A Deep Convolutional and Recurrent Approach for Large

Vocabulary Arabic Word Recognition

Faten Ziadi

1,2 a

, Imen Ben Cheikh

1 b

and Mohamed Jemni

1 c

University of Sousse, ISITCom, 4011, Sousse, Tunisia

Latice Laboratory, ENSIT, University of Tunis, Tunis, Tunisia

Keywords: CNN, LSTM, Arabic Writing, Large Vocabulary, Linguistic Knowledge, APTI Dataset.

Abstract: In this paper, we propose a convolutional recurrent approach for Arabic word recognition. We handle a large

vocabulary of Arabic decomposable words, which are factored according to their roots and schemes.

Exploiting derivational morphology, we have conceived as the first step a convolutional neural network,

which classifies Arabic roots extracted from a set of word samples int the APTI database. In order to further

exploit linguistic knowledge, we have accomplished the word recognition process through a

recurrent network, especially LSTM. Thanks to its recurrence and memory cabability, the LSTM model

focuses not only prefixes, infixes and suffixes listed in chronological order, but also on the relation between

them in order to recognize word patterns and some flexional details such as, gender, number, tense, etc.

1 INTRODUCTION

Recognition is an area that covers various fields such

as facial recognition, fingerprint recognition, image

and character recognition, number recognition, etc.

For decades, considerable progress has been made in

handwriting recognition.

Our work focuses primarily on this field of

research, specifically Arabic writing recognition.

Arabic is a mother Semitic tongue (Ibrahim,

Bilmas and Babiker, 2013) spoken by hundreds of

millions of people in Middle Eastern countries.

The following figure (Figure 1) illustrates the 28

characters of the Arabic alphabet.

The morphological complexity of written Arabic

and its cursivity remains a very broad subject area

which has known in recent years a great progress in

diverse fields.

Indeed, the complexity of the recognition process

depends on the type of script (printed or handwritten),

the approach (holistic, pseudo-global or analytic)

(Touj,Ben Amara and Amiri, 2007) (Avila, 1996) and

the vocabulary size (reduced, large).

https://orcid.org/0000-0002-1966-5630

https://orcid.org/0000-0003-2119-1312

https://orcid.org/0000-0001-8841-5224

Figure 1: The 28 character of the Arabic alphabet.

Numerous approaches have been proposed,

dealing with letters and / or pseudo-words, and

several works have experienced statistical, neural,

stochastic methods on different types of Arabic

documents.

Among these approaches, we will focus on neural

networks by presenting two variants: convolutional

networks and recurrent networks with LSTM

memory.

In this context, we have chosen to combine these

two architectures in order to recognize a large

vocabulary of Arabic words.

Ziadi, F., Ben Cheikh, I. and Jemni, M.

A Deep Convolutional and Recurrent Approach for Large Vocabulary Arabic Word Recognition.

DOI: 10.5220/0010814800003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 213-220

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

213

The remainder of this paper is structured as

follows: Section 2 displays the main characteristics of

the Arabic script. Section 3 describes not only the

performance of convolutional network and LSTM,

but also various related works using this two models

in the recognition field.

Section 4 reviews the related works suggested in

the literature using respectively CNN model, LSTM

model, and the hybridation of these two models for

writing recognition. Section 5 introduces our

suggested approach. Then, section 6 reports our

preliminary results. Finally, section 7 summarizes the

paper and highlights major directions for future

research.

2 ARABIC LANGUAGE

CHARACTERISTICS

Thanks to its stability, Arabic is a rich language in

terms of its morphology, phonology and vocabulary

(Cheriet and Beldjehem, 2006) (Ben Hamadou, 1993)

(Kanoun,Alimi and Lecourtier, 2005) (Kammoun and

Ennaji, 2004). An Arabic word can be decomposed

into smaller units. If it derives from a root, it is said

to be decomposable into morphemes (root, prefix,

infix and suffix). Figure 2 illustrates our vision of

Arabic words derived from several roots and

conjugated according to different schemes.

