See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271828246

The State of the Art Handwritten Recognition of Arabic Script Using Simplified
Fuzzy ARTMAP and Hidden Markov Models

Article · August 2014

CITATIONS READS

10 299

3 authors:

Dr. Yusuf Perwej Shaikh Abdul Hannan

Professor Goel Institute of Technology & Management (GITM), Dr. Babasaheb Ambedkar Marathwada University
107 PUBLICATIONS 1,403 CITATIONS 62 PUBLICATIONS 590 CITATIONS

SEE PROFILE SEE PROFILE

Nikhat Akhtar
Department of Information Technology (IT), Goel Institute of Technology & Manage…
60 PUBLICATIONS 517 CITATIONS

SEE PROFILE

All content following this page was uploaded by Nikhat Akhtar on 04 February 2015.

The user has requested enhancement of the downloaded file.

International Journal of Computer Science and Telecommunications [Volume 5, Issue 8, August 2014] 26

The State of the Art Handwritten Recognition of Arabic

Script Using Simplified Fuzzy ARTMAP and Hidden
Markov Models
ISSN 2047-3338

Yusuf Perwej1, Shaikh Abdul Hannan2 and Nikhat Akhtar3

1,2
Department of Computer Science & Engg., Al Baha University, Al Baha, Kingdom of Saudi Arabia (KSA)
3
Department of Computer Science & Engg., Integral University, Lucknow, India
1
[email protected], [email protected], [email protected]

Abstract– In this paper, we present recognition of handwritten Qur’anical Arabic is not used in conversation or in non-
characters of Arabic script. Arabic is now the 6th most spoken religious writing and Modern Standard Arabic is the official
language in the world and is spoken by more than 200 million language of the Arabic world. Colloquial Arabic refers to
people worldwide. The 7th Century A.D., Arabic started to Arabic that is spoken with a dialect [3].
spread to the Middle East as many people started to convert to
The modern Arabic language writing system runs from
Islam. During this time of religious conversions, Arabic replaced
many South Arabian languages, most of which are no longer right to left and is a cursive script [4]. There are twenty eight
commonly spoken or understood languages. The challenges in letters in the alphabet, but because the script of the alphabet is
Arabic handwritten character recognition wholly lie in the cursive, 22 of the letters take different shapes when they are
variation and disfigurement of Arabic handwritten characters, in initial, medial, final, or isolated positions [5]. There are six
since different Arabic people may use a different style of letters in the alphabet which have only two presumable forms
handwriting, and direction to draw the same shape of the because you only connect to them; they cannot be connected
characters of their known Arabic script. Though various new from. The three long vowels are represented within the
propensity and technologies come out in these days, still alphabet. However, the three short vowels are not. Short
handwriting is playing an important role. To recognize Arabic
vowels can be indicated by optional diacritical markings [6],
handwritten data there are different strategies like Simplified
Fuzzy ARTMAP and Hidden Markov Models (HMM). In this but these are most often not written. Those texts in which they
paper, we are using Simplified Fuzzy ARTMAP, which is an are written are usually of a religious nature and they are
updated version of Predictive Adaptive Resonance Theory. It included to ensure that the proper pronunciation is made for
also has a capacity to adjust clusters, as per the requirements all the words. Note that Arabic is particularly rich in uvular,
Arabic script, which is remunerative to mitigate noise. We have pharyngeal, and pharyngealized ("emphatic") sounds.
tested our method on Arabic scripts and we have obtained The Arabic handwritten script recognition is an open field
encouraging results from our proposed technique. of research which has a large amount of scope for
development. A some models are already enforcing for the
Index Terms– Hidden Markov Model (HMM), Arabic Script,
hand written character recognition system include framework
Handwriting, Fuzzy ARTMAP, Recognition and Feature
Extractor based models, support vector machines, stochastic models,
and learning-based models etc. In this paper using a Hidden
Markov Models because Hidden
I. INTRODUCTION Markov Models are mainly sequence [7] classifiers and are
frequently used for recognition of Arabic handwritten script.
They are stochastic models and can encounter with noise and
A RABIC is the official language of many countries in the
Middle East such as Saudi Arabia, Jordan, Lebanon,
Libya, Egypt, Iraq, Morocco, and Sudan. It is also one of
also give diversifications in Arabic handwriting.
progressist viewpoint has been proposed to reform simplified
A

