Multi-stage Off-line Arabic Handwriting Recognition Approach using

Advanced Cascading Technique

Taraggy Ghanim

1,2

, Mahmoud I. Khalil

and Hazem M. Abbas

Faculty of Computer Science, Misr International University, Egypt

Faculty of Engineering, Ain Shams University, Egypt

Keywords:

Arabic Handwriting Recognition, Random Forest, Kullback-Leibler Divergence, Pyramid Histogram of

Gradient, Support Vector Machine.

Abstract:

Automatic Recognition of Arabic Handwriting is a pervasive ﬁeld that has many challenging complications to

solve. Such complications include big databases and complex computing activities. The proposed approach

is a multi-stage cascading recognition system bases on applying Random Forest classiﬁer (RF) to construct

a forest of decision trees. The constructed decision trees split big databases to multiple smaller data-mined

sets based on the most discriminating computed geometric and regional features. Each data-mined set include

similar database classes. RF match each test image with one of the data-mined sets. Afterwards, the matching

classes are sorted relative to test image using Pyramid Histogram of Gradients and Kullback-Leibler based

ranking algorithm. Finally, the classiﬁcation process is applied on the highly ranked matching classes to

assign a class membership to test image. Adjusting the classiﬁcation process to only consider the highly

ranked database classes reduced the computing classiﬁcation and enhanced the overall performance. The

proposed approach was tested on IFN-ENIT Arabic database and achieved satisfactory results and enhanced

sensitivity of decision trees to reach 93.5% instead of 86.5% (Ghanim et al., 2018).

1 INTRODUCTION

Automatic Arabic Handwriting recognition is a chal-

lenging computer vision application. It is useful for

analyzing and digitizing handwritten documents, re-

serving and documenting old manuscripts. It requires

building robust hybrid classiﬁers that merge differ-

ent but complementary methods to achieve high au-

tomatic recognition rates.

Handwritten recognition process is categorized

into online or ofﬂine process(Amin, 1998). Ofﬂine

recognition is based solely on visual images and pixel

information which is more challenging and concerned

in our work.

Arabic is a cursive language and one of the ma-

jor worldwide document sources (Amin, 1998). Lan-

guage speciﬁcations and description are provided by

(Abandah and Khedher, 2009) and (AbdelRaouf et al.,

2008). It is the native language of more than 420 mil-

lion people around the world, the sixth most spoken

language, and used in around 27 different languages

(Campbell and Grondona, 2008) and can be repre-

sented in different handwriting styles.

The proposed approach is divided into three cas-

caded stages. The ﬁrst is matching each test image

with set of database classes. The second is ranking

the set of matching classes and ﬁnally classiﬁcation as

described and analyzed in Section 3.In the introduced

approach, recognition is done without character seg-

mentation to decrease computing time and errors due

to wrong segmentation.

Previous related work is summarized in section 2.

The approach is described in section 3. The experi-

ments and analysis of achieved results are proposed

in section 4 . Conclusion and Future work are ﬁnally

in sections 5 and 6.

2 LITERATURE REVIEW

Lawgali (Lawgali, 2015) provided a survey on Arabic

handwritten automatic recognition. (Ghanim et al.,

2018) summarized the different applied classiﬁers in

this research area.

Hidden Marcov Model (HMM) was applied by

(Hicham et al., 2016), (AlKhateeb et al., 2011),

(Jayech et al., 2015) and (Dreuw et al., 2011) on

IFN/ENIT Arabic database (Pechwitz et al., 2002).

532

Ghanim, T., Khalil, M. and Abbas, H.

Multi-stage Off-line Arabic Handwriting Recognition Approach using Advanced Cascading Technique.

DOI: 10.5220/0007374605320539

In Proceedings of the 8th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2019), pages 532-539

ISBN: 978-989-758-351-3

Localized density features and statistical-type fea-

tures (Hicham et al., 2016) were extracted achieving

78.95% recognition rate. Intensity features on mir-

rored word image (MWI) (AlKhateeb et al., 2011)

were extracted and achieved 86.73% success rate

without re-ranking and 89.24% with re-ranking. Sta-

tistical and structural features (Jayech et al., 2015)

achieved 91.1% recognition rates. Multi-Stream

Hidden Markov Model was applied for handwriting

recognition (Mezghani et al., 2014) using Gaussian

Mixture Models (GMMs) to recognize Arabic and

French handwritten words.

