An Ensemble Learning Approach using Decision Fusion for the

Recognition of Arabic Handwritten Characters

Rihab Dhief

, Rabaa Youssef

1,2

and Amel Benazza

University of Carthage SUP’COM, LR11TIC01, COSIM Lab., 2083, El Ghazala, Tunisia

INSAT, University of Carthage, Tunisia

Keywords:

Handwritten Arabic Character Recognition, Skeletonization, Freeman Chain Code, Heutte Descriptors,

Feature Extraction, Supervised Machine Learning Algorithms, Deep Learning.

Abstract:

The Arabic handwritten character recognition is a research challenge due to the complexity and variability of

forms and writing styles of the Arabic alphabet. The current work focuses not only on reducing the complexity

of the feature extraction step but also on improving the Arabic characters’ classiﬁcation rate. First, we lighten

the preprocessing step by using a grayscale skeletonization technique easily adjustable to image noise and

contrast. It is then used to extract structural features such as Freeman chain code and Heutte descriptors.

Second, a new model using the fusion of results from machine learning algorithms is built and tested on

two grayscale images’ datasets: IFHCDB and AIA9K. The proposed approach is compared to state-of-the-art

methods based on deep learning architecture and highlights a promising performance by achieving an accuracy

of 97.97% and 92.91% respectively on IFHCDB and AIA9K datasets, which outperforms the classic machine

learning algorithms and the deep neural network chosen architectures.

1 INTRODUCTION

Arabic is an international language widely spoken in

the world. The Arabic alphabetic contains 28 letters

written from right to left and they are highly similar to

each other. The Arabic language has ﬂourished in sev-

eral eras, and gained a scientiﬁc, literary and religious

weight. Digitizing this heritage is very important for

better archiving and exploration.

Optical Character Recognition (OCR)

(Borovikov, 2014) is the process of converting

images of handwritten or printed text into digital,

machine-editable text. Despite the attention given

to the optical character recognition ﬁeld and the

interesting results of the literature (Alaei et al., 2012;

Rajabi et al., 2012; Siddhu et al., 2019; Althobaiti

and Lu, 2017; Altwaijry and Al-Turaiki, 2021;

Boulid et al., 2017; KO and Poruran, 2020; Balaha

et al., 2021b), the recognition of Arabic handwritten

characters still has its challenges and difﬁculties. In

fact, its cursive writing style generates a variation

in shape, curve angles and size of each character,

depending on its position in the word. Furthermore,

various characters in the Arabic alphabet have the

same main body but can be differentiated only

by the position and number of the diacritics (Lutf

et al., 2010). Many Arabic datasets are provided by

open-access resources. Some datasets include only

simpliﬁed binary characters (El-Sawy et al., 2017;

Altwaijry and Al-Turaiki, 2021), while some others

gather grayscale images (Torki et al., 2014; Mozaffari

et al., 2006).

In this work, we propose to investigate a new ar-

chitecture that brings together many algorithms in or-

der to take advantages of each of them. First, we pro-

pose to combine structural and statistical features. To

obtain the structural features, the Self-Noise and Con-

trast Controlled Thinning algorithm (Youssef et al.,

2016) is implemented. This algorithm lighten the pre-

processing step by improving the model robustness to

noise and low-contrast. In fact, the SCCT algorithm

has proven its efﬁciency in the medical ﬁeld (Mallat

and Youssef, 2016), when directly applied on X-Ray

images. Second, the principle contribution of the cur-

rent work consists of implementing a decision fusion

based on the most efﬁcient machine learning classi-

ﬁers, namely SVM, KNN and RF.

The structure of this paper is described as follows.

First, Section 2 reviews the pertinent works in Ara-

bic handwritten recognition. Second, Section 3 de-

scribes the datasets used in this work. Then, Sec-

tion 4 highlights the feature extraction step, using the

SCCT skeletonization. Section 5 details the proposed

model: the fusion of machine learning classiﬁers. Ex-

perimental results are showed in Section 6. Finally,

conclusions are drawn in Section 7.

Dhief, R., Youssef, R. and Benazza, A.

An Ensemble Learning Approach using Decision Fusion for the Recognition of Arabic Handwritten Characters.

