A neuro-fuzzy inference engine for Farsi numeral characters recognition

Abstract

Character recognition of Farsi and Arabic texts as an open and demanding problem needs to encounter sophisticated specifications of the characters such as their shapes, continuity, dots and also, different fonts. Utilizing fuzzy set theory as a tolerant approach toward uncertainty and vagueness and artificial neural networks as a machine learning method in this paper, we propose a neuro-fuzzy inference engine to recognize the Farsi numeral characters. This engine takes holistic approach of character recognition through the comparison of the unknown character’s features with the features of the existing characters that itself is characterized through Mamdani inference engine on fuzzy rules which is largely enhanced with a multi layer perceptron neural network’s learning on features of the different fonts’ characters which leads to more comprehensive recognition of Farsi numeral characters in the proposed system. Having applied this novel engine on a dataset of unknown numeral characters consisted of 33 different Farsi fonts, it yielded more accurate results than the corresponding researches. The recognition rates of unknown numeral characters are greater than 97% except for Farsi character 4, so as the proposed schema could not score a better result than 95% for this numeral character which implies its recognition is still in need of more enhancement.

Full text

Page 1

Published in Expert Systems with Applications, 2010. 37(9): p. 6327-6337

A Neuro-fuzzy Inference Engine
for Farsi Numeral Characters Recognition

Gholam Ali Montazer*
Information Technology Dept., School of Engineering, Tarbiat Modares University,
Tehran, Iran
Hamed Qahri Saremi
Information Systems Dept., DeGroote School of Business, McMaster University,
Hamilton, ON, Canada
Vahid Khatibi
Information Technology Dept., School of Engineering, Tarbiat Modares University,
Tehran, Iran


Abstract
Character recognition of Farsi and Arabic texts as an open and demanding problem needs to
encounter sophisticated specifications of the characters such as their shapes, continuity, dots and
also, different fonts. Utilizing Fuzzy Set Theory as a tolerant approach toward uncertainty and
vagueness and Artificial Neural Networks as a machine learning method in this paper, we
propose a neuro-fuzzy inference engine to recognize the Farsi numeral characters. This engine
takes holistic approach of character recognition through the comparison of the unknown
character’s features with the features of the existing characters that itself is characterized through
Mamdani inference engine on fuzzy rules which is largely enhanced with a multi layer perceptron
neural network’s learning on features of the different fonts’ characters which leads to more
comprehensive recognition of Farsi numeral characters in the proposed system. Having applied
this novel engine on a dataset of unknown numeral characters consisted of 33 different Farsi
fonts, it yielded more accurate results than the corresponding researches. The recognition rates of
unknown numeral characters are greater than 97% except for Farsi character 4, so as the proposed
schema could not score a better result than 95% for this numeral character which implies its
recognition is still in need of more enhancement.

Keywords: Character recognition, Fuzzy rules, Multi layer perceptron neural network, Neuro-
fuzzy Inference Engine, Farsi Numbers

1. Introduction

Recent developments in electronic publishing has brought up the need to transformation
of paper-based and printed documents to digital formats, hence, text recognition has been
highlighted further (Broumandnia & Shanbehzadeh, 2007). Also, extensive applications
of character recognition in recognizing the numbers in bank checks (Morita, Lethelier,
Yacoubi, Bortolozzi, & Sabourin, 2000), the car plate numbers (Pan, Ye, & Zhang,
2005), etc. have caused varieties of new systems such as optical character recognition
(OCR) system to develop (AI-Alawi, 2004). In character recognition literature, several

* Corresponding Author. Tel.: +98(21)82883990, Fax: +98(21)82883990.
Email Addresses: montazer@modares.ac.ir (GH.A. Montazer), QahriSH@mcmaster.ca (H. Qahri Saremi),
khatibi@modares.ac.ir (V. Khatibi).