Support Vectored Machines (SVMs) achieved ef-

fective results in this ﬁeld when applied on the

IFN/ENIT database. (Elleuch et al., 2016) applied

convolution neural network (CNN) and SVM to clas-

sify only 56 classes with error rate 7.05%. (Al-

Dmour and Abuhelaleh, 2016) proposed a multi-stage

model using SURF feature descriptor and K-means

clustering algorithm to recognize 85% using SVM.

(Khaissidi et al., 2016) applied Histograms of Ori-

ented Gradients (HoGs) with SVM on Ibn-Sina data-

set (Moghaddam et al., 2010).

(Sa

ıdani and Echi, 2014) extracted a combination

of Pyramid HOG and co-occurrence Matrix of HOG.

(Elfakir et al., 2015) extracted HOG features using

Sobel edge detectors. SVM was applied on the nor-

malized features for classiﬁcation.

Different types of Neural Networks (NN) were

applied in this area. Pseudo-Zernike moments fea-

tures (Leila et al., 2011) with Fuzzy ARTMAP neu-

ral network classiﬁed 96 word classes written by hun-

dred writers with 93.8% recognition rate. (Lawgali

et al., 2014) extracted Discrete Cosine Transform

(DCT) features and applied Artiﬁcial Neural Net-

work to classify a subset of IFN/ENIT database with

90.73% accuracy. (Benjelil et al., 2012) used a steer-

able pyramid decomposition method with k-NN clas-

siﬁer and achieved 97.5% success rate on a database

of 800 printed and handwritten words.

Random forest (RF) was applied for handwriting

recognition. (Shamim et al., 2018) presented a com-

parative study for ofﬂine digit recognition and SVM

with RF classiﬁer achieved highest results. (Do and

Pham, 2015) computed GIST features with RF on

USPS, MNIST data-sets. (Zamani et al., 2015) ap-

plied RF and convolutional neural network (CNN)

for Persian handwritten digit recognition. The sys-

tem was tested on Hoda data-set (Khosravi and Kabir,

2007). The concept of cascading classiﬁers was ap-

plied on different recognition problems and achieved

satisfactory results (Ghanem et al., 2009), (Mohamed

et al., 2018).

3 THE PROPOSED APPROACH

In the proposed approach three main consecutive

stages are applied as presented in Figure 1. All stages

are complementary to each other. The output of

each stage passes as an input to the next subsequent

stage. Each stage passes smaller set of chosen training

classes to its next consecutive stage. Reducing num-

ber of concerned database classes per stage eases the

classiﬁcation task and equate classiﬁer growing com-

plexity.

Figure 1: The proposed system overview.

Multi-stage Off-line Arabic Handwriting Recognition Approach using Advanced Cascading Technique

533

3.1 Stage 1: Matching Test Image with

Similar Database Training Classes

This stage match each test image with a set of sim-

ilar database classes. First, preprocessing and seg-

mentation are applied (section 3.1.1), a set of features

(Dileep, 2012) are then computed (section 3.1.2).

Random Forest classiﬁer vote for the optimum set of

computed features and match each test image with a

set of database classes. Matching sets include similar

database classes together and is represented by a de-

ﬁned range of selected feature values. This introduces

a data-mined database.

3.1.1 Preprocessing and Segmentation

The ﬁrst stage starts by passing each test image and

training database classes through preprocessing and

segmentation as shown in Figure 1. It is an essential

stage for consistent post analysis and classiﬁcation.

Images are binarized, cropped, normalized and

thinned to one pixel wide to remove variations in

handwritten images as shown in Figure 2. Binariza-

tion concentrates computations on regions of interest.

(a) original image

(b) cropped image

(d) image thinning

Figure 2: A sample image during preprocessing.