Figure 2: Our vision of an Arabic word.

Based on this principle, we were able to enlarge our

vocabulary not only by combining words derived

from the same root with several schemes, but also by

applying conjugation rules (tense, number, gender,

etc.).

Figure 3 presents several samples of words that

are derived from the root  and conjugated

according to different schemes. For instance, the root

 derived according to the scheme  gives the

verb word. The latter conjugated in the past

tense with the masculine plural pronoun “they” gives

the verb اوفرصنا.

Figure 3: Various words derived from the root “”.

In this context, we aim to classify Arabic word

roots using a convolutional network. The output of

this model will be used by a recurrent network

(LSTM) to recognize whole word.

3 CNN AND LSTM OVERVIEW

3.1 Convolutional Neural Network

Convolutional Neural Network (CNN), proposed by

LeCun (LeCun and Bengio, 1995), is a neural

network model that combines three key architectural

ideas, namely local receptive fields, shared weight

and spacial subsampling. This network receives 2-D

input (e.g. an image, a voice signal) and extracts high-

level characteristics through a series of hidden layers.

These layers consist of:

▪ Convolution layer includes a set of filters whose

parameters need to be learned. These filters have

the same shape as the input but with smaller

dimensions. In the learning process, the filter of

each convolutional layer is tiled across the input

space (in the case of an image, for example, the

filter is slid across the width and height of the

input) and the input product is calculated. This

calculation of the entire input leads to a filtering

feature map.

▪ Pooling layer operates upon each feature map. It

reduces the spatial size of the representation in

order to reduce the number of parameters and

computational time, and hence to also control

over-fitting. Max-Pooling is a common

approach that partitions the input space into

non-overlapping regions and chooses the

maximum value of each region.

▪ Fully connected layer takes the output of

convolution/pooling and predicts the best label

to describe the image.

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

214

3.2 LSTM Network

Recurrent neural networks (RNN) have recurrent

connections. At time t, the RNN takes into account a

number of past states, following the short term

memory principle. We can deduce that RNNs are

adapted for analyzing contextual applications, and

more particularly for processing sequences.

However, this “short-term memory” is not anymore

sufficient in most applications involving long time

differences (typically the classification of video

sequences). In fact, “classic” RNNs are only able to

memorize the so-called near past, and start to forget

after multiple iterations.

These networks take as input: 1) the current input

and 2) what they have observed previously over time.

The decision reached by a recurrent neuron at time t-

1 affects the learning process at instant t. However,

RNNs suffer from the vanishing gradient problem.

As more layers using activation functions are

added, the gradients of the loss function begin to

approach to zero. This problem hampers learning of

long data sequences and therefore makes the network

hard to train. LSTM networks provide a solution to

this problem.

In addition to its external connections, A LSTM

neuron has a recurrent-self connection with a constant

coefficient equal to 1. This allows us to save the

successive states of the neuron, varying, from one

iteration to another. Multiplication gates (input-gate,

output-gate and forget-gate) protect the current state

of memory (the cell state) and the memory of the

whole network. The structure of the LSTM neuron is

illustrated in Figure 4.

Figure 4: LSTM node architecture.

Equations (1) and (2) are used to calculate the

current memory state (Ct) and current output (ht),

respectively.

)

**(

1 ttttt

CiCfC +=

−



(1)

ttt

oCh *)tanh(=

(2)

Where C

t-1

is the previous memory cell, C

is the

past value of the current memory. i

, o

and f

stand for

input-gate, output-gate and forget-gate, respectively.

These outputs are estimated by the current neuron

input and the previous neuron output.

4 RELATED WORKS

4.1 CNN for Writing Recognition

A special focus is put on Arabic word recognition.