DOI: 10.5220/0010839500003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 51-59

ISBN: 978-989-758-549-4; ISSN: 2184-4313

2 RELATED WORK

For years, the problem of classiﬁcation of Arabic

characters has existed, several studies have been car-

ried out to ﬁnd as much precision as possible. Three

important steps in the recognition system should be

treated : Preprocessing , Feature Extraction and Clas-

siﬁcation.

2.1 Preprocessing

Preprocessing aims at removing unnecessary infor-

mation without modifying the form of the object.

This is a fundamental step to ensure good classiﬁca-

tion results. Traditional preprocessing methods are

generally ﬁltering and noise removal (Althobaiti and

Lu, 2017; Sahlol et al., 2014). Thining and Ob-

ject contour are also often used techniques especially

for identifying the object structure (Boufenar et al.,

2018). These techniques are efﬁcient in the character

recognition ﬁeld but since they are based on binary

image, they still risk losing useful information during

the binarization step.

2.2 Feature Extraction

Feature engineering is a primordial step for every

ML learning model. It consists of transforming im-

age data into features that better represent the under-

lying problem to the predictive models, resulting in

improved model accuracy on unseen data. Recently,

many researchers use deep features for the feature

extraction step. This method does not need further

preprocessing and achieves high performance accord-

ing to the literature (Altwaijry and Al-Turaiki, 2021;

Balaha et al., 2021a). Despite its effectiveness, it re-

quires a higher computational costs than traditional

methods. Traditional feature extraction is performed

through two main shape description approaches: sta-

tistical and structural. Studies based on supervised

learning using only structural features (Althobaiti and

Lu, 2017), or statistical approaches (Alaei et al., 2012;

Rajabi et al., 2012), and also the combination of both

(Alaei et al., 2012; Zanchettin et al., 2012; Sahlol

et al., 2014; Sahlol et al., 2016; Boufenar et al.,

2018; Siddhu et al., 2019) were proposed in the litera-

ture with the latter achieving more interesting results.

Combining both type of features explains the need

for merging different classiﬁers since each classiﬁer

needs different form of features for better results.

2.3 Classiﬁcation

Previous works propose two main streams of ap-

proaches to deal with the Arabic handwritten recog-

nition problem, namely Deep Neural Network (DNN)

architectures (Altwaijry and Al-Turaiki, 2021; KO

and Poruran, 2020; Balaha et al., 2021a) and classi-

cal Machine Learning (ML) techniques (Zanchettin

et al., 2012; Alaei et al., 2012; Rajabi et al., 2012;

Sahlol et al., 2014; Sahlol et al., 2016; Boufenar

et al., 2018; Siddhu et al., 2019; Ali et al., 2020).

On the one hand, and regarding the use of Deep Neu-

ral Networks, authors of (Altwaijry and Al-Turaiki,

2021) propose a convolutional neural network ap-

proach (CNN) for the recognition of Arabic hand-

written characters using a small binary dataset and

achieving an accuracy result of 97%, while authors

of (Boulid et al., 2017; KO and Poruran, 2020; Bal-

aha et al., 2021a) conduct a research on grayscale

character images with a result of 96%. On the other

hand, and regarding the use of classical Machine

Learning (ML) approaches, three main models have

been widely applied, namely Support Vector Ma-

chine (SVM) (Zanchettin et al., 2012; Alaei et al.,

2012; Rajabi et al., 2012; Sahlol et al., 2014; Siddhu

et al., 2019), Random Forest (RF) (Sahlol et al., 2016;

Rashad and Semary, 2014) and K-Nearest Neighbors

(KNN) (Zanchettin et al., 2012; Rajabi et al., 2012;

Sahlol et al., 2014; Boufenar et al., 2018), with a

highlight on SVM results in most of the cited exper-

imental results. Furthermore, the idea of combining

multiple classiﬁers emerged in the past few years us-

ing both DL (Bosowski et al., 2021) and ML algo-

rithms (Zhao and Liu, 2020; Kaoudja et al., 2019).

