New Approach Based on Substantial Derivative and LSTM for

Online Arabic Handwriting Script Recognition

Hasanien Ali Talib Alothman

1,2,3

, Wafa Lejmi

1,3

and Mohamed Ali Mahjoub

ISITCom, Higher Institute of Computer Science and Communication Technologies of Sousse,

University of Sousse, 4011 Sousse, Tunisia

College of Education for Pure Science, Computer Science Department, University of Mosul, Iraq

LATIS - Laboratory of Advanced Technology and Intelligent Systems, University of Sousse, 4011 Sousse, Tunisia

Keywords: Handwriting, Arabic, Script, Text, Character, Descriptor, Substantial Derivative, Feature, Extraction,

Acceleration, ADAB Dataset, Recognition, LSTM.

Abstract: As some tasks easily performed by man seem to be hard to be accomplished by the machine, the Artificial

Intelligence field examines more and more the reproduction of thinking methods and human intuition by

studying some mental faculties and substituting them by calculating approaches. Among the major fields of

such interest, we can focus on recognizing handwritten characters. However, most handwritten characters

are written in Latin, which makes the recognition of Arabic characters handwriting a delicate process

compared to other languages, due to the specificity of Arabic words. In this paper, we aim to conceive a

framework that offers the ability to recognize online Arabic handwriting applied to a dataset named ADAB

(Arabic DAtaBase), using a particular descriptor based on a substantial derivative and a neural network

handling Arabic handwritten characters features and then electing the appropriate output for the final

decision.

1 INTRODUCTION

Handwriting recognition is among the oldest

problems faced by artificial intelligence, since its

advent in the 1950s (Mori et al., 1992). An essential

playground for new learning algorithms, it

represents a real scientific and technical challenge

and an imminent need requested by many business

sectors.

1.1 Applications

Recognizing handwritten text is being applied in

several various human activity fields, including:

 Education (Wu et al., 2021): through the

recognition and translation of texts, such as

texts in Braille, and writing learning.

Photosensors and tactile simulators are used

for the blind and low vision (BLV) persons

with a sound output.

 Banks and Insurance companies (Singh et al.,

2015): through the check authentication

(correspondence between amounts and

denomination, and between the identity of the

signatory and his signature), and verification

of insurance contracts.

 Post services (Charfi et al., 2012): through

reading postal addresses and automatic sorting

of mails.

 Business and Industry (Nagy, 2016): through

inventory management and technical

documents recognition (electronic diagrams,

technical drawings, architectural plans, etc.).

 Office automation (Chherawala & Cheriet,

2014): through indexing and automatic

archiving of documents, and for electronic

publication.

 Automatic reading of administrative

documents and recognition of cartographic

maps (Velázquez & Levachkine, 2004).

1.2 Modes of Character Writing

Two writing modes of character writing are used:

static mode for characters already written and

dynamic mode for handwritten characters to be

recognized while writing. Below, we explain both

modes in more detail:

Alothman, H., Lejmi, W. and Mahjoub, M.

New Approach Based on Substantial Derivative and LSTM for Online Arabic Handwriting Script Recognition.

DOI: 10.5220/0012385000003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 689-698

ISBN: 978-989-758-680-4; ISSN: 2184-433X

689

 Static mode (ElKessab et al., 2013): It mainly

uses scanners. The scanner scans the text line

by line and digitizes each line into a larger or

smaller series of dots. The resolution of a

scanner, expressed in number of dots per inch

(dpi), refers to its ability to digitize fine lines.

In OCR, the most common values range from

200 to 400 dpi (Abuhaiba, 2004); a higher

resolution does not increase the precision of

the digitization; on the contrary it increases

the number of points to be processed and

generates noise (grain of the paper). Several

types of scanners exist on the market offering

several input modes: flatbed, drum, or

handheld, providing a choice of raster images,

binary images, grayscale images and color

images.

 Continuous mode (Begum et al., 2021): It uses

a graphic tablet which sends coordinates of

points to the contact of the pen. Capturing fine

lines depends on the regular support of the pen

so as not to cause discontinuities in the lines.

This mode is quite valuable in OCR because it

gives very useful information for recognition

such as the number of pencil strokes and

therefore the reading direction, and the

number of points per pencil stroke significant

of the curvature of the line.

1.3 Categories of Handwritten

Characters Recognition

Several kinds of handwritten character recognition

exist such as:

 Online or offline (Kannan et al., 2008; Tappert

et al., 1990): Online recognition is a dynamic

recognition that takes place during writing. A

slight delay of a word or a character allows the

recognition not to encroach on the input. The

continuous response of the system allows the

user to correct and modify writing directly and

instantaneously. Offline or delayed

recognition starts after the acquisition of the

entire document. It is suitable for printed

documents and already written manuscripts.

This mode allows the instantaneous

acquisition of a large number of characters but

imposes costly pre-processing to find the

reading order.