the six official languages of the United Nations. Arabic is a fuzzy ARTMAP Neural Network [8] performance for
“Semitic,” language and is most closely related to Aramaic character recognition of handwritten Arabic script. The few
and Hebrew. Semitic languages are based on a consonantal fuzzy values are used for the similar Arabic script to
root system. Every word in Arabic is derived from one or ameliorate recognition. The fuzzy
dissimilar root word [1]. By the 7th Century A.D., Arabic ARTMAP beforehand gives preferable execution for other
started to spread to the Middle East as many people started to characters [9]. In this paper available the all information
convert to Islam. During this time of religious conversions, keeps together by both labeled and unlabeled Arabic patterns,
Arabic substitute many South Arabian languages, most of it is essential to intermingle supervised and unsupervised
which are no longer commonly [2] spoken or understood learning in a single training algorithm. We are using a
languages. There are three forms of Arabic; Qur’anical simplified fuzzy ARTMAP and hidden Markov models in
Arabic, Modern Standard Arabic, and Colloquial Arabic. Arabic script. We have acquired spanking outcome.

Journal Homepage: www.ijcst.org

 Yusuf Perwej et al. 27

II. THE HIDDEN MARKOV MODELS

The Hidden Markov Model (HMM) is a sinewy statistical

tool for modeling productive sequences that can be
characterized by an underlying process productive a
noticeable sequence [10]. In today's scenario hidden Markov
model have found application in many areas take an interest
in signal processing, handwriting recognition, phrase Fig. 1. The Graphical model of a Hidden Markov Model
chunking, and character recognition and in particular speech
processing. Andrei Markov gave his name to the In the Markov Model we could write out the thorough
mathematical theory of Markov processes in the early likelihood, since all the variables have been look on. In
twentieth century, but it was Baum and his colleagues that describing the likelihood [14] of the Hidden Markov Model,
developed the theory of hidden markov model in the we have to dislodge all unpredictability on the hidden states
1960s [11]. X, which is done by conjecture. To calculate the likelihood of
In a hidden Markov model, one does not know anything the set of observations D = {y1, y2, y3,……yN} using the
about what originates the observation sequence. The number above rules gives:
of states, the transition prospect, and from which state an
observation is originated are all unknown [12]. Instead of
intermingling each state with a cogent output; each state of
the hidden Markov model is associated with a prospect
function. At time t, an observation ot is originate by a prospect The aloft expression can also be written using the
function bj (ot), which is associated with state j, with the annotation of transition matrix, observation matrix and initial
likelihood, state likelihood:

An HMM is composed of a five tuple: (S, K, ∏, A, B)

Where bxn = Sj (yn =ci) = bj,i is the emission likelihood (the
1. S = {1……..N} is the set of states. The state at time t is likelihood of emitting symbol ci have given we are in state Sj).
denoted st. The usual hidden markov model outlook framework is an
unsupervised learning technique which assent new patterns to
2. K = {k1…….kM} is the output alphabet. In a discrete be found. It can handle variable lengths of input sequences,
observation density case, M is the number of observation without having to inflict a template to cognize.
pick. In the above example, M equals the number of
presumable movements. III. THE PROBLEM STATEMENT