3.1.2 Regional and Geometric Features

Extraction

A set of regional and geometry based features

(Dileep, 2012) are computed. Based on the applied

feature selection technique (section 3.1.3), Extent fea-

ture is an effective regional feature that represents the

normalized area of the skeleton. It is a scalar mea-

surement that speciﬁes the ratio of handwritten word

area to area of the word imaginary bounding ellipse

as shown in Figure3.

Figure 3: Word Bounding Ellipse.

The geometric features; descriptors of image con-

tours, are computed after dividing image into six dif-

ferent zones (Dileep, 2012). Zoning considers the po-

sition of line segments as a feature and satisfy the con-

cept of feature localization. Eight features are com-

puted per zone. First four are the normalized number

N of horizontal, vertical, left diagonal, right diagonal

lines, as deﬁned in equation 1.

N = 1 −

(1)

The second four features are the normalized length

of all line types as deﬁned in equation 2.

L =

(2)

where A

is the zone area. Line types are deter-

mined using the concavity features (Theodoridis and

Koutroumbas, 2006) as shown in Figure 4.

Figure 4: Concavity features.

3.1.3 Random Forest for Feature Selection and

Decision Trees Construction

Random Forest classiﬁer perform feature selection to

choose most effective features from the computed set

of regional and geometric features. A forest of deci-

sion trees are then constructed to vote for a group of

similar classes corresponding to each test image.

Database classes are split into two branches at

each tree node. Splitting is based on features’ impor-

tance. Importance of features is estimated from their

impact on model accuracy. The measured importance

of the computed features is shown in Figure 5. The

x-axis is the feature, values from 1 till 5 are the com-

puted regional features including number of PAWs,

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

534

Figure 5: Feature Importance.

number of holes, eccentricity, extent and orientation

respectively , other values are the computed set of ge-

ometric features (Dileep, 2012). It is clearly shown

from Figure 5 that extent;feature 4, is the most rele-

vant regional features, while the whole set of geomet-

ric features are all important and affect performance

positively.

Finally, based on the measured features impor-

tance, the selected set of features includes extent and

the 8 geometric features. Based on the selected fea-

tures, Apriori pruning algorithm (Rokach and Mai-

mon, 2014) match each test sample with a set of sim-

ilar database classes; called item-sets.

It is a challenging aspect to mine item-sets from

large database (Han et al., 2011). If a long item-set is

frequent, then all its subsets are frequent as well. For

example, a k-item-set ( i.e. of length k items ) contains

total number of frequent item-sets deﬁned by









+ ··· +





= 2

− 1 (3)

Each matching set include similar classes with

similar features. Different sets may include common

classes. Some smaller sets may be a subset of other

bigger sets which serve the matching process to be

implemented as a binary search tree. The test image

is now matched with a set of similar database classes

and both are ready for ranking stage.

3.2 Stage 2: Ranking Stage (PHoG &

KL-divergence)

Each test image and its matching set of similar classes

pass through stage 2. Kullback-Leibler divergence

(Ghanim et al., 2018) is computed between Pyra-

mid Histogram of Gradient features of input images

(Sa

ıdani and Echi, 2014). Accordingly, the match-

ing classes are ranked from the nearest to the furthest

(Ghanim et al., 2018) relative to test image. The ob-

jective of this stage is to pass only subset of highly

ranked classes to classiﬁcation stage as presented in

Figure 1.

3.2.1 Statistical Features Extraction: Pyramid

Histogram of Gradients PHoG

Histogram of oriented gradients HoGs is a statistical-

type descriptor that performs orientation analysis at

different levels (Ghanim et al., 2018). It captures ﬁne

details and more discriminating information about

words skeletons than the ordinary HoGs (Sa

ıdani and

Echi, 2014).

Oriented gradients are extracted per image zone as

shown in Figure 6 by Canny edge detector (Gonzalez

and Woods, 2007) as deﬁned in equation 4.