 With or without learning (Stremler &

Karácsony, 2016): A system with learning

includes a module for introducing reference

character models. Significant samples of

writing (printed or handwritten) in sufficient

numbers (tens or even hundreds of samples

per character) are entered in manual or

automatic mode. In the first case, the user

indicates to the system the identity of each

sample allowing the system to organize its

classes according to the vocabulary which is

being studied. For the handwritten script, the

ideal would be to learn characters directly

from a text, but this presents obvious

segmentation problems. In the second case,

the samples are grouped automatically from

morphological analyses, without knowledge of

their name. In a non-learning system, a

knowledge base is built into the system and

particular analysis algorithms are used.

 Direct, scaled or with labelling (Zhang et al.,

2023): In systems with direct recognition, the

learning includes a reference model per

character. The recognition compares the

candidate character to each of these models

and retains the closest model. In scaled

recognition, learning is divided into

progressively selective classification levels.

Each level contains separation tests for the

different classes represented, allowing

recognition to refine its decision if necessary.

Labelling in the latter case consists of

identifying the different characters in a text,

thus distinguishing patterns, and then

recognizing them. This mode has the

advantage of limiting recognition only to the

patterns found and being able to quickly make

corrections, but it has the disadvantage of

perpetuating any recognition error through all

the characters of the same model.

1.4 Challenges

It is not obvious to be able to accurately recognize

human writing. Indeed, the form of a handwritten

character often varies, reflecting the style of writing,

the state of mind, and the personality of the writer,

which makes it difficult to characterize.

Among the variations that affect it, we can

mention the contour distortion that produces curls

and rounding, as well as the inclination that

produces a rotation of the base of the character or a

flattening of its shape, and the asymmetry generating

an occlusion of the character parts or a closure of its

cavities, and the poor connection of the lines leading

to their sudden extensions and interruptions.

The remnant of this paper is structured as

follows: In the following section we provide an

overview around the related works and the most

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

690

known datasets in terms of Arabic handwritten text

recognition. Then, we present an innovative

approach for features extraction step followed by

applying LSTM deep neural model for classification

step. Afterwards, we describe the experimentation

implemented. A summarized conclusion is provided

in the last section.

2 RELATED WORKS

2.1 Standardization and Famous

Datasets

Despite the artifacts of handwriting, attempts of

standardization have been elaborated. Indeed,

committees from different countries have proposed

standards of handwritten characters for recognition.

In 1974, the American National Standards Institute

(ANSI) (Alphabetic Handprint Reading, 1978)

developed a character set introducing additional

features to remove ambiguities. We can also

mention the “Japanese Industry Standard” proposed

by the “Japanese OCR Committee” and the

standards of the “OCR Committee of Canada”. In

order to be able to check the performance of

handwritten character recognition systems and to

compare them, test datasets were established,

containing several samples of different handwritten

characters, including uppercase and lowercase

letters, digits, and punctuation marks. These

databases were created to support research in the

field of pattern recognition and machine learning

and have been widely used in numerous research

papers and projects. The most frequently used

datasets are those of Munson (Munson, 1968) and

Highleyman (Highleyman, 1961).

The Munson database includes the 46 characters

of Fortran II, represented by matrices of size 24 X

24. It includes 12,760 samples of which 6,762 are

written by different writers on special sheets. The

others were taken either from particular authors or

from documents. Munson instructed scripters to

strike through the characters O and Z, to print 1

without a slash, and to place horizontal strokes over

the I. Highleymann's dataset consists of the 36

alphanumeric characters represented by 12 X 12 size

matrices. It includes 1,800 letters and 500 numbers

from 50 writers who were instructed to write on grid

paper in a writing frame.

We also mention many other recent datasets such

as Mori handwriting dataset (Mori et al., 1984)

which is a collection of handwritten Japanese

characters created by the Mori Laboratory at the

University of Tokyo. It contains over 3,000

handwritten characters written by 100 different

writers and is widely used in the development of

handwriting recognition systems. Moreover, the

Modified National Institute of Standards and

Technology (MNIST) dataset was introduced in

1998 (Lecun et al., 1998) and is a widely used as a

benchmark dataset for handwritten digit recognition.

It consists of 60,000 training images and 10,000

testing images of a 28 × 28 size of handwritten digits

from 0 to 9. Furthermore, EMNIST (Cohen et al.,

2017) dataset is an extension of the MNIST dataset

and includes handwritten letters in addition to digits.

It contains 240,000 training images and 40,000

testing images. Also, IAM Handwriting dataset

(Marti & Bunke, 2002) contains handwritten words

and sentences in English including over 5,000 pages

of handwritten text from 657 writers.

CEDAR dataset (Hull, 1994) is composed of

handwritten forms, including surveys,

questionnaires, and tax forms. It contains over 1,000

forms with over 50,000 fields. Besides the

mentioned, ICDAR is a collection of datasets (Lucas

et al., 2003; Lucas, 2005; Shahab et al., 2011) for

handwritten text recognition, including Chinese,

Japanese, and Arabic text. RIMES (Reconnaissance

et Indexation de données Manuscrites et de fac

similÉS / Recognition and Indexing of handwritten

documents and faxes) (Grosicki et al., 2009) is a

modern dataset of handwritten words in French with

over 1 million words written by over 1,300 writers.