3. Initial state delivery ∏ = {πi}, i ϵ S. πi is defined as: The Arabic character recognition is exceedingly arduous to
automate. The humankind being can diagnosticate variegated
objects and make cognition out of volumetric amount of
4. State transition prospect delivery A = {aij}, i, j, ϵ S. visual information, seemingly requiring very diminutive
attempt. The emulate task execution by humankind to
diagnosticate to the extent allowed by physical barricades will
be extremely gainful for the system. The difficulty contains in
5. Observation symbol prospect delivery B = bj (o t). The the real world data handwritten Arabic alphabets, where
prospect function for each state j is: handwritten Arabic characters are the input to the system,
while in print characters will be goal output of the system.
In this paper collection of Arabic character data. We are
Since modeling a problem as a hidden Markov model, and particular piece of paper has to be designed for the Arabic
acknowledging that some set of data was originated by the data collection. The Arabic data are collected from many
hidden Markov model, we are potentially viable to calculate people from various ages and realm. The Arabic character
the prospect of the observation sequence [13] and the data obtaining is done manually i.e., The bond piece of paper
potential fundamental state sequences. We can train the model was provided to the respondent and asked to inscribe the
parameters based on the comply data and get a more actual Arabic characters from to for one time. Because bond paper is
model. Then use the trained model to predict unappreciated a strong, high quality, durable writing paper similar to bank
data. paper, but having a weight greater than 50 g/m2. After that
A Hidden Markov Model is that the states are look on bond piece of paper are scanned using HP Scanjet 5590
directly on the Markov Model, and look on indirectly with Digital Flatbed Scanner at 2400 dpi optical resolution, which
unpredictability in the hidden markov model. The above is gives low noise and good quality image. The digitized images
best be an example using the graphical model representation are stored in BMP file shown in Fig. 2.
see in Fig. 1.
International Journal of Computer Science and Telecommunications [Volume 5, Issue 8, August 2014] 28

or binary input patterns. It endows a unique solution to the

stability-plasticity suspense faced by autonomous learning
systems. Because fuzzy ARTMAP can perform fast, stable,
on-line, unsupervised or supervised, incremental learning, it
can learn from novel events encountered in the field, yet
overcome the problem of catastrophic forgetting associated
with many renowned neural network classifiers [21].
FAM is composed of a pair of fuzzy ART modules, ARTa
and ARTb connected by an inter-ART module called
Mapfield. ARTa and ARTb are used for coding the input and
Fig. 2. The Handwritten Arabic Script Data Collection
output patterns serially, and Mapfield allows mapping
between inputs and outputs. The ART a modules [22] include
the input layer Fa1 and the competitive layer Fb2. A
IV. PRAPOSED SOLUTION BY SIMPLIFIED FUZZY preprocessing layer Fa0 is also added before. Analogous layers
ARTMAP FOR ARABIC SCRIPT appear in ARTb. Fuzzy ARTmap architecture is shown in
Fig. 3.
The ART was introduced by Stephen Grossberg in The preparative input vectors have the form: a = (a1, . . . ,
1976.The term resonance refers to the so called resonant state an) ∈ {0, 1}n. A data preprocessing technique called
of network in which a category [15] prototype vector matches complement coding is featured in the two fuzzy art modules
the current input vector so close enough that the orienting by the Fa0 (and Fb0 serially) layer in order to avoid diffusion of
system will not generate a reset signal in the other nodes. Each input vector a produces the normalized vector A
consideration layer. The networks learn only in their resonant = (a, 1 − a) whose L1 norm is constant: |A| = n.
states [16]. The architecture of ART is based on the idea of
adaptive resonant feedback between two layers of nodes as
developed by Grossberg (1988).
ART permitted for cluster input by unsupervised learning.
One can current input patterns [17] in any order. Each time
when a pattern is presented, a suitable cluster unit is selected
and the cluster’s weights are adjusted to let the cluster unit to
cognize the pattern. There is one drawback with HMM also
that is it takes lots of training time in particular at a higher
number of states the Arabic characters. Also the recognition
time is elongate as the increase in states. Many researchers
thought that Backpropagation as a solution for the Pattern
recognition. But the Fuzzy ARTMAP based on Predictive
ART (ARTMAP) can also be used for pattern recognition of
characters. This network gives optimum performance on
image the Arabic character recognition [18]. Fig. 3. The Fuzzy ARTMAP Architecture
Therefore, in this paper, we are applied to Arabic character
recognition too. The network will be more accomplished for Let Ma be the number of nodes in Fa1 and Na be the number
the recognition of similar characters. ARTMAP is of nodes in Fb2. Due to the preprocessing step, Ma = 2n. Wa is
supervised, self learning, real time, a self organizing network, the weight vector between Fa1 and Wabj . Each Fa2node
which supports slow and fast Arabic characters learning, represents a class of inputs grouped together, notify as a
whereas, Backpropagation is a supervised and supports slow category. Each Fa2category has its own set of adaptive weights
learning. This model employs simple learning equations with stored in the form of a vector Waj, j= 1……Na whose
a single Arabic user selectable parameter and can learn single geometrical interpretation is a hyper-rectangle inside the unit
training pattern within a small number of training iterations. box. Similar notations and affirmations are valid for ART b
The Fuzzy ARTMAP efficient, high speed, accuracy in both that receives the m-dimensional input vectors. For a
of Arabic character recognition like online as well as offline classification problem, the class index is the same as the
[19]. It has a number of parameters and it requires no problem category number in Fb2, thus ARTb can be simply substituted
conspicuous crafting or liking of initial weights or parameters. an Nb dimensional vector.
Simplified Fuzzy ARTMAP is the structure developed by The Mapfield module allows fuzzy ARTmap for execution
Kasuba in the year 1993. It is a vast simplification of of heteroassociative tasks, establishing many-to-one links
Carpenter and Grossberg’s fuzzy ARTMAP [20]. This has a between various categories from ART a and ARTb, serially.
diminished computational overhead and architectural The number of nodes in Mapfield is equal to the number of
exorbitance when compared to its predecessor fuzzy nodes in Fb2. Each node j from Fa2 is linked to every node
ARTMAP. The fuzzy ARTMAP neural network architecture from Fb2via a weight vector Wabj. The learning algorithm is
is competent of self-organizing stable recognition categories sketched below. Training pattern, the watchfulness parameter
in response to pleasing one's own mind sequences of analog factor is set equal to its baseline value, and all nodes are not
 Yusuf Perwej et al. 29