Ensemble learning is an efﬁcient way to take advan-

tage of different classiﬁers especially when they have

heterogeneous inputs (different features type). In fact,

the authors of (Zanchettin et al., 2012) combine the

SVM and KNN classiﬁers, also, RF and KNN com-

bination is proposed in (Zhao and Liu, 2020) for

numeral recognition, and multi-classiﬁer system for

Arabic calligraphy recognition is built in (Kaoudja

et al., 2019), merging three classiﬁers namely: Mul-

tilayer Perceptron (MLP), SVM, and KNN. Although

state-of-the-art methods for recognising binary Arab

handwritten characters have achieved satisfactory re-

sults(Alaei et al., 2012; Rajabi et al., 2012; Zanchettin

et al., 2012; Sahlol et al., 2014; Zhao and Liu, 2020;

Kaoudja et al., 2019), further improvements remain

conceivable by adopting new approaches and method-

ologies on grayscale image datasets.

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

3 DATASETS

Two datsets are used in the curret work, namely

Isolated Farsi Handwritten Character Data Base

(IFHCDB) and AlexU Isolated Alphabet (AIA9K).

Both datasets contain grayscale images.

3.1 The IFHCDB Dataset

The IFHCDB dataset (Mozaffari et al., 2006) in-

cludes isolated Farsi and Arabic handwritten charac-

ters. In our project, only Arabic letters are considered.

The Arabic alphabet contains 28 letters and thus, our

model contains 28 classes. The main issue with this

dataset is the lack of balance between the classes as-

sociated to each character. In total, IFHCDB dataset

contains 51029 Arabic character with different num-

ber of observation varies from 40 to 10000 per class.

To reduce this unbalance, we merge classes based on

same character body and variable diacritics. We end

up with 18 classes presented in Figure 1.

Figure 1: The character bodies.

3.2 The AIA9K Dataset

The total size of AIA9K dataset (Torki et al., 2014) is

8736 divided in 28 classes with number of observa-

tions varies from 251 and 278 per class. For sake of

clarity, we merged classes with same character body,

as for IFHCDB dataset.

4 HANDCRAFTED FEATURE

EXTRACTION

Two types of features are extracted: structural ones

from the skeleton and statistical ones from the char-

acter body. Heutte et al. (Heutte et al., 1998) col-

lected a set of statistical and structural features that

describes the character globally (projection, moments

and proﬁles) and locally (Intersection with straight

lines, holes, concave arcs, junctions, endpoints and

extrema). These descriptors browse almost all what

could characterize a character. We brieﬂy describe

these features in the next two subsections.

4.1 Features Extracted from the

Character Body

Hu Moments: The seven equations are detailed in

(Hu, 1962) and are invariant to position, size and ori-

entation of the character.

Projections: Vertical and horizontal projections de-

rived from the histograms of the character image are

calculated and the maxima from each is extracted to

locate the most signiﬁcant object pixels number verti-

cally and horizontally.

Proﬁles: The proﬁle correspond to the set of differ-

ences between two consecutive pixels (between two

ordinates in the right and left proﬁles or two abscissas

otherwise). The proﬁles provide information about

the harmony of the character.

Heights and Widths: They describe the character

in terms of height (respectively width) in particular lo-

cations (1/5, 1/2 and 4/5). They are extracted from the

difference between the left and the right raw proﬁles

(respectively bottom and top) of the character bound-

ing box.

Extrema: Top, bottom, left and right extrema of the

character are extracted by browsing the image from

left to right, right to left, top to bottom and bottom to

top. Each time, the ﬁrst pixel which does not have an

8-connected object neighbour is recorded.

Concave Arcs: They are extracted from the object

contours. A concave arc is the set of three consecutive

points that form an angle of less than 180 degrees.

Ratio: The height to width ratio is extracted and

aims at characterizing the spread of the character.

4.2 Features Extracted from the

Skeleton

Self-Noise and Contrast Controlled Thinning (SCCT)

was developed by the authors of (Youssef et al.,

2016). It generates a smooth silhouette using two

thinning parameters: contrast and noise. This study

proposes to relax the topology preservation prop-

erty of homotopic thinning by considering local noise

and contrast, as shown by Figure 2. Applying this

skeletonization method improves the skeleton’s qual-

ity compared to binary skeletonization (Zhang and

An Ensemble Learning Approach using Decision Fusion for the Recognition of Arabic Handwritten Characters

(a) (b) (c)

Figure 2: Comparison:(a) Character body, (b) Binary thin-

ning, (c) Grayscale thining.

