution specifics of the selected system characters. For the setting architecture, we need to
are given in the next subdivision. determine about SVM classifier essentially two parameters
of the RBF kernel; Gamma (y) and C.
Pre-processing: The pre-processing phase where the
database utilized in this experimental study does not We selected the pertinent parameters for SVM model on
require to be normalized. It is noticeable that a few the basis of empirical tests. We inaugurated an experimental
fundamental pre-processing activities are vitally to be study so that we can assess the suggested model. Our
conducted throughout the database development. selection of parameters is on the basis of the criterion of the
error classification rate on the train dataset. We also used the
Parameters setting: For the setting architecture, we must one-versus-all method with 66-way for the multi-class RBF
determine the optimal kernel parameter and penalty kernel SVM as it provides a more valuable discrimination
parameter of SVM. The value of the tradeoff parameter than the linear kernel. Yet, a less parameter than the
C and a parameter in SVM are chosen empirically. polynomial kernel was utilized.
Feature extraction: Features are the information Tn this work, Gabor features is the input to the SVM
extracted from the image of a word or character, and they classifier. Nevertheless, the SVM assumes the handcraft
are used to build classifiers for classification. The features in the role of a feature vector to recognize the
challenge is to determine which features are more unrevealed handwriting text. And consequently by
suitable for classification. Tn this paper, Gabor filter [16] analyzing the error classification rate on the Arabic
was used in this experiment. handwritten character classification task.
Gabor filters are used widely in image processing for its We indicated that in our experiments, LTBSVM [27] tool
good characteristics as; the optimal joint spatial/spatial is considered to set up multi-class SVM classifier with RBF
frequency localization and the fitness to simulate the kernel. LTBSVM implements the SMO-based algorithm. We
receptive fields of simple cells in the visual cortex [23, 24]. utilized a soft-margin SVM marked by the first parameter C
They depict an image in different levels of frequency. which controls the margin from a class boundary and the 0"
Therefore, various features are extracted, depending on parameter which controls the variance in the RBF kernel.
everyone's response of filter and its frequency. Tn [25] The parameter C is the tradeoff between training set and the
Daugman incited the use of Gabor filters as a representation error in the separation.
for receptive fields and their exploit in recognition systems.
The choice of parameters (C, 0") was proved to be
Two-dimensional Gabor filter is a complex sinusoidally experimentally effective by applying a grid search with the
modulated by a Gaussian function with the response in the 5-fold stratified cross validation method. We synthesized the
spatial domain (equation 6) and in frequency domain most favorable usefulness of principal parameters obtained
(equation 7), defined as follows [16, 26]: after the tests on the training Arabic handwritten text
h(x,y;:t,,<Jx,<Jy)
1
exp
{I [
- - -2 +-2
R\ 2 R ]} x exp
[ 27fR\ ]
.
/--
database HACDB as follows: a = 0.05 and C = 30.
The suggested system founded on RBF kernel SVM
27f<Jx<Jy 2 <Jx <Jy :t
already described, provides an error classification rate
(6) (ECR) of 11.23% on the testing dataset with 66 classes (see
fig. 5) while SVM with polynomial kernel attained an error
Where R1 , R2 ,Ie, <1>, ax and ay are explain as: classification rate of 29.18%.
R1 = xcosr/J+ ysinl/l
R2 = -xsinl/l+ ycosl/l
554 2015 15th International Conference on Intelligent Systems DeSign and Applications (ISDA)
to increase the recognition rate.
2015 15th International Conference on Intelligent Systems DeSign and Applications (ISDA) 555
[ 1 6] J. Chen, H. Cao, R. Prasad, A Bhardwaj, P. Natarajan, "Gabor
features for offline arabic handwriting recognition," In Proceedings
of the 9th IAPR International Workshop on Document Analysis
Systems (DAS),pp. 53-58,20 10.
[ 1 7] C. Cortes, V. Vapnik, "Support vector networks," Machine Learning,
vol. 20,pp. 273-297, 1995.
[ 18] C. Burges, "A tutorial on support vector machines for pattern
recognition," Data Mining Knowledge Discovery,vol. 2(2), pp. 1 2 1 -
167, 1998.
[ 1 9] B. Scholkopf, A Smola, "Learning with Kernels," MIT Press,
Cambridge, USA,2002.
[20] J. Shawe-Taylor, N. Cristianini, "Kernels Methods for Pattern
Analysis," Cambridge University Press, Cambridge,2004.
[2 1] A Lawgali,M. Angelova, A Bouridane, "HACDB: Handwritten
Arabic characters database for automatic character recognition,"
EUropean Workshop on Visual Information Processing (EUVIP), pp.
255-259,20 13.
[22] P. Simard, D. Steinkraus, 1. C. Platt, "Best Practices for
Convolutional Neural Networks Applied to Visual Document
Analysis," International Conference on Document Analysis and
Recognition (ICDAR),pp. 958-962,2003.
[23] 1.G. Daugman, "Uncertainty relation for resolution in space, spatial
frequency, and orientation optimized by two dimensional visual
cortical fIlters," J. Opt. Soc. Am.,vol. 2 (7), 1985.
[24] AK. Jain, F. Farrokhnia, "Unsupervised texture segmentation using
Gabor fIlters," Pattern Recog.,vol. 24( 12),pp. 1 167- 1 186, 199 1 .
[25] 1 . Daugman, "Complete discrete 2 D Gabor transforms b y neural
networks for image analysis and compression," IEEE Transactions
on Acoustics, Speech and Signal Proc., vol. 36(7), pp. 1 169- 1 179,
1988.
[26] X. Wang,X. Ding,and C. Liu,"Gabor fIlter-based feature extraction
for character recognition," Pattern Recognition,vol. 38,pp. 369-379,
2005.
[27] c.c. Chang, C.J. Lin, "LTBSVM: A Library for Support Vector
Machines," Software Available at
http://www.csie.ntu.edu.tw/-cjlin/libsvm. 200I.
[28] A Lawgali, A Bouridane, M. Angelova and Z. Ghassemlooy,
"Handwritten Arabic character recognition: Which feature extraction
method?," International Journal of Advanced Science and
Technology,vol. 34,pp. 1 -8,20II.
[29] M. Elleuch,N. Tagougui,and M. Kherallah, "Towards Unsupervised
Learning for Arabic Handwritten Recognition Using Deep
Architectures," Neural Information Processing - 22nd International
Conference, ICONIP 20 15, Istanbul, Turkey, part. (I), pp. 363-372,
20 15.
[30] Applied Media Analysis, Arabic-Handwritten-I.O,
http://appliedmediaanalysis.comlDatasets.htm. 2007.
[3 1] M. Pechwitz,S. S. Maddouri, V. Margner,N. Ellouze,and H. Amiri,
"IFNIENIT database of handwritten Arabic words," In Colloque
International Francophone sur l'Ecrit et Ie Document (CIFED), pp.
127- 136,2002.
556 2015 15th International Conference on Intelligent Systems DeSign and Applications (ISDA)