inhibited. For each input A, a fuzzy choice function is used to each of which determine some quantifiable property of an
get the response for each Fb2 category: Arabic object, and is computed such that it quantifies some
valued Arabic characteristics of the object. We classify the
various features [24] currently employed as follows:
• General features: Application unattached features such as
Let J be the node with the highest value computed as in color, texture, and shape. According to the abstraction level,
(1). If the resonance condition from eq. 2 is not fulfilled, they can be further divided into three types, firstly the pixel
Then the Jth node prevents such that it will not take part to level features these features calculated at each pixel, e.g.,
further competitions for this pattern and a new search for a color, location. Secondly the local features these features
resonant category is performed. This might lead to the calculated over the results of subdivision of the image band
creation of a new category in ARTa. on image segmentation or edge detection. Lastly the global
features: these features calculated over the overall image or
just a regular sub - area of an image.
• Domain-specific features: Application unattached features
such as human faces, fingerprints, and conceptual features.
An identical process occurs in ART b and let K be the These features are frequently a synthesis of low-level features
winning node from ARTb. The Fb2output vector is set to: for a specific domain. On the other hand, all features can be
coarsely classified into low-level features and high-level
features. Low-level features can be extracted directed from
the actual images, whereas high-level feature extraction must
be based on low level features [25].
This task is an example many facets of a typical Arabian
An output vector X ab is formed in Mapfield: X ab = y b pattern recognition problem, including feature selection,
∧Wabj . A Mapfield vigilantly test controls the match dimensionally deficiency and the use of prestigious
between the predicted vector X ab and the target vector yb : descriptors jiffy are the extracted features derived from raw
measurements. Where jiffy are used to obtain Arabic script
Scaling (ASS), Arabic script Rotation (ASR), Arabic script
Translation (AST) unalterable. The disposition of
immutability to ASS, ASR, AST transforms may be derived
using the function of jiffy. The jiffy transformation of an
Where ρab∈ {0, 1} is a Mapfield vigilantly parameter. If
Arabic image function ĄƗƑ(x,y) is given by:
the test from (4) is not passed, then a sequence of steps called
match tracking is initiated (the vigilantly parameter ρ a is
 
 
m n
increased and a new resonant category will be solicited for lmn = ĄƗƑ(x,y) m , n= 0,1,2,……,∞
 