Suen, 1984) and consequently, the ﬁnal classiﬁcation

results.

Figure 2 presents two examples from class 14 and

6. The binary skeletonization (Zhang and Suen, 1984)

fails at recognizing the character’s hole for the letter

”sad” while creating one in the wrong place for letter

”ha”. This confusion is problematic since hole detec-

tion remains an essential feature to separate classes.

Endpoints and Junctions: An endpoint is an ob-

ject pixel that has only one 8-connected object neigh-

bor. A junction is deﬁned as a pixel having at least

three 8-connected object neighbors that separate the

background into three or more 4-connected compo-

nents.

Holes and Intersections with Straight Lines: A

hole is detected when the image background contains

more than one 4-connected component. The intersec-

tions extraction are deﬁned as follows: two horizontal

lines (1/3 height and 2/3 height) and a vertical line

crossing the character’s centre of gravity.

Freeman Chain Code: Freeman’s chain converts a

skeleton image into a directional code. Our process

of extracting the Freeman chain code is described as

follows: the starting point is the ﬁrst encountered end-

point pixel. Then, based on the position of the current

pixel neighbour, we pick the appropriate Freeman di-

rection to start constructing the chain code. Figure

3 (a) states the common choice of the 8 directions.

An example in Figure 3 (b) details the Freeman chain

code constructed for letter ”Lam”.

(a)

(b)

Figure 3: (a) Freeman 8 directions, (b) The chain code of

the letter ”lam”.

4.3 Features Extracted from Diacritics

Diacritics are obtained after removing the character

body from the image. The number, size and position

of each diacritic are extracted to differentiate between

characters having similar bodies but differing in di-

acritics. Figure 4 shows some example of diacritics

encountered in the AIA9K dataset.

Figure 4: Different types of diacritics.

5 THE PROPOSED

METHODOLOGY: ENSEMBLE

LEARNING CLASSIFIER

USING DECISION FUSION

Once these statistical and morphological features are

computed, they should undergo a classiﬁcation in or-

der to recognize the underlying character. In this re-

spect, we adopt a decision fusion strategy based on

the most efﬁcient machine learning classiﬁers used

in the character recognition ﬁeld, namely weighted

SVM, weighted RF and K-NN (Ayodele, 2010).

5.1 The Classiﬁers

Weighted Support Vector Machine: SVM is ini-

tially used for binary classiﬁcation. It consists of

deﬁning a hyperplane that separates two classes.

As we are faced to a multi-classiﬁcation problem

with m > 2 classes, the strategy one versus one is

adopted in order to apply m(m − 1)/2 binary classi-

ﬁers. Weighted SVM is adopted in our work because

of the datasets unbalance.

Weighted Random Forest: Random forest consists

of a set of decision trees. Every decision tree gives a

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

Figure 5: The proposed architecture for Arabic handwritten characters recognition.

predicted class. Then, the most frequently predicted

classes is chosen. RF and SVM are used to classify

characters using numeric statistical and structural fea-

tures.

K-Nearest Neighbors: K-NN is essentially based

on the calculation of metrics (distance) between ob-

servations. For each new observation, we can pre-

dict its class by looking at the classes of its nearest

neighbors. The number of considered neighbors in

the K-NN classiﬁer is the parameter K which is set

empirically from the beginning. The K-NN classiﬁes

the Freeman chains by using the Levenshtein distance

(Levenshtein et al., 1966), which is a string metric for

measuring the difference between two sequences.

5.2 Ensemble Learning: Decision-fusion

Principle

In this section, the suggested methodology for the

recognition of Arabic handwritten characters is pre-

sented. Figure 5 illustrates the proposed system in

a block diagram. First, a preprocessing step is pre-

sented, where two information sources of the char-

acter are described: the binary character itself and

its skeleton graph. The aim of this work is to test

the SCCT method, which makes it possible to avoid

part of the pretreatments, and to see its effect on the

classiﬁcation rate.The second step consists in extract-

ing features from each character form. Structural fea-

tures are extracted from the skeleton, while statistical

ones are derived from the character body. Besides, the

number, position and size are also calculated for the

diacritics.