Using CNN, various methods have been proposed

and high recognition rates have been reported for

handwriting recognition in English (Bai,Chen,Feng

and Xu, 2014) (Yuan, Bai, Jiao and Liu, 2012),

Chinese (Wu, Fan, He, Sun and Naoi, 2014) (Zhong,

Jin and Feng, 2015) (Yang, Jin, Xie and Feng, 2015)

(He, Zhang, Mao and Jin, 2015) (Zhong, Jin and Xie,

2015), Hangul (Kim and Xie, 2015), Malayalam

(Anil, Manjusha, Kumar and Soman, 2015),

Devanagari (Acharya, Pant, and Gyawali, 2015)

(Singh,Verma and Chaudhari, 2016), Telugu (Soman,

Nandigam and Chakravarthv, 2013).

The majority of these methods are based on

Convolutional Neural Networks. The first CNNs date

back to the 1980s with the work of K. Fukushima

(Fukushima, 1980), but it was in the 1990s that these

networks were popularized with the work of Y. Le

Cun et al. about character recognition (LeCun, Boser,

Denker, Henderson, Howard, Hubbardet and Jackel,

1990). In a study conducted by (Poisson and Gaudin,

2001), a particular neuron network TDNN (Time

Delay Neural Network) was implemented for

handwriting recognition. Their recognition results

show the input of a hidden Markov model.

4.2 LSTM for Writing Recognition

Unidirectional and bidirectional LSTM networks

have been tested in many handwriting applications.

In (Graves, Fernandez, Liwicki, Bunke and

Schmidhuber, 2008), a combined BLSTM-CTC

network was put forward in an online handwriting

dataset for unconstrained online handwriting

recognition. In a study carried out by (Graves and

Schmidhuber, 2009), a multidimensional LSTM was

combined with CTC for offline handwriting

recognition. (Su and Lu, 2015) set forth a

segmentation-free approach in order to recognize text

A Deep Convolutional and Recurrent Approach for Large Vocabulary Arabic Word Recognition

215

in scene images. A RNN was adapted to classify the

word sequential characteristics obtained using

Histograms of Oriented Gradients (HOG). (Graves,

Fernandez, Liwicki, Bunke and Schmidhuber, 2007)

offered a system that is able to manage online

handwritten data. It is based on a RNN with an output

layer designed for sequence labeling, associated with

a probabilistic language model. (Mioulet, 2015)

combined a bidirectional LSTM with a connectionist

temporal classification (BLSTM-CTC) for

handwriting recognition and keyword detection.

4.3 A Hybrid CNN-LSTM Network for

Arabic Recognition

Several works have combined CNN and LSTM

networks for Arabic language treatment.

(Hamdi et al., 2019) suggested an online Arabic

character recognition system based on hybrid Beta-

Elliptic model (BEM), CNN, bidirectional LSTM and

SVM. Online handwriting trajectory features are

extracted using BEM, while generic features are

extracted using CNN. The classification is achieved

by combining DBLSTM and SVM models.

To bridge the communication gap between deaf

and hearing communities, (Agrawal and Urolagin,

2020) introduced a 2-way Arabic sign language

translator using CNN-LSTM and NLP models for

translating texts into signs and vice versa. CNN layers

perform feature extraction and LSTM layers provide

temporal sign prediction.

(Al Omari, Al-Hajj, Sabra and Hammami, 2019)

recommended a combined CNN and LSTM model for

Arabic sentiment classification. This model was

tested on the Arabic Health Services (AHS) dataset.

It achieved more interesting results on the Main-

AHS and the Sub-AHS datasets compared with those

of the previous methods (Alayba, Palade, England

and Iqbal, 2018).

Since the combination of CNN and LSTM has

proven its effectiveness for improving the quality of

Arabic writing analysis, we use this hybridization of

different linguistic knowledge in order to recognize

Arabic decomposable words.

5 APPROACH PROPOSAL

With the recent emergence of deep learning

technology, the aforementioned research studies

described in previous sections inspired our work.

Thus, we have tried to combine two neural

network variants, namely CNN and LSTM models.

The first step of our work consists in

implementing a convolutional neural network that

learns and recognizes roots from word images. In the

second step, we feed a LSTM with word samples.