Additionally, the Street View House Numbers

(SVHN) (Netzer et al., 2011) is a Google’s dataset

of street view house numbers, which includes

handwritten digits from 0 to 9. It contains over

600,000 images. One more dataset that we should

consider is the ADAB dataset (Arabic DAtaBase)

(Kherallah et al., 2011; Tagougui et al., 2012;

Boubaker et al., 2012) made up of 15,000 Arabic

names of Tunisian towns and villages, handwritten

by more than 166 different writers. Figure 1 shows

the appearance of some sample images from ADAB

database displaying handwritten Arabic names of

some Tunisian towns and villages.

Figure 1: Appearance of some sample images from ADAB

dataset.

New Approach Based on Substantial Derivative and LSTM for Online Arabic Handwriting Script Recognition

691

2.2 Works on Arabic Handwritten

Recognition

Various approaches have been used to automate the

process of identifying and converting handwritten

Arabic text into digital format, which is essential for

many applications related to document analysis, text

mining, and machine translation. This task mainly

relies on using advanced algorithms and machine

learning techniques to recognize the unique features

of Arabic handwriting, such as the shape and size of

letters, the direction of strokes, and the spacing

between words. It has become an important research

area in the field of computer vision and natural

language processing and has the potential to

revolutionize the way we interact with Arabic

language documents.

There have been numerous research studies and

developments in the field of Arabic handwritten

recognition. Among the notable works, the

comprehensive review of the state-of-the-art

techniques and methodologies used in Arabic

handwritten recognition (Cheriet, 2007) as well as

many studies presented throughout the last decade

such as a survey (Tagougui et al., 2012) where

different approaches and techniques used for online

Arabic handwriting recognition have been exposed.

Another system that combines different feature

extraction methods and classifiers for Arabic

handwritten text recognition was proposed (Al-

Maadeed, 2012).

Furthermore, the literature was examined on the

most significant work in Arabic optical character

recognition (AOCR) and a survey of databases on

Arabic offline handwritten character recognition

system was provided (Abdalkafor, 2018) as well as a

recognition system of Arabic cursive handwriting

using embedded training based on hidden Markov

models (Rabi et al., 2017).

Later, various papers (Noubigh et al., 2020;

Altwaijry & Al-Turaiki, 2020; AlJarrah et al., 2021;

Ali & Mallaiah, 2022) presented some deep learning

approaches such as convolutional neural networks

(CNNs) for Arabic handwritten text recognition.

These works demonstrate the ongoing research

and development in the field of Arabic handwritten

recognition, and the potential for further

advancements in this area.

3 PROPOSED METHOD

Enlightened by a main concept emerging from the

physics of fluid mechanics previously explained

(Lejmi et al., 2020), we suggest an innovative

framework to recognize handwritten text from

ADAB dataset. The overall framework phases of the

proposed model are depicted in figure 2.

Figure 2: Framework phases of online Arabic handwriting

recognition.

Specifically, we suggest performing the features

extraction step using the optical flow and the

substantial derivative (SD).

The latter describes the rate of change of a

particle while in motion with respect to time.

Analogically to the particle derivative stemming

from the physics of fluid mechanics, we estimate

local and convective accelerations from ADAB

dataset frames. In fact, the local or temporal

acceleration represents the increase rate of a pixel’s

speed over time at a specific point of the flow. The

convective acceleration describes the increase rate of

speed due to the change in pixel position.

To estimate local and convective accelerations,

we need first to calculate the optical flow (Lejmi et

al., 2017). It represents a set of vector fields that

relates a frame to an upcoming one (figure 3). Each

vector field describes the obvious movement of each

pixel from frame to frame.

Figure 3: Plot of optical flow vector for an image from

ADAB dataset.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

692

Considering the “Brightness Conservation

Theorem” which means that "The brightness of an

object is constant from one image to another.”:



x,y,t





x+ dx, y+dy, t+dt



(1)

‘I’ represents an image sequence, ‘dy’ and ‘dx’

represent displacement vectors for the pixel with

coordinates



x,y



and ‘t’ and ‘dt’ the frame and

temporal displacement of the image sequence. The

optical flow equation is derived from the above

equation as follows:



U+f



V+ f



=0 (2)

‘fx’ , ‘fy’ represent pixel intensity gradients and

‘f



’ represents the first temporal derivative.

If we solve (2) we obtain a couple of flow vector

maps U and V that dictate perceived motion in both

the x and y coordinate plane.

For each frame











, the optical flow











represents each pixel’s velocity in x and y

directions:



x,y



=v









 (3)

When we apply (3), the local acceleration gets

the value of the rate of change of velocity over time

at a fixed point in a flow field.

Generally, the rate of change of the quantity

undergone by an observer who moves with the flow

is described in (4) and (5).









