A Refined Multilingual Scene Text Detector Based on YOLOv7

Houssem Turki

1,3 a

, Mohamed Elleuch

2,3 b

and Monji Kherallah

National Engineering School of Sfax (ENIS), University of Sfax, Tunisia

National School of Computer Science (ENSI), University of Manouba, Tunisia

Advanced Technologies for Environment and Smart Cities (ATES Unit), University of Sfax, Tunisia

Keywords: Multilingual Scene Text Detection, YOLOv7, Specific Data Augmentation, Deep Learning.

Abstract: In recent years, significant advancements in deep learning and the recognition of text in natural scene images

have been achieved. Despite considerable progress, the efficacy of deep learning and the detection of

multilingual text in natural scene images often face limitations due to the lack of comprehensive datasets that

encompass a variety of scripts. Added to this is the absence of a robust detection system capable of

overcoming the majority of existing challenges in natural scenes and taking into account in parallel the

characteristics of each writing of different languages. YOLO (You Only Look Once) is a highly utilized deep

learning neural network that has become extremely popular for its adaptability in addressing various machine

learning tasks. YOLOv7 is an enhanced iteration of the YOLO series. It has also proven to be effective in

solving complex image-related problems thanks to the evolution of its 'Backbone' responsible for capturing

the features of images to overcome the challenges encountered in a natural environment which leads us to

adapt it to our text detection context. Our first contribution is to over-come environmental variations through

the use of specific data augmentation based on improved basic techniques and a mixed transformation method

applied to “RRC-MLT” and “SYPHAX” multilingual datasets which both contain Arabic scripts. The second

contribution is the refinement of the 'Backbone' block of the YOLOv7 architecture to better extract the small

details of the text which particularly stand out in Arabic scripts in punctuation marks. The article highlights

future research directions aimed at developing a generic and efficient multilingual text detection system in

the wild that also handles Arabic scripts, which is a new challenge that adds to the context, which justifies the

choice of the two datasets.

1 INTRODUCTION

Text detection is vital in the field of computer vision,

utilizing various Machine Learning (ML) and deep

learning models to improve the effectiveness of text

detection and associated tasks. Detecting text in the

wild within diverse environments poses several

challenges, including issues such as mixed languages,

complex character designs, and various types of

distortions like shape, size, multi-scale characters,

orientation, gradient illumination, blur, noise, low

contrast, complex background, etc. (see Figure 1).

The complexity is further heightened when dealing

with multilingual scripts (Bai et al., 2018). These

challenges exist in datasets specific to our context, but

there are few multilingual datasets (Saha et al., 2020)

https://orcid.org/0009-0001-8472-3622

https://orcid.org/0000-0003-4702-7692

https://orcid.org/0000-0002-4549-1005

and particularly limited studies that contain Arabic

text (Boujemaa et al., 2021) which also presents

totally different challenges from Latin or other

scripts; it comprises calligraphy, punctuation marks,

ligatures, skewed, multi-level baselines, multi-

position joining, etc. (Boukthir et al., 2022) (see

Figure 2). As well only a few studies concentrate on

the central issue of multilingual and varied scripts

which can exist in the same image with the aim of

developing a generic and resilient system capable of

taking into account this varied multitude of

challenges (Turki et al., 2023a). Among the many

algorithms developed for object and text detection,

the YOLO framework distinguishes itself as an

extensively utilized algorithm (Nourali et al., 2024).

During its evolution, the YOLO series of algorithms

512

Turki, H., Elleuch, M. and Kherallah, M.

A Reﬁned Multilingual Scene Text Detector Based on YOLOv7.