This LSTM is specialized in recognizing words

derived from the predicted root. It is conceived to

recognize the schemes and the flexional details of the

candidate word.

The CNN model comprises these different layers:

▪ A first convolutional layer with 64 filters,

each with the size of (3*3), followed by a

“relu” activation function.

▪ A max-pooling layer with a filter size (2*2).

▪ A second convolutional layer with 64 filters

each of size (3*3) and a “relu” activation

function.

▪ A max-pooling layer with a filter size (2*2).

▪ A flattening operation that flatten all data into

a single vector.

▪ A Dense layers where each neuron is

connected to all neurons in the previous layer.

“relu” is the activation function of this layer.

▪ A Dense layer with 10 neurons and a

“softmax” activation function to classify the

10 roots.

Note that the “Dropout” technique is used after

each layer to prevent overfitting.

Figure 5 shows the implemented CNN

architecture and decisions made at each layer.

Figure 5: CNN architecture.

As mentioned before, after the classification

phase, the recognized root and the whole word image

will be passed to the LSTM model. We have started

with a static segmentation where each word is split

into vertical bands of a constant number of columns

(25) as shown in Figure 6. Each band is converted into

a 150 sized vector (image width), where each value

corresponds to the sum of band row . At each time

step, LSTM will be fed by six vectors.

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

216

Figure 6: Splitting the word “ ”.

Figure 7: An example of the word “  ” recognition by

the LSTM network.

Figure 7 displays multiple information

propagation that allow the LSTM network to learn

word derivations and flexions with their

corresponding roots (CNN output).

In this example, at time t, the cell accepts two

information: 1) the root  (It/he becomes

abundant) and 2) the first letter “ ” of the verb word

to be recognized “ ” (i.e. She has been

reproduced).

To recall these two relevant data, the forget gate

is activated (=1). Now, this neuron is going the try to

predict the conjugation time, the scheme, etc. and its

memory will save these information which will be

exploited by the next cell.

At t+1, the second letter (“ ”) is the input of the

LSTM cell, the last memory will be used to go on,

while the current memory will ignore the current

input (it is actually the first root consonant) and

decide to save conjugation information to propagate

to the following cell at t+2.

At t+2, the following cell will determine the time

(. i.e. past tense) and the scheme () thanks to

the stored information and relevant input “ ”. Details

concerning the second and third root letters at t+3 and

t+4 are not memorized.

Finally, at time t+5, the last letter of the word

(suffix) will determine the personal pronoun (. i.e.

She). Once the root, time, scheme and different

conjugation elements are known, the word is

reconstituted and well-recognized.

6 EXPERIMENTAL RESULTS

6.1 Dataset

Our approach is tested using a database that contains

Arabic printed word images, of various sizes and

fonts, derived from three consonant roots (see Figure

8), displays samples extracted from the APTI

database.

Figure 8: Samples of words extracted from the APTI

database.

The corpus includes a set of Arabic words images.

We manipulated 17600 samples of 1100 words

derived from tri-consonant roots following different

schemes. For data collection, we have gathered words

from the APTI database and organized them into 10

sets. Each set corresponds to one root and contains

many words that follow several schemes conjugation

A Deep Convolutional and Recurrent Approach for Large Vocabulary Arabic Word Recognition

217

forms (e.g. different tenses and pronouns). Each

word is given in 16 samples. The distribution of the

words samples is as follows : 60% for the training,

25% for the validation and 15% for the test.

6.2 CNN Results

Both CNN and LSTM architectures have been created

based on the Keras Framework. Before fitting the

CNN model, we must compile it. Several parameters

are necessary to carry out this phase like the

optimizer, the loss function and optionally the

evaluation metrics.

To compile the model,we choose the Root Mean

Square Propagation (RMSPROP) optimizer,

Categorical Cross-Entropy loss (also called Softmax

Loss) and accuracy metric to evaluate our model.