∗v







∗v







=v

⃗

grad



⃗

f +





(4)





≡v.∇f+





(5)

Let a





be the local acceleration in an “x”

direction and a





the local acceleration at “y”

direction as detailed below:









−v





(6)









−v





(7)

The calculation of the local acceleration a



of a

couple of successive optical flows is obtained as

follows:

















(8)

To obtain the rate of change of velocity with

respect to position at a fixed time in a flow field, the

convective acceleration should be calculated. It is

combined with spatial velocity gradients in the flow

field. We consider a



as the convective acceleration

in an “x” direction and a



as the convective

acceleration in a “y” direction:



















∗v



(9)



















∗v



(10)

Let a



be the convective acceleration magnitude

defined as follows:













(11)

Overall, the physical interpretation of the

substantial derivative (SD) is highlighted in figure 4

indicating the total acceleration of the pixel moving

along its trajectory.

Figure 4: Physical interpretation of the SD concept.

Figure 5: Bidirectional-LSTM classification network (SD-

LSTM).

Afterwards, the classification algorithm will be

implemented using a recurrent neural network (Long

Short-Term Memory LSTM) (Lejmi et al., 2020),

which can process both isolated data as well as

sequences. This helps avoid long-term dependency

issues, by interacting through four layers of neural

network and gates indicating which data is useful to

keep and which is not. Thus, only relevant data

New Approach Based on Substantial Derivative and LSTM for Online Arabic Handwriting Script Recognition

693

passes through the sequence chain to facilitate

prediction. The LSTM deep learning classification

technique allows to classify the generated features

and to calculate a prediction value for each word.

Figure 5 illustrates how to create a Bidirectional-

LSTM (Bi-LSTM) classification network.

We concatenate the extracted features and feed

them to a Softmax classifier through a fully

connected operator. The classification ability of the

model will be later evaluated on confusion matrices

which will present the system predictions and their

actual labels.

4 EXPERIMENTS

We resorted to MATLAB R2022a software to carry

out experimental work on an Intel (R) Core (TM) i9-

11980HK, 2.6 GHz and 32 GB RAM under

Windows 11 operating system (64-bit).

4.1 ADAB Dataset and Preprocessing

In addition to input files in Tag(ged) Image File

Format (TIFF) of ADAB dataset, one more

alternative is considering the files in the Ink Markup

Language (InkML) format that was created by the

W3C (Watt & Underhill, 2011; Xavier et al., 2014)

as a standard for data storage ink. Indeed, the

features inside them are called “traces” where each

one represents a continuous writing curve, as shown

in figure 6. Traces consist of series of points. Each

Figure 6: Schematization of one handwritten script from

ADAB dataset generated by connecting the dots of 7

stored InkML traces. The displayed lines of the inkml file

below represent the 7 traces of this Arabic word.

represents a number of coordinate values whose

meanings are provided by a <traceFormat> element.

These coordinates are able to give us information

on values for quantities like pen position, angle, tip

force, button state, etc. More clearly, the data is

recorded in consecutive elements within <trace>

tags.

Each one includes a comma-separated

"coordinate" tuple representing points pairs (x , y).

For each point, the pair (x , y) describes the position

of the pen relative to the origin (0 , 0) placed in the

upper left corner of the screen. Thus, we read the

files into the ADAB dataset composed of three sets

(set1, set2 and set3) containing the names of 937

Tunisian towns and villages written in Arabic and in

more than one handwriting, and presented in Ink

Markup Language (InkML) files.

The hypothesis of our research is to read the

ADAB online Arabic handwriting dataset in order to

perform the main procedures below:

 Calculating substantial derivative equations

and then getting features using optical flow

and taking advantage of the results to compute

convective and local acceleration in order to

get local and convective features which will

be concatenated to obtain the total ones.

 Using total features with taxonomic

information in order to train and test the

obtained characteristics before utilizing them

for classification based on the LSTM recurrent

neural network.

A first difficulty in obtaining meaningful

features (Al-Helali & Mahmoud, 2016; Wilson-

Nunn et al., 2018) using optical flow, is that it

accepts input datasets only in video format. We,

therefore, need to find a technique to convert the

INKML files of our dataset into a suitable form that

can efficiently compute and calculate the optical

flow.

One more important issue is the need for an

approach to prepare files names, so that we can read

them and associate each INKML file with its

intended label. The latter can be found in the UPX

file having the same name, precisely in the

"alternative value tag", for example in the case of the

name "ﺔ

ﻄﺤﻤﻟﺍ ﻥﻭﺪﻌﺳ ﺏﺎﺑ", or as depicted in line 17 of

the example previously shown in figure 5. We also

need to take into consideration, when preparing the

file name, the association of the INKML file with all

the other ones having the same label, i.e. the same

"alternative value tag". In addition, we need to

prepare these names based on their occurrences in

the dataset so that they are easy to read and to use

during the training and classification phases.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