Page 2

Published in Expert Systems with Applications, 2010. 37(9): p. 6327-6337

2
different methods such as decision tree(Amin, 2000; Nagabhushan & Pai, 1999), fuzzy
set theory (Cheok, Jian, & Chng, 2008; Chi, Suters, & Yan, 1996; Hanmandlua, Mohanb,
Chakrabortyc, & Goyald, 2003; Patil & Sontakke, 2007), artificial neural networks
(Goltsev & Rachkovskij, 2005; Lauer, Suen, & Bloch, 2007), support vector machines
(Camastra, 2007; Malon, Uchida, & Suzuki, 2008), hidden Markov model (Amor &
Amara, 2005; Bunke, Roth, & Schukat-Talamazzini, 1995; Khorsheed, 2003;
Theeramunkong & Wongtapan, 2005) or any other significant hybridation of these
methods (Nagabhushan & Pai, 1999; Ping & Lihui, 2002) are exploited.
Document image analysis has been accentuated as one of the main research issues
for about three decades and several methods for recognition of Indian(Al-Omari & Al-
Jarrah, 2004; Pal & Chaudhuri, 2004), Thai(Theeramunkong & Wongtapan, 2005),
Chinese (Hung, Luk, Yeung, Chung, & Shu, 2005; Li, Tan, Ding, & Liu, 2004; Su,
Zhang, Guan, & Huang, 2009; Zhao, Chi, Shi, & Yan, 2003), and Japanese (Itoh, 1995;
Kimura, Wakabayashi, Tsuruoka, & Miyake, 1997; Srihari, Hong, & Srikantan, 1997)
documents have been proposed. Also, in spite of numerous researches on recognition of
Farsi and Arabic texts (AI-Alawi, 2004; Al-Muhtaseb, Mahmoud, & Qahwaji, 2008;
Azmi & Kabir, 2001; Broumandnia & Shanbehzadeh, 2007; Dehghan, Faez, Ahmadi, &
Shridhar, 2001; Mozaffari, Faez, Märgner, & El-Abed, 2008; Solimanpour, Sadri, &
Suen, 2006; Ziaratban, Faez, & Faradji, 2007), the researchers’ inability in reaching a
consensus has made these researches to be continued, so as it is still an open and
demanding problem, resulting in new methods of Farsi and Arabic texts recognition.
Considering the fact that more than 30% of the world populations talking in the various
types of Farsi, Arabic and Urdu dialects use Farsi alphabets and numbers in their written
texts, so analyzing and recognizing the Farsi texts are highly crucial (Baghshah,
Shouraki, & Kasaei, 2005). On the other hand, the specifications of characters in respect
of their shapes, continuity, dots and different fonts in these languages make their
recognition more complicated in comparison to other ones in the world (Broumandnia &
Shanbehzadeh, 2007).
The character recognition researches can be classified to two various types as:
analytical approach and holistic approach (Mitchell & Yan, 2004). The former comprises
the pattern verification, momentum measurement and mathematical conversions. In this
approach, the pattern is presented as a multi-dimensional vector in a so-called “pattern
space” and the purpose will be segmenting the pattern space to some regions called
“class”. On the other hand, in the holistic approach, the character is considered as a
unique and invisible entity. In this approach, the aim is recognizing the character using its
features. This approach comprises two phases; in the first phase, the character features
are identified and during the second one, the character is recognized by comparing its
features with the features of the existing characters (Ho & Leedham, 2001).
The analytical approach itself is consisted of two other approaches: statistical and
deterministic. In statistical approach, the probability density functions are estimated for
the patterns distribution in each class, using parametric and non-parametric methods, and
consequently the unrecognized samples are classified using such criteria as the minimum
error and the minimum risk. On the other hand, in deterministic approach, no specific
statistical feature is considered for the patterns and instead, the distance and discriminant
functions are usually utilized. Some classification methods such as minimum distance,
template matching and K-nearest neighborhood could be mentioned in this regard.

Page 3

Published in Expert Systems with Applications, 2010. 37(9): p. 6327-6337

3
Overall, the analytical approach mostly depends on the use of statistical methods such as
Markov models, principle component analysis and hierarchical cluster analysis (Bunke,
Roth, & Schukat-Talamazzini, 1995; Fang & Alouani, 2002).
Not considering the alphabets and numbers structure in their identification is the
prime weakness of the statistical methods. Conversely, the holistic approach, comprising
structure and syntax analysis, identifies the new patterns by the help of the previous
structures of the alphabets and numbers as well as the features of these structures (Fang
& Alouani, 2002). In this approach, each pattern is considered as a set of sub-patterns or
principal elements and their interrelations (Dimauro, Impedovo, Pirlo, & Salzo., 1997).
Pattern definition in this approach is similar to sentence definition in a language
grammar. Namely, the principal elements are like alphabets and a set of syntactic rules
helps to define each pattern of a class through various combinations of those elements.
Accordingly, the geometrical characteristic of the pattern is one of the essential features
to be considered in this approach and artificial neural networks (ANNs) are sensible
instruments for this purpose. Comparing the strengths and weaknesses of two discussed
approaches, analytical and holistic, we choose the holistic approach in this paper. As
discussed before, this approach determines the identity of the characters by comparing
the features of their new samples with those of the old ones.
The main objective of this paper is to recognize the printed Farsi numeral characters,
considering 33 different Farsi fonts through proposing a neuro-fuzzy inference engine.
Considering a distinct class for each number and having particular structures in shape,
some numbers incorporate uncertainty in their assignment to a specific class. Regarding
this fact, we exploit fuzzy set theory to quantify the significant existing uncertainty. The
proposed engine first extracts the numbers’ features through the holistic approach which
compares the unknown character’s features with the features of the existing characters that itself
is characterized through Mamdani inference engine on fuzzy rules which is largely enhanced with
a multi layer perceptron neural Network’s learning on features of the different fonts’ characters
which leads to more comprehensive recognition of Farsi numeral characters in the proposed
system. In fact, we have utilized a neuro-fuzzy system to recognize the printed Farsi
numeral characters, considering 33 different Farsi fonts.
This paper is organized as follows. In Section 2, the structure analysis methodology
is discussed. Then, we propose a neuro-fuzzy system and its usage results on 33 different
Farsi fonts are analyzed. At last, the whole discussion is concluded, comparing the new
results with the previous studies and analyzing the improvements for Farsi numeral
characters recognition.