In order to adjust the parameters of a model, a set

of hyperparameters must be chosen appropriately. In

our work we have used hyperparameters that we have

adjusted manully. In fact, we choose them according

to our experience. We then train the model, evaluate

its accuracy and repeat the process. This loop is

repeated until a satisfactory accuracy is noted.

After the training phase, estimated at about ten

hours, the CNN model achieved a 97% classification

rate after 30 epochs.

Some false predictions were detected during the

classification phase, such as the images presented in

the following figure (Figure 9) where the predicted

label is the root “”, while the true root is “”

which is too similar to the predicted label.

Figure 9: Examples of word predictions derived from the

root.

Figure 10 and Figure 11 represent curves

illustrating the two metrics: accuracy and loss,

respectively.

Based on the two figures, we notice that:

▪ The classification rate is 97%.

▪ The plot of training loss decreases to a point

of stability.

▪ The plot of validation loss decreases to a point

of stability and has a small gap with the

training loss.

We can deduce that this is the case of a good fit.

Figure 10: Performance curves on training and validation

data.

Figure 11: Representative curves of loss function on

training and validation data.

6.3 LSTM Results

Before fitting it in the LSTM, each word is split into

vertical bands with a constant number of columns

(25). In the short term, we intend to operate the

vertical projection of a word and send bands

delimited by local minima.

As first, the LSTM model was evaluated using a

corpus composed of 5000 samples. The first results

are promising. In fact, the recognition rate of the

LSTM algorithm is equal to 87.2% .

In fact, false results are detected mainly because

of the segmentation phase. As mentioned, in this

work we started with a static segmentation where

each word sample is divided into 6 bands to detect

each letter of the word. Since we have words

composed of more than 6 letters as well as less than 6

letters, letters are not properly fed to the LSTM.

In addition, we will expand the lexicon by using

more than 100 roots. Hence, we build not only

vocabulary of about 4000 words, but also a corpus of

around 60000 samples for training and testing

a hybrid CNN-LSTM model.

We notice that: 1) the LSTM model browses the

entire word from right to left and learns to focus on

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

218

schemes and flections rather than on the consonants

of the root. 2) We are confident that our model is

reliable thanks to the ability of memory cells. The

latter are able to store information and help neighbour

neurons acquire relevant knowledge. Our model fits

well with the Arabic lexicon philosophy by taking

into account the significant coherence between

prefixes and suffixes and their complicity in

incarnating the time of an action verb and its

corresponding pronoun.

6.4 Comparison with Related Works

Table 1 illustrates models that combined CNN and

LSTM for Arabic recognition. It also indicates the

rates obtained on specific datasets.

Table 1: Arabic rates recognition.

Model

Dataset

Result

BEM + CNN +

BLSTM + SVM for

an online Arabic

character recognition

(Hamdi et al.,

2019)

LMCA

99.11 %

Online-Khatt

93.98 %

CNN-LSTM for

Arabic sign language

translator (Agrawal

and Urolagin,

2020)

Minisi MN.

Arabic sign

language

dictionary

88.67 %

CNN-LSTM for

textual reviews (Al

Omari, Al-Hajj,

Sabra and

Hammami, 2019)

Main-AHS

88 %

Sub-AHS

96 %

Ar-Twitter

84 %

ASTD

79 %

OCLAR

90 %

CNN-LSTM for

an Arabic sentiment

analysis (Alayba,

Palade, England

and Iqbal, 2018)

Main-AHS

94 %

Sub-AHS

95 %

Our CNN model

for classification

APTI Dataset

97 %

Our CNN-LSTM

for word recognition

APTI Dataset

87.2 %

7 CONCLUSIONS

In this work, we proposed a deep recurrent approach

based on a hybrid CNN-LSTM model for tri-

consonantal Arabic word recognition.

Preliminary experiments were carried out on a

corpus including 17600 samples collected from the

APTI Dataset. The CNN model achieved a 97%

classification rate in root word recognition.