θ = arctan

(4)

Gx and Gy are gradients in x and y direction, θ is ori-

entation [0,360

◦

] which are discretized to eight val-

ues (Saidani et al., 2015). PHoG features are com-

puted and normalized per image zone for the 8 de-

ﬁned angular bins. KL-divergence is computed be-

tween feature vectors of test image and its matching

classes 3.2.2.

3.2.2 Divergence Measure: Kullback-Leibler

(KL)

KL divergence measure has been popularly used in

data-mining (Ghanim et al., 2018). It is a non-

symmetric measure of difference between any two

probability distributions; as PHoGs, for orientation

analysis. It is a non-symmetric metric measure. The

measure from one distribution q(x) to another p(x) is

not equal to the measure from q(x) to p(x). It is a

non-negative value that equals to zero if and only if

p(x) = q(x).

This measurement ranks the members of matching

set from the nearest to the furthest relative to the test

image. Classes of high ranks pass to ﬁnal classifying

stage instead of passing all database.

3.3 Stage 3: Classiﬁcation Stage

Finally each test image pass through classiﬁcation

stage with only highly ranked best nearest neighbors

classes to vote for ﬁnal membership class.

3.3.1 Final Feature Vector: Statistical &

Geometric Features Extraction

Final feature vector is a combination of the selected

features from stage 1 and the PHoG features from

stage 2. Final features vector is used to classify the

test image as described next section 3.3.2.

Multi-stage Off-line Arabic Handwriting Recognition Approach using Advanced Cascading Technique

535

(a) Level=1 (b) Level=2 (c) Level=3

Figure 6: Pyramid histogram of gradients (F:frequency, G:oriented gradient).(Ghanim et al., 2018)

3.3.2 Classiﬁcation: Multi-class Support Vector

Machines

A bigger feature vector is now ready to train a multi-

class SVM. Only the highly ranked best matching ref-

erence classes are used in training SVM. The output

ﬁnally determines the test image membership training

class.

Many surveys (Kumar and Rao, 2013) demon-

strated effectiveness of SVM. There are two schemes

to convert SVM to series of binary SVM’s (Abdi-

ansah and Wardoyo, 2015), one-versus-rest and one-

versus-one. In the proposed solution, the one-versus-

one approach is applied. It is symmetric SVM model

and does not suffer from the unbalanced classiﬁcation

problem.

4 EXPERIMENTS

4.1 Database

The approach is tested on IFN/ENIT database (Pech-

witz et al., 2002) of 937 distinct classes. Data-set is

composed of 5 parts named a, b, c, d and e.

4.2 Matching Set Selection

Matching sets vary in sizes. Sensitivity is measured

by degree of correct class inclusion in the matching

set. Random Forest for decision trees construction

enhanced sensitivity to be 93.5% instead of 86.5%

in (Ghanim et al., 2018). The weighted average

matching set size computed from equation 5 is ap-

proximately 172.6 different classes, which is approx-

imately 18% of the whole database number of classes.

¯x =

∑

i=1

∗ w

)

∑

i=1

(5)

¯x is the weighted average, x is the set size and w is the

normalized weights of sets. Largest set contains 423

different classes (worst case) which is 44.7% of the

946 total database classes.

Figure 7 shows the effect of the different regional

and geometric features and Random forest (RF) on

the overall matching error. Figure 7a shows that num-

ber of holes, PAWs, eccentricity and word orientation

cause 21% misclassiﬁcation error. Figure 7b shows

that Extent feature cause an exponential reduction in

misclassiﬁcation error to be 9%. Including the ge-

ometric features to Extent feature reduce error from

9% to 6%, Figure 7c.

4.3 The Ranking Stage

The approach categorized database into three main

parts;labelled 30, 20 and 5, according to average ra-

tio between training and testing samples. Figure 8

shows that increasing PHoG level causes proceeding

of correct class in early ranks. This cause fast graph

saturation where high recognition rates are achieved.

Fastest saturation achieved with label 30 due to ex-

istence of enough training samples relative to testing

ones. This leads to passing as small sets of classes

as possible to ﬁnal stage and so less training time as

shown in Figure 8b and 8c. The correct class rank

in worst case is 100 inside matching set. Accord-

ingly, only 10.5% of the total database classes passes

to classiﬁcation stage.