DOI: 10.5220/0013157100003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 512-519

ISBN: 978-989-758-737-5; ISSN: 2184-433X

has undergone iterative enhancements, its use in the

context of text detection in natural scenes images has

given remarkable results going from version 5 to

version 7 (Turki et al., 2023b). This inspires us to

employ the YOLOv7 and evaluate the detection

capabilities of this neural network architecture when

used for text detection in natural environments. Our

selection of YOLOv7 is based on the latest

advancements in the field and the promising

outcomes achieved (Khan et al., 2023). We make two

notable contributions:

Firstly, our primary contribution entails the novel

experimental study utilizing two multilingual datasets

comprising text in natural scenes images featuring the

Arabic scripts. The contribution involves selecting

and enhanced a specific data augmentation

techniques tailored for natural scenes, incorporating a

mixed transformation method designed specifically

for detecting text in dark area and low-light

conditions. In the second contribution, we will be

improving the architecture of YOLOv7 by refining

the ‘Backbone’ block through replacing in the first

step two ELAN modules with SwinV2_TDD

modules (Yang et al., 2023), which includes adding a

convolutional layer and an upsampling layer at the

beginning of each stage in the Swin Transformer v2

(Liu et al., 2022), the model is enhanced. This

modification strengthens the extraction of local

features and avoids over-compressing feature maps,

thereby enhancing the accuracy of detecting text and

small details, such as those found in Arabic writing.

In the second step we introduce a BiFormer Attention

Mechanism (Zhu et al., 2023) to direct the network's

attention toward more important features without

increasing the model complexity and to ensure the

model focuses more on text information during the

detection process.

The paper is organized into four sections as

follows. In Section 2, an overview of various related

works and their methodologies are presented,

highlighting frequently used methods. Section 3

presents the methodology and provides the

experimental assessment. Finally, Section 4

concludes the paper.

Figure 1: Samples of different challenging problems of text

in the wild.

Figure 2: Samples of different challenges in Arabic scripts.

2 RELATED WORK

Many applications take advantage of the capabilities

offered by deep learning and computer vision

algorithms to improve text detection in learning

processes (Turki et al., 2017). Object detection

encompasses the recognition of specific object

instances in images, videos, or live applications,

including the detection of text in diverse

environments of the wild (Amrouche et al., 2022).

YOLO is a commonly utilized technology for various

applications involving the detection of objects and

text (Redmon et al., 2016). It simplifies the detection

process by partitioning the input image into a grid,

enabling the prediction of bounding boxes and class

probabilities for each cell in the grid. YOLO has

become highly popular in the field of object detection

because of its simple architecture and low

complexity, ease of implementation (Ravi et al.,

2022). This algorithm adeptly recognizes objects and

their spatial positions by leveraging the information

provided by the bounding boxes (Diwan et al., 2023).

Our main goal is to adapt a recent YOLO architecture

for text detection, for this reason our contributions

focus on the development of a new refined

architecture of the YOLOv7 framework, effective for

multilingual text detection capable of overcoming the

different challenges of a natural environment as well

as the variety of scripts including those of the Arabic

language.

Implementing YOLO for text detection involves

improving its initial architecture and training the

algorithm to detect and locate text regions within an

image. In natural scenes, such as identifying text on

road signs or billboards in images related to

autonomous driving applications, can be achieved

using YOLO (Cui et al., 2024). This is valuable for

guidance and decision-making processes in

autonomous vehicles. In video surveillance systems,

YOLO can be used for real-time text detection

(Azevedo et al., 2024). Its application extends to

recognizing text in surveillance videos, such as

detecting license plates (Ramajo-Ballester et al.,

2024), identification of vehicle identification

numbers (VIN), traffic sign recognition (Kunekar et

al., 2024). In medical Image Analysis, YOLO can

A Reﬁned Multilingual Scene Text Detector Based on YOLOv7

513

assist in locating and extracting text from medical

images, including radiology reports, pathology slides,

and diagnostic imagery (Julia et al., 2024). In the

realm of augmented reality applications, YOLO can

contribute to text detection (Chu et al., 2024). It has

the capability to identify text in frames captured in

real-time. In Natural Language Processing (NLP) and

Chatbot Applications, YOLO can be adapted to detect

and extract text from images in the context of NLP

applications (Karakaya et al., 2024). In chatbots or

virtual assistants, YOLO-based text detection can be

used to understand and respond to textual information

present in images. In mobile Applications, YOLO can

be integrated into mobile applications to enable on-

device text detection for various purposes, such as

translating text in real-time or extracting information

from images (Hendrawan et al., 2024). Liu et al. (Liu

et al., 2022) improved YOLOv7 through the

incorporation of an attention mechanism into its main

structure, which allows extracting crucial features.