ARTa), otherwise learning occurs in ART a, ARTb and
Mapfield: In the case of a spatially calumniate 9×11 (M×N) character
denoted by ĄƗƑ(i,j)is approximated as,

 
5 7
lmn = im jn ĄƗƑ(i,j)
i 0 j 0
(And the analogous in ARTb) and Wabjk = δkk, where δij is
Now the value of ĄƗƑ(i,j)0 or 1 keep faith upon whether the
Kronecker’s delta. With respect to βa, there are two learning
(i,j)th pixel The moment severity is represented by several
modes firstly the speedy learning for βa = 1 for the entire
prospects like the way of handwriting, ink used for Arabic
training process, and secondly the speedy commit and gently
written character i.e. 0 ≤ ĄƗƑ(i,j)≤ 1 point to that the severity
recode learning corresponds to setting βa = 1 when creating a
[26] falsehood amid the ends of a spectrum. However ĄƗƑ(i,j)
new node and βa < 1 for subsequent learning.
is static over any pixel realm and ponder it as pivotal jiffy.
In batch supervised learning mode, fuzzy ARTMAP may
The pivotal jiffy is given by:
also be accomplished in that its asymptotic [23] generalization
  ( i - iˆ )m (
error can be attained for a moderate time and space 5 7
mn=
i 0 j 0
j - ĵ )n ĄƗƑ(i,j) Where
complexity. They have been swimmingly applied in complex
real-world pattern recognition tasks such as the recognition of
l10 ˆ l 01
radar signals multi-sensor image fusion [24], remote sensing iˆ  , j
and data mining recognition of handwritten characters and l 00 l 00
signature verification . The pivotal jiffies are so far sentient to ASR and ASS
transformation. The scaling unalterable may be procured by
V. WORKING OF FEATURE EXTRACTOR FOR ahead normalizing μmn as:
ARABIC SCRIPT

mn = μmn m+n=2,3,…..
The assortment of two-dimensional objects from Arabic
mn
visual image data is vital Arabic pattern recognition task. The μ00 1
feature is defined as a function of one or more measurements, 2
International Journal of Computer Science and Telecommunications [Volume 5, Issue 8, August 2014] 30

VI. THE ALGORITHEM FOR TRANING AND Phase 10: If CFn (BIx) < ƥ then If few much top down
INFERENCE PHASES FOR ARABIC SCRIPT weight nodes exist, then Contemplate the next topmost
winner WN among the top-down weight nodes.
The algorithms for Training and Inference phases of Go to Phase 8;
SFARTMAP on the basis of the prolonged definition of Else go to Phase 11;
complementary distribution are as follows:
Phase 11: Make a recently top-down weight node RWfirst such
A) The Training phase of Arabic Script that Wfirst= BIx and link the node to the ranking RN;
Phase 1: Select a suitable value for the vigilantly parameter
(0 < ƥ < 1) and a small value for š. The conglomeration Phase 12: If nope, more input patterns, then go to Phase 14;
number of training epochs to the desired number of training
epochs and Enumerate of training epochs to 0. Phase 13: Otherwise x ← x +1
Go to Phase 3;
Phase 2: x← 1;
Enumerate of training epochs = Enumerate of training epochs Phase 14: GOTO Phase 2;
+1; While (Enumerate of training epochs ≤ number of training
epochs) B) The Inference phase of Arabic Script
Repeat Phases 3 – 12 else 13;
Phase 1: Let Wy , y=1,2,3,………..,m allude m top-down
Phase 3: Input the pattern vector Ix = (bx1, bx2, bx3, bx4,…….., weight vectors procured after training the network with a
bxd) of dimension d and its ranking Rx. [28] given set of training patterns;