694

4.2 Features Extraction Using

Substantial Derivative and

Acceleration Features

The traces marks are selected in order to convert

data into video with an AVI (Audio Video

Interleave) format and then deduce the dimensions

of the x and y coordinates from the InkML files so

that the video starts drawing the word from the first

point of its intersection and continues until this word

is completed as a whole. First, we managed to use

the *.inkml files and extract x and y coordinates

from traces which represent the words, in order to

store the source files names and their coordinates in

the file 'inkmlFileName_X_Y.mat' and use them

later. This is performed for dataset set1 while

making sure, through two subsequent steps for set 2

and set 3 of this same dataset, to store the results

later in the matrices ‘inkmlFileName_X_Ys2.mat’

and ‘inkmlFileName_X_Ys3’. Then, the generated

inkmFileName_X_Y.mat will be used to read and

convert data from inkml files to AVI video file with

the same name, using the coordinates for each file to

extract each video in a new way: The graphics axis

limits will be set according to the actual x and y

point limits without affecting the size of the output

video. This starts by exploiting the current

coordinates and changing them according to the

following ones to adapt simultaneously to the

successive coordinates as depicted in figure 7.

Figure 7: An example of converting the data read from

inkml file in case of the word “

ﻱﺩﺍﻭ ﻞﻣ

ﺮﻟﺍ

” which

represents one file.

Through the first part of the following figure, we

highlight a new way of reading Arabic writing in

order to benefit from the largest possible number of

video image coordinates, unlike the usual way where

most of the video image remains empty or

stationary, particularly for short words, as shown for

example in the case of the word “ﺓﺩﺎﻣﺭ” (figure 8).

Figure 8: Example of empty part of video image in case of

ADAB short words.

This helps us to collect more features in the next

step. This new method is beneficial in terms of

homogeneity of the data used for analysis, as all the

video files and their frames have the same size and

line thickness. This will help us to achieve more

accurate results in the classification process.

4.3 Classification Results

This section presents our experimentation in the

field of multilingual online handwriting script

recognition. First, we introduce the datasets used in

our study. Then, we provide an in-depth discussion

of the results. Both kinds of acceleration features are

mixed and then divided into a Training set and a test

set. In this context, it is noteworthy to mention the

number of frames that have been processed.

To deal with challenging aspects of online

handwriting recognition (Mahmoud et al., 2018), it

is recommended to use a standardized database that

accommodates different writing styles and

encompasses various classes in the target language.

To evaluate the performance and effectiveness of

our proposed system, we used ADAB dataset

(Boubaker et al., 2012), tailored for Arabic words

which collectively serve as pivotal components of

our experimentation.

Figure 9: LSTM minimizes loss function for total

acceleration features.

This ADAB dataset, consisting of over 33,000

handwritten Arabic words by 170 different writers,

has been widely utilized in the literature. It contains

937 names of Tunisian towns and villages. This

database is segmented into six distinct sets,

originally collected for the ICDAR 2011 online

Arabic handwriting recognition competition

(Elleuch et al., 2016).

To improve the training accuracy, we applied

some data augmentation techniques (Hamdi et al.,

2021) and we focused on sets 1, 2, and 3 which

altogether contain more than 45,158 words for the

training process. When testing the network, the

recognition rate was quite high for the Bi-LSTM

New Approach Based on Substantial Derivative and LSTM for Online Arabic Handwriting Script Recognition

695

neural network trained with an SGDM optimizer and

a learning rate of 0.01 during 100 epochs (figure 9).

Indeed, the accuracy was around 90% and 95%.

Specifically, for the Training set, it reached 99%.

Based on features extraction results, further

implementation is on the way to enhance the

classification performance. Indeed, we are planning

to increase the size of the ADAB dataset as it

currently has a considerable imbalance in terms of

content occurrences. We believe that this will

significantly improve the results. Besides, we

performed a comparative analysis of the suggested

system against other ones that have been

experimented in the field of online Arabic character

recognition. The results are summarized in Table 1.

Regarding the ADAB dataset, our results are like the

state of the art.

Table 1: Comparison of recognition rate between the

suggested approach and some existing models.

Models

Recognition

rate

(

)

Elleuch et al.

(

2016

)

97,5

Abdelaziz an

Abdou

(

2014

)

97,1

Tagougui and Kherallah (2017) 96,2

SD-LSTM 95,3

5 CONCLUSIONS

In this paper, we presented various well-known

datasets and crucial works underlying the

approaches of handwritten recognition and we

specially outlined those used to process identifying

Arabic handwriting as well. Then, enlightened by a

main concept of fluid mechanics, we presented a

novel model based on an initial phase of spatio-

temporal features extraction using the optical flow

and the substantial derivative to calculate local,

convective and total accelerations. Afterwards, we

suggested a classification model relying on the deep

learning LSTM neural network. The last part was

mainly devoted to the experiments we performed on

ADAB dataset as well as the implementation of the

descriptor and the technical tricks that we proposed

in order to effectively classify the Arabic characters.

We believe that further research on this can lead to

fruitful results.

REFERENCES

Mori, S., Suen, C. Y., & Yamamoto, K. (1992). Historical

review of OCR research and development.