The ﬁnal step is the classiﬁcation, which is imple-

mented in two main steps :

1. First stage: The character body classiﬁcation: As

described in subsection 5.1, three main classiﬁers

are used, namely SVM, RF, K-NN and for which

we choose different features as input. In fact,

K-NN uses only the Freeman chain code feature

since this chain requires the use of a speciﬁc met-

ric which is here the Levenshtein distance. Re-

garding SVM and RF, we choose to implement

them using the remaining features described in

this work. Since the classiﬁcation error is differ-

ent from a classiﬁer to another, a comparison be-

tween their results is made, and a vote that merges

the respective decisions is built. In fact, we choose

the most common predicted class between the

three classiﬁcation results. If the three predicted

classes are different, the class corresponding to

the best global accuracy is picked.

2. Second stage: Separate merged classes using the

diacritics: in fact, due to the similarity between

classes and the notable unbalance of the datasets,

merging classes having the same body but differ-

ing by the diacritics’ forms was adopted to im-

prove the global accuracy. In this step, we sepa-

rate the merged classes using information related

to the diacritics. Each character is classiﬁed by

adding the features related to its diacritics, and in

this case, the SVM classiﬁer is used since it gen-

erates better results when dealing with unbalanced

data.

An Ensemble Learning Approach using Decision Fusion for the Recognition of Arabic Handwritten Characters

Table 1: Comparison between binary and grayscale methods using IFHCDB dataset.

SCCT Thome Zhang

93.6% 92.9% 93.3%

Table 2: Parameter tuning for the three classiﬁers.

Weighted SVM Weighted RF KNN

Kernel = Linear Max−depth = 50 Distance=Levenshtein(Levenshtein

et al., 1966)

IFHCDB C = 16 Number of estimators = 900 K = 5

AIA9K C=12 Number of estimators = 900 K = 7

6 EXPERIMENTAL RESULTS

Our contribution lies in the use of the SCCT skele-

tonization and the decision fusion of the three classi-

ﬁers. Thus, in this section, we evaluate the results of

the two contributions separately.

6.1 SCCT Contribution

In this part, a comparison between binary and

grayscale skeletonization is made on the IFHCDB

database. Since the Freeman chain code is the de-

scriptor that fully exploits the skeleton: its shape,

endpoints and junctions, we choose to use it for test-

ing the efﬁciency of the SCCT skeletonization and

thus to implement the K-NN classiﬁer. Alongside the

SCCT, two binary methods are used in the compar-

ison : Thome skeletonization (Merad et al., 2010)

and Zhang skeletonization (Zhang and Suen, 1984).

The adjustment of the SCCT parameters regarding the

contrast and noise of the image aims at ﬁnding a com-

promise between preserving the topology of the ob-

ject and removing noise related information. For this

purpose, the two following parameters must be set in

the SCCT skeletonization technique:

• The standard deviation of the background noise:

Since all the images in the dataset are acquired un-

der the same conditions, we can precalculate noise

standard deviation empirically by choosing a re-

gion from a set of image background, on which

we calculate the standard deviation.

• Test conﬁdence level: This parameter is inti-

mately linked to contrast level. According to the

authors (Youssef et al., 2016), and for images that

are correctly contrasted, a conﬁdence level of 0.01

is adequate.

According to Table 1, the results provided by

SCCT skeletonization comfort us in our choice, since

we can remove all preprocessing steps and, at the

same time, improve classiﬁcation results. By doing

so, the risk of damaging signiﬁcant information is re-

duced. These results support our ﬁrst observations in

Figure 2.

6.2 Classiﬁers Decision Fusion

Contribution

Parameter Tuning: Cross-validation was used to

compare the performance of different predictive mod-

els: weighted SVM, weighted RF and K-NN. A 5-

fold cross-validation was conducted on the training

dataset in order to choose the best parameters conﬁg-

uration. Table 2 details the chosen parameters values

for each model on respective dataset. All parameters

were chosen empirically.