Chen et al. (Chen et al., 2022) improved YOLOv7

and then used it effectively in the recognition field to

develop an automated selection technique. Zheng et

al. (Zheng et al., 2022) improves the backbone of

YOLOv7 by accelerating model convergence and

making improvements including enhancing the non-

maximum suppression technique through the

application of grouped anchor frames during

ensembles of training data.

Through the training of YOLO on adequately

annotated datasets tailored for text detection, the

algorithm becomes proficient in accurately

identifying and localizing text regions within images.

This ability opens up opportunities for the automation

of tasks related to text in diverse domains.

3 METHODOLOGY

This study includes various experiments to obtain the

best improved model of YOLOv7 to ensure the best

detection of multilingual text in images of natural

scenes containing the Arabic script. The first

experiments of the work focus on the preparation of

the two selected multilingual datasets not chosen

randomly, through the use of targeted specific data

augmentation techniques. In this context, our second

refinement experiments focus on strengthening the

extraction of features related to the 'Backbone' block.

the details of the improvements will be described in

the following sections following a balanced strategy

aimed at obtaining a better generic detection system.

3.1 Datasets Employed in Experiments

For our experiments we chose the two datasets RRC-

MLT (Nayef et al., 2017) and SYPHAX (Turki et al.,

2023c). The first dataset RRC-MLT (Robust Reading

Challenge on Multi-lingual Text) is associated with

the ICDAR (International Conference on Document

Analysis and Recognition) competitions, specifically

the ICDAR 2019 Robust Reading Challenge on

Multi-lingual Text Detection and Recognition (RRC-

MLT). This dataset is designed to evaluate and

benchmark text detection and recognition methods on

multi-lingual and multi-script text in a variety of

scenarios, including different languages, fonts, and

text orientations. The challenges in the RRC-MLT

competition typically include tasks related to locating

and recognizing text in complex, real-world images.

The second dataset SYPHAX, was gathered in the

city of "Sfax," the second-largest city in Tunisia after

the capital. It encompasses a total of 2008 images,

with 401 allocated for testing and 1607 for training

across 7 different classes. The dataset was compiled

with a focus on addressing the dynamic challenges

associated with detecting text in natural scenes. The

images predominantly feature text in both Arabic and

Latin scripts. The SYPHAX dataset introduces a

novel and significant challenge in the realm of text

detection in the wild.

we chose the two datasets RRC-MLT and

SYPHAX because they contain, compared to other

benchmarks in the field, more scripts varied in terms

of language, challenges and text exposure. RRC-

MLT dataset contains texts in 9 different languages

with an equal number of images with varying degrees

of complexity at the script level and contain 2000

images with Arabic scripts, moreover this database is

part of an ICDAR2019 competition from which we

can compare our results obtained. SYPHAX dataset

contains three languages but focuses more on Arabic

language scripts and these related challenges; all the

images contain Arabic scripts. Table 1 summarizes

the comparison between the two selected datasets and

Figure 3 illustrates samples of their multilingual text

in the wild.

(a) RRC-MLT (b) SYPHAX datase

Figure 3: Samples of multilingual text in the wild from

RRC-MLT, and SYPHAX dataset.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

514

Table 1: Comparison between the two datasets RRC-MLT

and SYPHAX.

Dataset RRC-MLT SYPHAX

Size 18000 2008

Script Arabic, French,

Englis, Germa,

Bangla, Chines,

Italian, Japanese

and Korean

Arabic,

English and

French

Images with

Arabic scripts

2000 2800

Type of content Text line, Words Text line,

Words

Availability Public Public

Source Camera, Web

Camera

Text shape Linear, bending Linear,

ending

We carefully select each specific data

augmentation technique with consideration, ensuring

its alignment with the images found in natural settings

and its relevance to the characteristics of the chosen

dataset. The improvement of the chosen values of

each transformation is based mainly on

experimentation. This meticulous approach aims to

achieve effective results when applying text detection

methods. We choose 5 principal data augmentations

techniques based on basic image manipulations

(Shorten et al., 2019) (see Figure 4) and we add a

mixed transformation method.