2. Structural Analysis Methodology

Doing the preprocessing work, each number goes through a three stepped process ; noise
elimination, thinning and feature extraction, successively (Broumandnia &
Shanbehzadeh, 2007; Mitchell & Yan, 2004). Noise in number recognition is a small blob
(in the size of a pixel or a dot) which is resulted from various hand-writing styles or small
dirt around the hand-written number. Utilizing printed numbers as the inputs to our
system in this paper, we sensibly assume that our numbers are free from noise and so,
two steps from preprocessing work can be omitted. Thinning is a necessary and important
step for a correct stroke extraction. After thinning, feature points are extracted (Mitchell
& Yan, 2004).

Page 4

Published in Expert Systems with Applications, 2010. 37(9): p. 6327-6337

4
In general, the feature points comprise three types of points as: terminal points (T),
branch points (B) and cross points (C). Utilizing these feature points, each pattern can be
segmented to different regions. In number recognition process, the numbers can be
segmented to several regions, considering their feature points. Then, each region can be
described by a fuzzy set, considering their fuzzy characteristics. This method can
simplify and accurate the recognition process. Considering numeral characters as the
main target to identify, we focus on their recognition from now on.
Numbers’ geometrical characteristics are essential factors helping in their
recognition and classification. These characteristics can be fused with feature extraction
techniques, making the numbers classification based on their shapes possible. For any
feature extraction process, the data should be pre-processed beforehand. For gathering the
generic information and features about the data, their geometric characteristics and
criteria should be extracted. The geometric characteristics include such criteria as pixel
density, line, point and etc. We have utilized numbers’ structure analysis method for
recognizing the numbers. As discussed earlier, in numbers’ structure analysis, we need
the information about three regions of a number, namely: terminal, branch and cross
points (Fang & Alouani, 2002). Also, any number is made of three essential components:
straight line, loop and orientation. Considering this approach as well as segment approach
in structure analysis, we can define a fuzzy status for each number including orientation,
curvature and number’s weight parameters. We throw more light on these parameters in
following.

2.1 Orientation
Orientation is a parameter of arcs and lines, as described below:

• Arcs’ Orientation
Arc orientation parameter is indicated using an angle between the principal and
vertical axes and is a number between 0 and 360 degrees. By arc in this paper, we
mean the regions of Farsi numbers of 2, 3, 4, 5 and 6 which are written using some
arcs. Considering all possible orientation statuses in numbers, we define four fuzzy
numbers labeled as up, down, right and left, indicated in Fig. 1 by U, D, R and L,
respectively. Determining the arc orientation parameter, the discussed angle, we
connect the two end points of the arcs using a straight line and the resulting angle
between the line and the vertical axis is labeled θ . Finally, we measure the
resulting angle between the perpendicular line on the connecting line and the
horizon, as the intended parameter.
As presented in Fig. 1, if the orientation angle is between 0 and 15 degrees, the
orientation state belongs to “Right” fuzzy set with membership degree of 1. On the
other hand, if the aforementioned angle is between 165 and 195 degrees, the
orientation state belongs to “Left” fuzzy set with membership degree of 1. In order
to cover the marginal angles, like 45 degrees, which can be assigned to, for
example both right and up fuzzy sets simultaneously, we utilize both fuzzy sets’
parameters to describe them. Concerning the fact that the precise measurement of
the orientation is not possible, using fuzzy concept, we can reach sensible
approximation in our recognition work.

Page 5

Published in Expert Systems with Applications, 2010. 37(9): p. 6327-6337

5

Fig 1. Fuzzy sets presenting various arcs’ orientation statuses


On this basis, we can present a arc fuzzy set as:
,,,,()(
RDLUR Arc
μμμμμθμμ==
As an example, discussed above, for an orientation angle between 0 and 15 degrees,
we will have:
) 0 , 0 , 0 , 0 , 1 (
=
Arc
μ

Determining the lines’ statuses, we utilize four fuzzy sets of: Horizontal, Negative
slope, Vertical and Positive slope (Montazer & Jafari, 2006). We have lines in all
Farsi numbers except Farsi number 5. These lines are presented using the four
aforementioned fuzzy sets, as depicted in Fig. 2.


)
(1)

Fig. 2. Fuzzy sets presenting various lines’ orientation statuses