Proceedings of the IEEE. Institute of Electrical and

Electronics Engineers, 80(7), 1029–1058.

doi:10.1109/5.156468

Wu, Z., Yu, C., Xu, X., Wei, T., Zou, T., Wang, R., & Shi,

Y. (2021). LightWrite: Teach Handwriting to The

Visually Impaired with A Smartphone. In Proceedings

of the 2021 CHI Conference on Human Factors in

Computing Systems. CHI ’21: CHI Conference on

Human Factors in Computing Systems. ACM.

https://doi.org/10.1145/3411764.3445322

Singh, S., Kariveda, T., Gupta, J. D., & Bhattacharya, K.

(2015). Handwritten words recognition for legal

amounts of bank cheques in English script. In 2015

Eighth International Conference on Advances in

Pattern Recognition (ICAPR). 2015 Eighth

International Conference on Advances in Pattern

Recognition (ICAPR). IEEE. https://doi.org/10.

1109/icapr.2015.7050716

Charfi, M., Kherallah, M., El, A., & M., A. (2012). A New

Approach for Arabic Handwritten Postal Addresses

Recognition. In International Journal of Advanced

Computer Science and Applications (Vol. 3, Issue 3).

The Science and Information Organization.

https://doi.org/10.14569/ijacsa.2012.030301

Nagy, G. (2016). Disruptive developments in document

recognition. In Pattern Recognition Letters (Vol. 79,

pp. 106–112). Elsevier BV. https://doi.org/10.

1016/j.patrec.2015.11.024

Chherawala, Y., & Cheriet, M. (2014). Arabic word

descriptor for handwritten word indexing and lexicon

reduction. In Pattern Recognition (Vol. 47, Issue 10,

pp. 3477–3486). Elsevier BV. https://doi.org/10.

1016/j.patcog.2014.04.025

Velázquez, A., & Levachkine, S. (2004). Text/Graphics

Separation and Recognition in Raster-Scanned Color

Cartographic Maps. In Graphics Recognition. Recent

Advances and Perspectives (pp. 63–74). Springer

Berlin Heidelberg. https://doi.org/10.1007/978-3-540-

25977-0_6

ElKessab, B., Daoui, C., & Bouikhalene, B. (2013).

Handwritten Tifinagh Text Recognition using Neural

Networks and Hidden Markov Models. In

International Journal of Computer Applications (Vol.

75, Issue 18, pp. 54–60). Foundation of Computer

Science. https://doi.org/10.5120/13354-0127

Abuhaiba, I. S. I. (2004). Discrete Script or Cursive

Language Identification from Document Images. In

Journal of King Saud University - Engineering

Sciences (Vol. 16, Issue 2, pp. 253–268). Elsevier BV.

https://doi.org/10.1016/s1018-3639(18)30790-6

Begum, N., Akash, M. A. H., Rahman, S., Shin, J., Islam,

M. R., & Islam, M. E. (2021). User Authentication

Based on Handwriting Analysis of Pen-Tablet Sensor

Data Using Optimal Feature Selection Model. In

Future Internet (Vol. 13, Issue 9, p. 231). MDPI AG.

https://doi.org/10.3390/fi13090231

Kannan, R. J., Prabhakar, R., & Suresh, R. M. (2008). Off-

line Cursive Handwritten Tamil Character

Recognition. In 2008 International Conference on

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

696

Security Technology. (SECTECH). IEEE. https://doi.

org/10.1109/sectech.2008.33

Tappert, C. C., Suen, C. Y., & Wakahara, T. (1990). The

state of the art in online handwriting recognition. In

IEEE Transactions on Pattern Analysis and Machine

Intelligence (Vol. 12, Issue 8, pp. 787–808). Institute

of Electrical and Electronics Engineers (IEEE).

https://doi.org/10.1109/34.57669

Stremler, S., & Karácsony, Z. (2016). Efficient

Handwritten Digit Recognition Using Normalized

Cross-Correlation. In The publications of the

MultiScience - XXX. MicroCAD International

Scientific Conference. MultiScience. University of

Miskolc. https://doi.org/10.26649/musci.2016.058

Zhang, Y., Li, Z., Yang, Z., Yuan, B., & Liu, X. (2023).

Air-GR: An Over-the-Air Handwritten Character

Recognition System Based on Coordinate Correction

YOLOv5 Algorithm and LGR-CNN. In Sensors (Vol.

23, Issue 3, p. 1464). MDPI AG. https://doi.

org/10.3390/s23031464

Alphabetic Handprint Reading. (1978). In IEEE

Transactions on Systems, Man, and Cybernetics (Vol.

8, Issue 4, pp. 279–282). Institute of Electrical and

Electronics Engineers (IEEE). https://doi.org/10.

1109/tsmc.1978.4309949

Munson, J. H. (1968). Experiments in the recognition of

hand-printed text, part I. In Proceedings of the

December 9-11, 1968, fall joint computer conference,

part II on - AFIPS ’68 (Fall, part II). ACM Press.

https://doi.org/10.1145/1476706.1476735

Highleyman, W. H. (1961). An Analog Method for

Character Recognition. In IEEE Transactions on

Electronic Computers: Vol. EC-10 (Issue 3, pp. 502–

512). Institute of Electrical and Electronics Engineers

(IEEE). https://doi.org/10.1109/tec.1961.5219239

Mori, S., Yamamoto, K., & Yasuda, M. (1984). Research

on Machine Recognition of Handprinted Characters.