Phase 4: Count the augmented input vector using the detailed Let Ix be the conjecture pattern set each of whose ranking is to
definition of complementation of Fuzzy set under the be drawn conclusion of the network;
presumable cases. x ← 1;
BIx = (bx1, bx2, bx3, bx4,…….., bxd, 1- bx1, 1- bx2, 1- bx3, 1- bxd)
Phase 2: Perceive input Ix;
Phase 5: If BIx is the first input in the given ranking Rx
conglomeration the top down weight vector W x as BIx Link Phase 3: Count the augmented input BIx;
Wx to the ranking Rx.
Go to Phase 12 else 13. Phase 4: for y ← 1 to m count the conative functions

BI x  Wx
Phase 6: If BIx is an input pattern vector whose ranking CFy  BI x  
previously exits, then count the conative function CFy(BIx) for š  Wy
each of the existing top-down weight nodes TDWy
Phase 5: Select the winner N among the M conative functions
BI x  Wx
CFy  BI x   Nc(BIx) = maxy Ny(BIx)
š  Wy
Phase 6: Output ranking RN linked to Nc(BIx) as the one to
Phase 7: Select that top-down weight node N which records which Ix pertain to.
the transcendent [29] conative function Nc(BIx) = maxy
Ny(BIx) Phase 7: If nope, more conjecture pattern vectors
Then exit
Phase 8: Count the correspond function CFn(BIx) of the Else x ← x +1;
vanquish node N; If CFn (BIx) > ƥ and Rx is same as that Go to Phase 2.
ranking RN linked to WN Then update weight vector WN as
WN recently = WN longstanding + (I ⋀ WN longstanding) VII. EXPRIMENTAL RESULTS

Phase 9: If CFn (BIx) > ƥ and Rx is not the ranking RN linked The performance of the Simplified Fuzzy ARTMAP and
to WN then Initiate harmonize tracing by setting to CFn (BIx) Hidden Markov Models (HMM) is given in figure 4. We look
and incrementing by a miniature value ƛ then on that up to noise level of 0.20 to 0.25 and handwritten
ƥ = CFn (BIx) + ƛ recognition of Arabic Script is 100%. The table 1 shows the
If few much top down weight nodes exist then contemplate feature value of handwritten Arabic characters. In the next
the next topmost winner W N among the top-down weight table 2 gives a recognition rate, which is based on the
nodes disparity between feature values of ideal and handwritten
Go to Phase 8; Arabic characters. In table 2 the last row notifies the average
Else go to Phase 11; of handwritten Arabic character recognition rate. We look on
that up to 96.38% recognition is instate for handwritten
Arabic Script.
 Yusuf Perwej et al. 31

results of simulations Arabic script database illustrate that a

precipitant, accurate and noise obstructive module has been
successfully designed by using Simplified Fuzzy ARTMAP
of the mentioned methods. An experimental result has shown
96.38% recognition rate that the Simplified Fuzzy ARTMAP
is patently preferable to the other model in recognizing the
handwritten Arabic script.

REFERENCES

[1]. Versteegh, Kees. 1997a. The Arabic Language. New York:

Columbia University Press.
[2]. Suleiman, Yasir, ed. 1994. Nationalism and the Arabic
Language: A Historical Overview, pages 3–24. Surrey: Curzon
Press Ltd.
[3]. Belnap, R. Kirk and Niloofar Haeri. Structuralist Studies in
Fig. 4. The Performance of Simplified Fuzzy ARTMAP and Hidden Markov Arabic Linguistics: Charles A. Fergusonís Papers, 1954-1994.
Models (HMM) Leiden: Brill, 1997.
[4]. MA. Attia , T. Salakoski , F. Ginter , S. T. Pyysalo
,Accommodating Multiword Expressions in an Arabic LFG
Table 1: The Handwritten Arabic Character of Feature Value
Grammar , In Finland Springer-Verlag Berlin Heidelberg, vol
4139, pp 87 – 98, 2006
[5]. Farghaly, Ali. December 2008. Arabic NLP: Overview, state
of the art, challenges and opportunities. In The International
Arab Conference on Information Technology, ACIT2008 .
Hammamat, Tunisia.
[6]. J. Dichy ,On lemmatization in Arabic. A formal definition of
the Arabic entries of multilingual lexical databases, ACL 39th
Annual Meeting. Workshop on Arabic Language Processing;
Status and Prospect. Toulouse, pp 23-30, 2001.
[7]. L. R. Rabiner and B. H. Juang. An introduction to hidden
Markov models. IEEE ASSP Mag., Vol.3, No.1, p.4-16, June
1986.
[8]. M. Taghi, V. Baghmisheh, and P. Nikola. A Fast Simplified
Fuzzy ARTMAP Network. Neural Processing Letters, 17,
Table 2: The Handwritten Arabic Character of Recognition Rate 2003, 273–316.
[9]. G. A. Carpenter, S. Grossberg, and D. B. Rosen, “Fuzzy ART:
fast stable learning and categorization of analog patterns by an
adaptive resonance system,” Neural Networks, 4:6, 759-771,
1991.
[10]. E. Fosler-Lussier : Markov Models and Hidden Markov
Models : A Brief Tutorial, Technical Report (TR-98-041),
December 1998, International Computer Science Institute,
Berkeley, California.
[11]. X. D. Huang, Y. Ariki, and M. A. Jack. Hidden Markov
Models for Speech Recognition. Edinburgh University Press,
Edinburgh, 1990.
[12]. S. Fine, Y. Singer, and N. Tishby. The hierarchical Hidden
Markov Model: analysis and applications. Machine Learning,
32, 1998.
VIII. CONCLUSION
[13]. R. Nag, K. H. Wong, and F. Fallside. Script recognition using
hidden markov models. In ICASSP 86, 1986.
In this paper, we proposed an improved on recognition of [14]. T. Jebara and A. Pentland. Maximum conditional likelihood
handwritten characters of Arabic script. Arabic is now the 6 th via bound maximization and the CEM algo- rithm. In
most spoken language in the world and is spoken by more Advances in Neural Information Processing Systems vol 11,
than 200 million people worldwide. The Arabic script pp 494 – 500, MIT Press, 1998
handwriting recognition is a very challenging task due to its [15]. G. A. Carpenter, S. Grossberg, N. Markuzon, J. H. Reynolds,
cursive nature. Because every Arabic handwritten character and D. B. Rosen, “Fuzzy ARTMAP: a neural network
has unlimited shapes, patterns and styles even when it is architecture for incremental supervised learning of analog
written by the same or a different Arabic person. The most multidimensional maps,” IEEE Trans. on Neural Networks,
Arabic character recognition systems can not read offended 3:5, 698-713, 1992.
[16]. E. Granger, et. al. “Classification of incomplete data using the
documents and handwritten Arabic script or character. In this fuzzy ARTMAP neural network,” Proc. 2000 Int. Joint
study, we are putting forward an Arabic handwritten character Conference on Neural Networks, 4, 2000, pp. 35-40.
recognition system using Simplified Fuzzy ARTMAP. The
International Journal of Computer Science and Telecommunications [Volume 5, Issue 8, August 2014] 32