Results: The global accuracy for the three clas-

siﬁers are presented in Table 3. Concerning the

IFHCDB dataset, we need to further merge class ”ba”

and ”fa” since we notice an important confusion be-

tween the two classes: 18% from class ”fa” were pre-

dicted belonging to class ”ba”. In addition, we also

merge classes 2 and 15, classes 10 and 11, in the

AIA9K data for the same reason.

Globally, the IFHCDB dataset results are better than

those of the AIA9k database due to size issue.

IFHCDB dataset contains 50k images while AIA9K

presents only 8k images. According to Table 3,

weighted SVM works better with unbalanced data,

while RF gives higher results with small data. KNN

has the lowest accuracy, but provides different infor-

mation since it uses another type of feature, the ’Free-

man chain code’.

Table 3: Accuracy of different classiﬁers.

Models IFHCDB AIA9K

SVM 97.88% 87.56%

RF 96.92% 91.08%

KNN 94.43% 84.83%

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

Table 4: Accuracy results on IFHCDB and AIA9K dataset of separate classiﬁers compared to the decision fusion approach.

IFHCDB AIA9K

Class ”ha” Class ”ain” Class ”ba+noun” Class ”kef”

KNN 97% 94% 93% 62%

Weighted SVM 99% 91% 90% 79%

Weighted RF 100% 93% 88% 77%

Fusion 100% 96% 92% 89%

Table 5: Accuracy compared to the literature.

Papers Classiﬁer Database Number of

classes

Accuracy

(Alaei et al., 2012) SVM IFHCDB 32-classes 96.91%

(Boulid et al., 2017) Neural Network IFHCDB 28-classes 96%

(KO and Poruran,

2020)

CNN IFHCDB 28-classes 96.3%

Proposed model Fusion of classiﬁers’

decisions

IFHCDB 28-classes 97.97%

(Balaha et al., 2021a) Deep learning sys-

tem

AIA9K 28-classes 93.3%

Proposed model Fusion of classiﬁers’

decisions

AIA9K 28-classes 92.91%

In fact, and according to results exposed in Ta-

ble 4 we notice that each classiﬁer succeeds/fails in

different situations. For example, in the case of the

IFHCDB dataset, the global accuracy of K-NN is

more interesting on letter ”ain”, while RF achieves

100% on letter ”ha”. In the case of the AIA9K dataset,

same remark can be made since we notice a 10%

improvement in classiﬁcation result when using fu-

sion on class ”kef”, compared to the best classiﬁer

(Weighted SVM). This reveals that the classiﬁcation

error is different from a classiﬁer to another and thus,

support the idea of merging the decisions of the three

machine learning approaches.

The proposed decision fusion contribution has im-

proved the accuracy up to 98.74% for the IFHCDB

dataset, which is a signiﬁcant result. The accuracy

rate of different classes varies from 85% to 100%.

Almost all the classes have an accuracy greater than

94%. And, a F

-score of 97.5% is obtained. For the

AIA9K dataset, the fusion result gives an accuracy of

94.58% and an F

-score of 94.5%, which are interest-

ing results regarding the limited size of the dataset.

In order to process the 28 classes, we moved to

the second stage where features related to the dia-

critics are used: number, size, and position. The

Weighted SVM classiﬁer is implemented to separate

similar character bodies. In this step, the separation is

done with 100% of precision in most classes. How-

ever, there is an issue with very similar characters’

diacritics, such as letter ’ba’ and letter ’tha’, leading

to a small decrease in overall accuracy : 97.97% for

IFHCDB and 92.91% for AIA9K.

Comparison with State of the Art: Table 5

presents our results and the ones of previously cited

works, for instance, deep learning models (Boulid

et al., 2017; KO and Poruran, 2020) and SVM clas-

siﬁer of (Alaei et al., 2012). Since we performed the

classiﬁcation on the same datasets as the above cited

works, we use in this comparison the results detailed

in their respective papers.

By mixing structural and statistical features ex-

tracted from both the character body and the skele-

ton, and by combining traditional classiﬁers to bring

out the best of each, we obtain the highest accuracy

of 97.97% among the cited methods from literature.