In IEEE Transactions on Pattern Analysis and

Machine Intelligence: Vol. PAMI-6 (Issue 4, pp. 386–

405). Institute of Electrical and Electronics Engineers

(IEEE). https://doi.org/10.1109/tpami.1984.4767545

Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998).

Gradient-based learning applied to document

recognition. In Proceedings of the IEEE (Vol. 86,

Issue 11, pp. 2278–2324). Institute of Electrical and

Electronics Engineers (IEEE). https://doi.org/10.

1109/5.726791

Cohen, G., Afshar, S., Tapson, J., & van Schaik, A.

(2017). EMNIST: Extending MNIST to handwritten

letters. In 2017 International Joint Conference on

Neural Networks (IJCNN). IEEE. https://doi.org/10.

1109/ijcnn.2017.7966217

Marti, U.-V., & Bunke, H. (2002). The IAM-database: an

English sentence database for offline handwriting

recognition. In International Journal on Document

Analysis and Recognition (Vol. 5, Issue 1, pp. 39–46).

Springer Science and Business Media LLC.

https://doi.org/10.1007/s100320200071

Hull, J. J. (1994). A database for handwritten text

recognition research. In IEEE Transactions on Pattern

Analysis and Machine Intelligence (Vol. 16, Issue 5,

pp. 550–554). Institute of Electrical and Electronics

Engineers (IEEE). https://doi.org/10.1109/34.291440

Lucas, S. M., Panaretos, A., Sosa, L., Tang, A., Wong, S.,

& Young, R. (n.d.). ICDAR 2003 robust reading

competitions. In Seventh International Conference on

Document Analysis and Recognition, 2003.

Proceedings. Seventh International Conference on

Document Analysis and Recognition. IEEE Comput.

Soc. https://doi.org/10.1109/icdar.2003.1227749

Lucas, S. M. (2005). ICDAR 2005 text locating

competition results. In Eighth International

Conference on Document Analysis and Recognition

(ICDAR’05). IEEE. https://doi.org/10.1109/icdar.

2005.231

Shahab, A., Shafait, F., & Dengel, A. (2011). ICDAR

2011 Robust Reading Competition Challenge 2:

Reading Text in Scene Images. In 2011 International

Conference on Document Analysis and Recognition.

IEEE. https://doi.org/10.1109/icdar.2011.296

Grosicki, E., Carré, M., Brodin, J.-M., & Geoffrois, E.

(2009). Results of the RIMES Evaluation Campaign

for Handwritten Mail Processing. In 2009 10th

International Conference on Document Analysis and

Recognition. IEEE. https://doi.org/10.1109/icdar.

2009.224

Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B. &

Ng, A. Y. (2011). Reading Digits in Natural Images

with Unsupervised Feature Learning

Kherallah, M., Tagougui, N., Alimi, A. M., Abed, H. E., &

Margner, V. (2011). Online Arabic Handwriting

Recognition Competition. In 2011 International

Conference on Document Analysis and Recognition.

(ICDAR). IEEE. https://doi.org/10.1109/icdar.

2011.289

Tagougui, N., Kherallah, M., & Alimi, A. M. (2012).

Online Arabic handwriting recognition: a survey. In

International Journal on Document Analysis and

Recognition (IJDAR) (Vol. 16, Issue 3, pp. 209–226).

Springer Science and Business Media LLC.

https://doi.org/10.1007/s10032-012-0186-8

Boubaker, H., Elbaati, A., Tagougui, N., El Abed, H.,

Kherallah, M., & Alimi, A. M. (2012). Online Arabic

Databases and Applications. In Guide to OCR for

Arabic Scripts (pp. 541–557). Springer London.

https://doi.org/10.1007/978-1-4471-4072-6_22

Cheriet, M. (2007). Strategies for visual arabic

handwriting recognition: Issues and case study. In

2007 9th International Symposium on Signal

Processing and Its Applications. (ISSPA). IEEE.

https://doi.org/10.1109/isspa.2007.4555620

Al-Maadeed, S. (2012). Text-Dependent Writer

Identification for Arabic Handwriting. In Journal of

Electrical and Computer Engineering (Vol. 2012, pp.

1–8). Hindawi Limited. https://doi.org/10.1155/

2012/794106

Abdalkafor, A. S. (2018). Survey for Databases On Arabic

Off-line Handwritten Characters Recognition System.

In 2018 1st International Conference on Computer

New Approach Based on Substantial Derivative and LSTM for Online Arabic Handwriting Script Recognition

697

Applications & Information Security (ICCAIS). IEEE.

https://doi.org/10.1109/cais.2018.8442001

Rabi, M., Amrouch, M., & Mahani, Z. (2017).