Preliminary experiments with the LSTM model

gave a score equal to 87.2%. The obtained results are

motivating. The combination of LSTM and CNN is

expected to be more and more interesting, due to the

significant coherence between the LSTM model,

(with a special focus on its concepts of forgetting and

memory), and the Arabic philosophy of constructing

words around the root, playing with prefixes, infixes

and suffixes and respecting Arabic derivations and

conjugation patterns.

Our future works will put special focus on: 1) We

will expand the vocabulary to handle with hundreds

of roots for training, validating and testing the

performance of the network on large datasets. 2).

Apply cross-validation to demonstrate the reliability

of our model. 3) Experiment the reliability of the

LSTM model to achieve the recognition, notably, of

the scheme, the tense and the pronoun, while

changing the way of word segmentation.

REFERENCES

Ibrahim, M.N., Bilmas, S.N, Babiker, A. (2013). A

framework of an online self-based learning for teaching

arabic as second language (TASL).Fifth International

Conference on Computational Intelligence, Modelling

and Simulation, pp. 255-260.

Touj, S., Ben Amara, N., Amiri, H. (2007). A hybrid

approach for off-line arabic handwriting recognition

based on a planar hidden markov modeling. ICDAR’07,

Brazil, 964- 968J.

Avila, J.M. (1996). Optimisation de modèles markoviens

pour la reconnaissance de l'écrit. PHD Thesis,

University of Rouen.

Cheriet, M., Beldjehem, M. (2006). Visual processing of

arabic handwriting: challenges and new

directions.SACH’06, India, pp 1-21.

Ben Hamadou, A. (1993). Vérification et correction

automatiques par analyse affixale des textes écrits en

langage naturel. PHD Thesis, Faculty of Sciences of

Tunis.

Kanoun, S., Alimi, A., Lecourtier, Y. (2005). Affixal

approach for arabic decomposable vocabulary

recognition: a validation on printed word in only one

font. ICDAR’05, Seoul.

A Deep Convolutional and Recurrent Approach for Large Vocabulary Arabic Word Recognition

219

Kammoun, W., Ennaji, A. (2004). Reconnaissance de

textes arabes à vocabulaire ouvert. CIFED, France.

LeCun, Y., Bengio, Y. (1995). Convolutional networks for

images, speech, and time series. The handbook of brain

theory and neural networks, 3361(10), 1995.Bai, J.,

Chen, Z., Feng, B. ,Xu, B. (2014). Image character

recognition using deep convolutional neural network

learned from different languages.IEEE International

Conference on Image Processing (ICIP), pp. 2560-

2564.

Yuan, A., Bai, G., Jiao, L., Liu, Y. (2012). Offline

handwritten english character recognition based on

convolutional neural network. Document Analysis

Systems (DAS), 10th IAPR International Workshop, pp.

125-129.

Wu, C., Fan, W., He, Y., Sun, J., Naoi, S. (2014).

Handwritten character recognition by alternately

trained relaxation convolutional neural network.

Frontiers in Handwriting Recognition (ICFHR), 14th

International Conference, pp. 291-296.

Zhong, Z., Jin, L., Feng, Z. (2015). Multi-font printed

chinese character recognition using multipooling

convolutional neural network.Document Analysis and

Recognition (ICDAR), 13th International Conference,

pp. 96-100.

Yang, W., Jin, L., Xie, Z., Feng, Z. (2015). Improved deep

convolutional neural network for online handwritten

chinese character recognition using domain-specific

knowledge. Proc. Document Analysis and Recognition

(ICDAR), 13th International Conference, pp. 551-555.

He, M., Zhang, S., Mao, H., Jin, L. (2015). Recognition

confidence analysis of handwritten chinese character

with CNN.Proc. Document Analysis and Recognition

(ICDAR), 13th International Conference, pp. 61-65.

Zhong, Z., Jin, L., Xie, Z. (2015). High performance offline

handwritten Chinese character recognition using

GoogLeNet and directional feature maps.Proc.

Document Analysis and R (ICDAR), 13th International

Conference, pp. 846-850.