Another important remark is that these results are ob-

tained on highly unbalanced dataset without using any

data augmentation technique. One can also notice

that the separation of merged classes in the case of

AIA9K dataset did not beneﬁt the overall accuracy,

since some of the separated classes has really few ob-

servations, which declined the overall accuracy.

7 CONCLUSIONS

In this paper, a method to recognize handwritten Ara-

bic character is presented. The proposed approach in-

cludes a data analysis step to extract each character’s

most accurate descriptors and a classiﬁcation step. In

An Ensemble Learning Approach using Decision Fusion for the Recognition of Arabic Handwritten Characters

the ﬁrst step, a new technique of skeletonization is

used to improve the feature extraction phase. In the

modelling step, a new classiﬁcation method is pro-

posed resulting in an interesting accuracy rate com-

pared to separate classiﬁers and Deep Learning ar-

chitectures when tested on the IFHCDB dataset. A

data augmentation technique should be done in future

works to improve the result on AIA9K dataset due to

its small size.

REFERENCES

Alaei, A., Pal, U., and Nagabhushan, P. (2012). A com-

parative study of persian/arabic handwritten charac-

ter recognition. In 2012 International Conference

on Frontiers in Handwriting Recognition, pages 123–

128. IEEE.

Ali, A. A. A., Suresha, M., and Ahmed, H. A. M. (2020).

A survey on arabic handwritten character recognition.

SN Computer Science, 1(3):1–10.

Althobaiti, H. and Lu, C. (2017). A survey on arabic opti-

cal character recognition and an isolated handwritten

arabic character recognition algorithm using encoded

freeman chain code. In 2017 51st Annual Conference

on Information Sciences and Systems (CISS), pages 1–

6. IEEE.

Altwaijry, N. and Al-Turaiki, I. (2021). Arabic hand-

writing recognition system using convolutional neu-

ral network. Neural Computing and Applications,

33(7):2249–2261.

Ayodele, T. O. (2010). Types of machine learning algo-

rithms. New advances in machine learning, 3:19–48.

Balaha, H. M., Ali, H. A., Saraya, M., and Badawy, M.

(2021a). A new arabic handwritten character recogni-

tion deep learning system (ahcr-dls). Neural Comput-

ing and Applications, 33(11):6325–6367.

Balaha, H. M., Ali, H. A., Youssef, E. K., Elsayed, A. E.,

Samak, R. A., Abdelhaleem, M. S., Tolba, M. M.,

Shehata, M. R., Mahmoud, M. R., Abdelhameed,

M. M., et al. (2021b). Recognizing arabic handwritten

characters using deep learning and genetic algorithms.

Multimedia Tools and Applications, pages 1–37.

Borovikov, E. (2014). A survey of modern optical

character recognition techniques. arXiv preprint

arXiv:1412.4183.

Bosowski, P., Bosowska, J., and Nalepa, J. (2021). Evolving

deep ensembles for detecting covid-19 in chest x-rays.

In 2021 IEEE International Conference on Image Pro-

cessing (ICIP), pages 3772–3776. IEEE.

Boufenar, C., Batouche, M., and Schoenauer, M. (2018).

An artiﬁcial immune system for ofﬂine isolated hand-

written arabic character recognition. Evolving Sys-

tems, 9(1):25–41.

Boulid, Y., Souhar, A., and Elkettani, M. Y. (2017). Hand-

written character recognition based on the speciﬁcity

and the singularity of the arabic language. Interna-

tional Journal of Interactive Multimedia & Artiﬁcial

Intelligence, 4(4).

El-Sawy, A., Loey, M., and El-Bakry, H. (2017). Ara-

bic handwritten characters recognition using convolu-

tional neural network. WSEAS Transactions on Com-

puter Research, 5:11–19.

Heutte, L., Paquet, T., Moreau, J.-V., Lecourtier, Y., and

Olivier, C. (1998). A structural/statistical feature

based vector for handwritten character recognition.

Pattern recognition letters, 19(7):629–641.

Hu, M.-K. (1962). Visual pattern recognition by moment

invariants. IRE transactions on information theory,

8(2):179–187.