4.4 The Classiﬁcation Stage

The SVM is applied with different kernel’s type at

different levels of Pyramid Histogram of gradients as

shown in Figure 9.

Final classiﬁcation is applied on highly ranked

classes. The Linear SVM with level 5 PHoG achieves

the highest recognition rate. Figure 9a shows SVM

output with Level 1 PHoG; original HOG. Figure 9b

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

536

(a) no. of holes, paws and eccentricity

and orientation.

(b) Effect of extent after excluding

no. of holes, PAWS, ecc, orientation.

tures.

Figure 7: Effect of selected features on Classiﬁcation Error, (N=number of decision trees), (ER% = error rate).

(a) Experiment on classes with label

30.

(b) Experiment on classes with label

20.

Figure 8: Effect of Ranking on the Expected Success Rate (Ghanim et al., 2018) (TPR: True positive rate), (R: rank of correct

class).

(a) Level 1 PHoG and Geometric fea-

tures with SVM.

(b) Level 3 PHoG and Geometric fea-

tures with SVM.

Figure 9: System Recognition Rate.

is for level 3 PHoG. The linear kernel gives the best

results with all sets but the polynomial kernel was bet-

ter than quadratic. Figure 9c displays the system per-

formance on level 5 PHoG. The linear kernel with set

labeled “30” outperforms all the other experiments.

4.5 Comparative Experiments and

Results

Table 1 represents similar systems’ experiments and

their results relative to proposed approach. Some

classiﬁcation systems applied word segmentation to

character level, while others didn’t. Segmentation

modify error rates by predicting the correct sequence

of word characters based on pre-deﬁned dictionary

but increase computation complexity. The proposed

approach is applied without segmentation and consid-

ered absence of pre-deﬁned dictionary. The results in

table 1 indicate the distinction of the proposed solu-

tion results than others.

4.6 Reasons of Misclassiﬁcation

Visual similarity between different classes is one of

the main reasons of misclassiﬁcation as shown in ta-

ble 2.

Multi-stage Off-line Arabic Handwriting Recognition Approach using Advanced Cascading Technique

537

Table 1: Comparative results of systems applied on IFN-ENIT.

Approach reference Training-testing Approach Segmentation Word error rate

(Abandah et al., 2014) abcd-e BLSTM-RNN without 24.28%

(Lawgali et al., 2014) abcd-e DCT-NN with 14.64%

(Jayech et al., 2015) abc-d MSHMM with 8.95%

without 16.9%

abcd-e without 27.6%

with 17.09%

(Hicham et al., 2016) abcd-e denstity-HMM with 21.05%

(Elleuch et al., 2016) 56 classes Conv-CNN SVM with 7%

(Al-Dmour and Abuhelaleh, 2016) 18 classes Surf-SVM without 15%

Proposed approach abc-d RF-PHOG-KL-SVM Without 3.6%

abcd-e Without 14.6%

Table 2: Sample of visual similarity between classes.

Class name Similar Class name

Ñm

'@ Õæ









G@ñË@ ú



æ@ñË@

èP@Q



«ñK

èX@Q«ñK

(a) Misclassiﬁcation due to bad handwriting styles

(b) Misclassiﬁcation due to inaccurate thinning results

Figure 10: Samples of Misclassiﬁed Images.

Bad handwriting styles and inaccurate thinning

also mislead the system as shown in Figure 10a and

10b. Misclassiﬁcation is sometimes due to lack of

training samples of some classes.

5 CONCLUSIONS

The proposed approach is a cascaded classiﬁer for of-

ﬂine Arabic handwritten recognition. It is applied on

IFN/ENIT database. Random Forest improved per-

formance of decision trees. High levels of PHoG in-

crease effectiveness of ranking stage.

6 FUTURE WORK

Future plan is to apply deep learning techniques to

improve classiﬁcation rates and matching process to

build more robust recognition systems.

REFERENCES

Abandah, G. A., Jamour, F. T., and Qaralleh, E. A.