[17]. R. Andonie and L. Sasu. Fuzzy ARTMAP with input Dr. Yusuf Perwej Assistant Professor in the
relevances. IEEE Transactions on Neural Networks, 17, 2006, Department of Computer Science & Engineering Al
929–941. Baha University, Al Baha , Kingdom of Saudi
[18]. Kwan, H. K., and Cai, Y. (1994). ‘‘A fuzzy neural network Arabia (KSA). He has authored a number of
and its application to pattern recognition.’’ IEEE Trans. on different journal and paper. His research interests
Fuzzy Systems, 2(3), 185–191. include Soft Computing, Artificial Neural Network,
[19]. C. P. Lim, H. H. Toh, and T. S. Lee, ”An evaluation of the Machine Learning, Pattern Matching, Pattern Recognition, Artificial
fuzzy ARTMAP neural network using offline and on-line Intelligence, Image Processing, Fuzzy Logic, Genetic Algorithm,
strategies,” Neural Network World, 4, 327-339, 1999. Robotics, Bluetooth and Network etc. He is a member of IEEE.
[20]. I. Dagher, M. Georgiopoulos, G. L. Heileman, and G. Bebis.
Fuzzy ARTVar: An improved fuzzy ARTMAP algorithm.In
Proceedings IEEE World Congress Computational
Intelligence, Anchorage, 1998, 1688–1693. Dr. Shaikh Abdul Hannan Assistant Professor
[21]. Vazquez-Lopez, J. A., Lopez-Juarez, I., Peña-Cabrera, M. in the Department of Computer Science &
2010. On the use of the FuzzyARTMAP Neural Network for Engineering Al Baha University, Al Baha ,
Pattern Recognition in Statistical Process Control using a Kingdom of Saudi Arabia (KSA). He has authored
Factorial Design. Int. J. of Computers, Communications & a number of different journal and paper. His
Control, Vol. V (2), pp. 205-215. research interests include Data Mining and Data
[22]. P. Henniges, E. Granger, and R. Sabourin, “Factors of Warehouse, Artificial Neural Network, Artificial
overtraining with fuzzy ARTMAP neural networks,” Int. Joint Intelligence, Image Processing etc.
Conference on Neural Networks, Montreal, Canada, 2005, pp.
1075-1080.
[23]. B. Lerner and B. Vigdor, ”An empirical study of fuzzy
ARTMAP applied to cytogenetics,” IEEE Convention of
Electrical and Electronics Engineers in Israel, 2004, pp. 301- Nikhat Akhtar Assistant Professor in the
304. Department of Computer Science & Engineering
[24]. M. Taghi, V. Baghmisheh, and P. Nikola. A Fast Simplified Integral University, Lucknow, India. She has
Fuzzy ARTMAP Network. Neural Processing Letters, 17, authored a number of different journal and paper.
2003, 273–316. Her research interests include Soft Computing,
[25]. Addison, J F D, Wermter, S and MacIntyre, J. 1999 Swarm Intelligence, Storage Technology, Artificial
Effectiveness of feature extraction in neural network Neural Network, Cryptography, Pattern Matching, Pattern
architectures for novelty detection, ICANN-99, Ninth Recognition, Artificial Intelligence, Network Security, Fuzzy Logic,
International Conference on Artificial Neural Networks”, Network and Database. She is a member of IEEE.
Edinburgh, UK, September 1999, pp976-981.
[26]. Yusuf Perwej , Dr. Ashish Chaturvedi, “Machine Recognition
of Hand Written Characters using Neural Networks” for
published in the International Journal of Computer
Applications (IJCA) ,USA , Vol. 14, No.2, January 2011,
Pages 6- 9, ISSN 0975 – 8887, DOI : 10.5120/1819-2380
[27]. E. Saber, A.M. Tekalp, ”Integration of color, edge and texture
features for automatic region-based image annotation and
retrieval,” Electronic Imaging, 7, pp. 684–700, 1998.
[28]. I. Dagher, M. Georgiopoulos, G. L. Heileman, and G. Bebis.
Fuzzy ARTVar: An improved fuzzy ARTMAP algorithm. In
Proceedings IEEE World Congress Computational Intelligence
WCCI’98,Anchorage, 1998, 1688–1693.
[29]. E. Gomez-Sanchez, Y. A. Dimitriadis, J. M. Cano-Izquierdo,
and J. Lopez-Coronado. Π ARTMAP: Use of mutual
information for category reduction in fuzzy ARTMAP. IEEE
Transactions on Neural Networks, 13, 2002, 58–69.

View publication stats