Kaoudja, Z., Kherﬁ, M. L., and Khaldi, B. (2019). An efﬁ-

cient multiple-classiﬁer system for arabic calligraphy

style recognition. In 2019 International Conference

on Networking and Advanced Systems (ICNAS), pages

1–5. IEEE.

KO, M. A. and Poruran, S. (2020). Ocr-nets: Variants of

pre-trained cnn for urdu handwritten character recog-

nition via transfer learning. Procedia Computer Sci-

ence, 171:2294–2301.

Levenshtein, V. I. et al. (1966). Binary codes capable of cor-

recting deletions, insertions, and reversals. In Soviet

physics doklady, volume 10, pages 707–710. Soviet

Union.

Lutf, M., You, X., and Li, H. (2010). Ofﬂine arabic

handwriting identiﬁcation using language diacritics.

In 2010 20th International Conference on Pattern

Recognition, pages 1912–1915. IEEE.

Mallat, K. and Youssef, R. (2016). Adaptive morphologi-

cal closing based on inertia tensor for structuring ele-

ment estimation. In 2016 International Symposium on

Signal, Image, Video and Communications (ISIVC),

pages 253–258. IEEE.

Merad, D., Aziz, K.-E., and Thome, N. (2010). Fast people

counting using head detection from skeleton graph. In

2010 7th IEEE International Conference on Advanced

Video and Signal Based Surveillance, pages 233–240.

IEEE.

Mozaffari, S., Faez, K., Faradji, F., Ziaratban, M., and

Golzan, S. M. (2006). A comprehensive isolated

farsi/arabic character database for handwritten ocr re-

search. In Tenth International Workshop on Frontiers

in Handwriting Recognition. Suvisoft.

Rajabi, M., Nematbakhsh, N., and Monadjemi, S. A.

(2012). A new decision tree for recognition of per-

sian handwritten characters. International Journal of

Computer Applications, 44(6):52–58.

Rashad, M. and Semary, N. A. (2014). Isolated printed ara-

bic character recognition using knn and random forest

tree classiﬁers. In International Conference on Ad-

vanced Machine Learning Technologies and Applica-

tions, pages 11–17. Springer.

Sahlol, A., Abd Elfattah, M., Suen, C. Y., and Hassanien,

A. E. (2016). Particle swarm optimization with ran-

dom forests for handwritten arabic recognition sys-

tem. In International Conference on Advanced In-

telligent Systems and Informatics, pages 437–446.

Springer.

Sahlol, A. T., Suen, C. Y., Elbasyoni, M. R., and Sallam,

A. A. (2014). Investigating of preprocessing tech-

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

niques and novel features in recognition of handwrit-

ten arabic characters. In IAPR Workshop on Artiﬁcial

Neural Networks in Pattern Recognition, pages 264–

276. Springer.

Siddhu, M. K., Parvez, M. T., and Yaakob, S. N. (2019).

Combining statistical and structural approaches for

arabic handwriting recognition. In 2019 International

conference on computer and information sciences (IC-

CIS), pages 1–6. IEEE.

Torki, M., Hussein, M. E., Elsallamy, A., Fayyaz, M.,

and Yaser, S. (2014). Window-based descriptors

for arabic handwritten alphabet recognition: a com-

parative study on a novel dataset. arXiv preprint

arXiv:1411.3519.

Youssef, R., Sevestre-Ghalila, S., Ricordeau, A., and Be-

nazza, A. (2016). Self noise and contrast controlled

thinning of gray images. Pattern Recognition, 57:97–

114.

Zanchettin, C., Bezerra, B. L. D., and Azevedo, W. W.

(2012). A knn-svm hybrid model for cursive hand-

writing recognition. In The 2012 International Joint

Conference on Neural Networks (IJCNN), pages 1–8.

IEEE.

Zhang, T. Y. and Suen, C. Y. (1984). A fast parallel algo-

rithm for thinning digital patterns. Communications

of the ACM, 27(3):236–239.

Zhao, H.-h. and Liu, H. (2020). Multiple classiﬁers fu-

sion and cnn feature extraction for handwritten digits

recognition. Granular Computing, 5(3):411–418.

An Ensemble Learning Approach using Decision Fusion for the Recognition of Arabic Handwritten Characters