(2014). Recognizing handwritten Arabic words us-

ing grapheme segmentation and recurrent neural net-

works. International Journal on Document Analysis

and Recognition (IJDAR), 17(3):275–291.

Abandah, G. A. and Khedher, M. Z. (2009). Analysis of

handwritten Arabic letters using selected feature ex-

traction techniques. International Journal of Com-

puter Processing of Languages, 22(01):49–73.

AbdelRaouf, A., Higgins, C. A., and Khalil, M. (2008). A

database for Arabic printed character recognition. In

International Conference Image Analysis and Recog-

nition, pages 567–578. Springer.

Abdiansah, A. and Wardoyo, R. (2015). Time complex-

ity analysis of support vector machines (SVM) in lib-

svm. International Journal Computer and Applica-

tion, 128(3).

Al-Dmour, A. and Abuhelaleh, M. (2016). Arabic hand-

written word category classiﬁcation using bag of fea-

tures. Journal of Theoretical and Applied Information

Technology, 89(2):320.

AlKhateeb, J. H., Ren, J., Jiang, J., and Al-Muhtaseb, H.

(2011). Ofﬂine handwritten Arabic cursive text recog-

nition using hidden markov models and re-ranking.

Pattern Recognition Letters, 32(8):1081–1088.

Amin, A. (1998). Off-line Arabic character recognition: the

state of the art. Pattern recognition, 31(5):517–530.

Benjelil, M., Mullot, R., and Alimi, A. M. (2012). Lan-

guage and script identiﬁcation based on steerable

pyramid features. In Frontiers in Handwriting Recog-

nition (ICFHR), pages 716–721. IEEE.

Campbell, L. and Grondona, V. (2008). Ethnologue: Lan-

guages of the world. Language, 84(3):636–641.

Dileep, D. (2012). A feature extraction technique based on

character geometry for character recognition. Depart-

ment of Electronics and Communication Engineering,

Amrita School of Engineering, Kollam, India.

Do, T.-N. and Pham, N.-K. (2015). Handwritten digit recog-

nition using gist descriptors and random oblique deci-

sion trees. In Some Current Advanced Researches on

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

538

Information and Computer Science in Vietnam, pages

1–15. Springer.

Dreuw, P., Heigold, G., and Ney, H. (2011). Conﬁdence-and

margin-based MMI/MPE discriminative training for

off-line handwriting recognition. International jour-

nal on document analysis and recognition, 14(3):273–

288.

Elfakir, Y., Khaissidi, G., Mrabti, M., and Chenouni, D.

(2015). Handwritten Arabic documents indexation us-

ing HOG feature. International Journal of Computer

Applications, 126(9).

Elleuch, M., Maalej, R., and Kherallah, M. (2016). A new

design based-SVM of the CNN classiﬁer architecture

with dropout for ofﬂine Arabic handwritten recogni-

tion. Procedia Computer Science, 80:1712–1723.

Ghanem, T. M., Moustafa, M. N., and Shahein, H. I. (2009).

Gabor wavelet based automatic coin classsiﬁcation. In

2009 16th IEEE International Conference on Image

Processing (ICIP), pages 305–308. IEEE.

Ghanim, T. M., Khalil, M. I., and Abbas, H. M. (2018).

Phog features and kullback-leibler divergence based

ranking method for handwriting recognition. In IAPR

Workshop on Artiﬁcial Neural Networks in Pattern

Recognition, pages 293–305. Springer.

Gonzalez, R. C. and Woods, R. E. (2007). Digital image

processing. Pearson.

Han, J., Pei, J., and Kamber, M. (2011). Data mining: con-

cepts and techniques. Elsevier.

Hicham, E. M., Akram, H., and Khalid, S. (2016). Using

features of local densities, statistics and HMM toolkit

(htk) for ofﬂine Arabic handwriting text recognition.

Journal of Electrical Systems and Information Tech-

nology, 3(3):99–110.

Jayech, K., Mahjoub, M. A., and Amara, N. E. B. (2015).