Kim, I.-J., Xie, X. (2015). Handwritten hangul recognition

using deep convolutional neural networks.

International Journal on Document Analysis and

Recognition (IJDAR), 18, pp. 1-3.

Anil, R., Manjusha, K., Kumar, S.S., Soman, K.P. (2015).

Convolutional neural networks for the recognition of

malayalam characters.Proc. Document Analysis and

Recognition (ICDAR), 13th International Conference,

pp. 1041-1045.

Acharya, S., Pant, A.K., Gyawali, P.K. (2015). Deep

learning based large scale handwritten devenegari

character recognition. 9th International Conference on

Software, Knowledge, Information Management and

Applications (SKIMA), pp. 1-6.

Singh, P., Verma, A., Chaudhari, N.S. (2016). Deep

convolutional neural network classifier for handwritten

devanagari character recognition. Information Systems

Design and Intelligent Applications: Proceedings of

Third International Conference INDIA, 2, pp. 551-561.

Soman, S.T., Nandigam, A., Chakravarthv, V.S. (2013). An

efficient multiclassifier system based on convolutional

neural network for offline handwritten Teugu character

recognition.Proc. Communications (NCC), 2013

National Conference, pp. 1-5.

Fukushima, K. (1980). Neocognitron: A self-organizing

neural network model for a mechanism of pattern

recognition unaffected by shift in position. Biological

cybernetics, 36(4):193–202.

LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard,

R.E., Hubbardet, W., Jackel, L.D. (1990). Handwritten

digit recognition with a backpropagation network.

Advances in Neural Information Processing Systems,

pages 396–404.

Poisson, E., Viard-Gaudin, C. (2001). Réseaux de neurones

à convolution: reconnaissance de l’écriture manuscrite

non contrainte.Polytechnic School of the University of

Nantes, France.

Graves, A., Fernandez, S., Liwicki, M., Bunke, H.,

Schmidhuber, J. (2008). Unconstrained online

handwriting recognition with recurrent neural

networks.Advances in Neural Information Processing

Systems, 20:1-8.

Graves, A., Schmidhuber, J. (2009). Offline handwriting

recognition with multidimensional recurrent neural

networks.Advances in Neural Information Processing

Systems, 21:545-552.

Su, B., Lu, S. (2015). Accurate scene text recognition based

on recurrent neural network. In: Cremers D., Reid I.,

Saito H., Yang MH. (eds) Computer Vision -- ACCV

2014. ACCV 2014. Lecture Notes in Computer Science,

vol 9003. Springer, Cham.

Graves, A., Fernandez, S., Liwicki, M., Bunke, H.,

Schmidhuber, J. (2007). Unconstrained online

handwriting recongnition with recurrent neural

networks.Advances in Neural Information Processing

Systems.

Mioulet, L. (2015). Reconnaissance de l’écriture

manuscrite avec des réseaux récurrents, PhD Thesis,

University of Rouen.

Hamdi, Y., et al. (2019). Hybrid DBLSTM-SVM based

Beta-elliptic-CNN Models for Online Arabic

Characters Recognition. 2019 InternationalConference

on Document Analysis and Recognition (ICDAR).

IEEE.

Agrawal, T., Urolagin, S. (2020). 2-way Arabic Sign

Language translator using CNNLSTM architecture and

NLP.Accepted in ACM, Ei Compendex and Scopus, ISI

Web of Science, International Conference on Big Data

Engineering and Technology BDET-20.

Al Omari, M., Al-Hajj, M., Sabra, A., Hammami, N.

(2019).Hybrid CNNs-LSTM Deep Analyzer for Arabic

Opinion Mining.2019 Sixth International Conference

on Social Networks Analysis, Management and Security

(SNAMS).

Alayba, AM., Palade, V., England, M., Iqbal, R. (2018). A

combined CNN and LSTM model for arabic sentiment

analysis.In International Cross-Domain Conference for

machine learning and knowledge Extraction, pp.197-

191.

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

220