Recognition of Cursive Arabic Handwritten Text

Using Embedded Training Based on Hidden Markov

Models. In International Journal of Pattern

Recognition and Artificial Intelligence (Vol. 32, Issue

01, p. 1860007). World Scientific Pub Co Pte Lt.

https://doi.org/10.1142/s0218001418600078

Noubigh, Z., Mezghani, A., & Kherallah, M. (2020).

Contribution on Arabic Handwriting Recognition

Using Deep Neural Network. In Hybrid Intelligent

Systems (pp. 123–133). Springer International

Publishing. https://doi.org/10.1007/978-3-030-49336-

3_13

Altwaijry, N., & Al-Turaiki, I. (2020). Arabic handwriting

recognition system using convolutional neural

network. In Neural Computing and Applications (Vol.

33, Issue 7, pp. 2249–2261). Springer Science and

Business Media LLC. https://doi.org/10.1007/s00521-

020-05070-8

AlJarrah, M. N., Zyout, M. M., & Duwairi, R. (2021).

Arabic Handwritten Characters Recognition Using

Convolutional Neural Network. In 2021 12th

International Conference on Information and

Communication Systems (ICICS). IEEE. https://doi.

org/10.1109/icics52457.2021.9464596

Ali, A. A. A., & Mallaiah, S. (2022). Intelligent

handwritten recognition using hybrid CNN

architectures based-SVM classifier with dropout. In

Journal of King Saud University - Computer and

Information Sciences (Vol. 34, Issue 6, pp. 3294–

3300). Elsevier BV. https://doi.org/10.1016/j.

jksuci.2021.01.012

Lejmi, W., Khalifa, A. B., & Mahjoub, M. A. (2020). A

Novel Spatio-Temporal Violence Classification

Framework Based on Material Derivative and LSTM

Neural Network. In Traitement du Signal (Vol. 37,

Issue 5, pp. 687–701). International Information and

Engineering Technology Association. https://doi.org/

10.18280/ts.370501

Lejmi, W., Mahjoub, M. A., & Ben Khalifa, A. (2017).

Event detection in video sequences: Challenges and

perspectives. In 2017 13th International Conference

on Natural Computation, Fuzzy Systems and

Knowledge Discovery (ICNC-FSKD). IEEE.

https://doi.org/10.1109/fskd.2017.8393354

Watt, S. M., & Underhill, T. (Eds.). (2011). Ink Markup

Language (InkML).

Retrieved from http://www.w3.

org/TR/InkML/

Xavier, C., Ambrosio, A. P., & Georges, F. (2014).

Written Assessments with Digital Ink. In 2014 14th

International Conference on Computational Science

and Its Applications (ICCSA). IEEE. https://doi.

org/10.1109/iccsa.2014.35

Al-Helali, B. M., & Mahmoud, S. A. (2016). A Statistical

Framework for Online Arabic Character Recognition.

In Cybernetics and Systems (Vol. 47, Issue 6, pp. 478–

498). Informa UK Limited. https://doi.org/

10.1080/01969722.2016.1206768

Wilson-Nunn, D., Lyons, T., Papavasiliou, A., & Ni, H.

(2018). A Path Signature Approach to Online Arabic

Handwriting Recognition. In 2018 IEEE 2nd

International Workshop on Arabic and Derived Script

Analysis and Recognition (ASAR). IEEE.

https://doi.org/10.1109/asar.2018.8480300

Hamdi, Y., Boubaker, H., & Alimi, A. M. (2021). Data

Augmentation using Geometric, Frequency, and Beta

Modeling approaches for Improving Multi-lingual

Online Handwriting Recognition. In International

Journal on Document Analysis and Recognition

(IJDAR) (Vol. 24, Issue 3, pp. 283–298). Springer

Science and Business Media LLC. https://doi.

org/10.1007/s10032-021-00376-2

Elleuch, M., Zouari, R., & Kherallah, M. (2016). Feature

Extractor Based Deep Method to Enhance Online

Arabic Handwritten Recognition System. In Artificial

Neural Networks and Machine Learning – ICANN

2016 (pp. 136–144). Springer International Publishing.

https://doi.org/10.1007/978-3-319-44781-0_17

Abdelaziz, I., & Abdou, S.M. (2014). AltecOnDB: A

Large-Vocabulary Arabic Online Handwriting

Recognition Database. ArXiv, abs/1412.7626.

Tagougui, N., & Kherallah, M. (2017). Recognizing

online Arabic handwritten characters using a deep

architecture. In A. Verikas, P. Radeva, D. P. Nikolaev,

W. Zhang, & J. Zhou (Eds.), SPIE Proceedings. SPIE.

https://doi.org/10.1117/12.2268419

Mahmoud, S. A., Luqman, H., Al-Helali, B. M.,

BinMakhashen, G., & Parvez, M. T. (2018). Online-

KHATT: An Open-Vocabulary Database for Arabic

Online-Text Processing. In The Open Cybernetics &

Systemics Journal (Vol. 12, Issue 1, pp. 42–59).

Bentham Science Publishers Ltd. https://doi.org/

10.2174/1874110x01812010042.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

698