Arabic handwriting recognition based on synchronous

multi-stream HMM without explicit segmentation.

In International Conference on Hybrid Artiﬁcial In-

telligence Systems, volume 9121, pages 136–145.

Springer.

Khaissidi, G., Elfakir, Y., Mrabti, M., Lakhliai, Z., Che-

nouni, D., and El yacoubi, M. (2016). Segmentation-

free word spotting for handwritten Arabic documents.

International Journal of Interactive Multimedia and

Artiﬁcial Intelligence, 4.

Khosravi, H. and Kabir, E. (2007). Introducing a very large

dataset of handwritten farsi digits and a study on their

varieties. Pattern recognition letters, 28(10):1133–

1141.

Kumar, K. V. and Rao, R. R. (2013). Online handwritten

character recognition for telugu language using sup-

port vector machines. International Journal of Engi-

neering and Advanced Technology (IJEAT), 3.

Lawgali, A. (2015). A survey on Arabic character recog-

nition. International Journal of Signal Processing,

Image Processing and Pattern Recognition, 8(2):401–

426.

Lawgali, A., Angelova, M., and Bouridane, A. (2014). A

framework for Arabic handwritten recognition based

on segmentation. International Journal of Hybrid In-

formation Technology, 7(5):413–428.

Leila, C., Ma

amar, K., and Salim, C. (2011). Combining

neural networks for Arabic handwriting recognition.

In Programming and Systems (ISPS), 2011 10th Inter-

national Symposium on, pages 74–79. IEEE.

Mezghani, A., Slimane, F., Kanoun, S., and M

argner,

V. (2014). Identiﬁcation of Arabic/French hand-

written/printed words using gmm-based system. In

CORIA-CIFED, pages 371–374.

Moghaddam, R. F., Cheriet, M., M. Adankon, M., Filo-

nenko, K., and Wisnovsky, R. (2010). Ibn sina: a

database for research on processing and understand-

ing of Arabic manuscripts images. Proceeding DAS

’10 Proceedings of the 9th IAPR International Work-

shop on Document Analysis Systems, pages 11–18.

Mohamed, H., Omar, R., Saeed, N., Essam, A., Ayman, N.,

Mohiy, T., and AbdelRaouf, A. (2018). Automated de-

tection of white blood cells cancer diseases. In Deep

and Representation Learning (IWDRL), 2018 First In-

ternational Workshop on, pages 48–54. IEEE.

Pechwitz, M., Maddouri, S. S., M

argner, V., Ellouze, N.,

Amiri, H., et al. (2002). IFN/ENIT-database of hand-

written Arabic words. In Proc. of CIFED, volume 2,

pages 127–136.

Rokach, L. and Maimon, O. (2014). Data mining with deci-

sion trees: theory and applications. World scientiﬁc.

ıdani, A. and Echi, A. K. (2014). Pyramid histogram

of oriented gradient for machine-printed/handwritten

and Arabic/Latin word discrimination. In Soft Com-

puting and Pattern Recognition (SoCPaR), 2014 6th

International Conference of, pages 267–272. IEEE.

Saidani, A., Kacem, A., and Belaid, A. (2015). Arabic/Latin

and machine-printed/handwritten word discrimination

using HOG-based shape descriptor. ELCVIA Elec-

tronic Letters on Computer Vision and Image Anal-

ysis, 14(2):1–23.

Shamim, S., Miah, M. B. A., Sarker, A., Rana, M., and

Al Jobair, A. (2018). Handwritten digit recognition

using machine learning algorithms. Global Journal of

Computer Science and Technology.

Theodoridis, S. and Koutroumbas, K. (2006). Pattern

Recognition. Academic Press.

Zamani, Y., Souri, Y., Rashidi, H., and Kasaei, S. (2015).

Persian handwritten digit recognition by random for-

est and convolutional neural networks. In Machine Vi-

sion and Image Processing (MVIP), 2015 9th Iranian

Conference on, pages 37–40. IEEE.

Multi-stage Off-line Arabic Handwriting Recognition Approach using Advanced Cascading Technique

539