3.1.1 The Basic Data Augmentations

Techniques

• Applying rigid geometric transformations

(Schaefer et al., 2006) is employed to generate

two additional skew angles, simulating difficult

camera angles on both the right and left sides.

• Applying a horizontal directional blur to

replicate the effect of capturing images in motion

from the camera (Naveed et al., 2024).

• Conducting color space transformations through

white balance adjustments to produce three

distinct color temperatures that simulate the

brightness variations during different times of the

day (He et al., 2024).

• Application of median filtering, coupled with

top-hat and bottom-hat transforms, serves to

eliminate Salt and Pepper Noise in images.

Additionally, the top-hat and bottom-hat

transforms are employed to enhance image

sharpness (Dong et al., 2024).

• Histogram Equalization (HE): Given that image

pixels typically cluster within specific intensity

ranges, histogram equalization is employed to

enhance contrast (Tiendrebeogo et al., 2024).

3.1.2 The Mixed Transformation Method

Certain images depicting natural scenes present a

challenge where text is situated in a dark region with

varying light and shadow, rendering the text

unreadable and challenging to locate. This difficulty

is particularly pronounced when dealing with close

color tones and grayscale distinctions between the

text and the background (refer to Figure 5). To

achieve these objectives, the mixed transformation

method serves a dual purpose. Firstly, it aims to

enhance image information by maximizing the

visibility of text in relation to the background.

Secondly, it artificially extends the size of the training

dataset. From a technical standpoint, the primary

function of the mixed transformation method relies on

the fusion of three maps: the original grayscale image

adding a sharpen filter, the Inversion map, and the

CLAHE map (refer to Figure 6). The Inversion step

facilitates the processed image to closely resemble the

grayscale map of the visible image, leading to a

notable improvement in recognition performance.

This adaptation improves the network's capability to

analyze the image, particularly boosting details in

white and gray within dark areas, thereby easing the

extraction of dark features (Joshi et al., 2023). On the

other hand, the CLAHE operation works to create a

more uniform grayscale distribution, simultaneously

boosting contrast and suppressing noise. This results

(a)

(b)

(c)

(d) (e)

(f)

(

)

(h)

(i)

Figure 4: Samples of the basic data augmentation

techniques: (a): original image (b): Right skew (c): left

skew (d): horizontal directional blur (e): color temperatures

1 (f): color temperatures 2 (g): color temperatures 3 (h):

median filtering & bottom-top-hat transforms (i): histogram

equalization.

A Reﬁned Multilingual Scene Text Detector Based on YOLOv7

515

in an increased level of detail information in the

image, with the processed pixel area becoming finer.

Consequently, the approach not only suppresses noise

but also enhances contrast (Mangal et al., 2024).

Figure 5: Samples of images with dark areas.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 6: Sample of a mixed transformation method: (a):

original image (b): grayscale image (c): Sharpen image (d):

Inversion map (e): CLAHE map (f): mixed combination.

3.2 Proposed Method for Backbone

Enhancement

3.2.1 The Network Architecture of YOLOv7

YOLOv7 is a real-time object detector belonging to

the YOLO family, known for its single-stage

architecture (Wang et al., 2023). The architecture of

YOLO is based on Fully Connected Neural

Networks. The YOLO framework comprises three

primary components: the Backbone, the Neck, and

the Head (see Figure 7). The main purpose of the

Backbone is to capture significant features present in

an image and transmit them to the Head via the Neck.

Initially, the images to be detected are processed in

the ‘Backbone’ section. Then, they undergo three

conventional convolutional operations as part of the

following process: convolution, batch normalization,

and SiLU activation. The MP module is a dual-

channel component responsible for performing

maximum pooling and convolution on the input

feature images. The SPPCSPC module represents a

spatial pyramid pooling structure designed to

overcome image distortion issues that may arise

during the scaling of input images. Following the

features extraction process within the Backbone

section, the image will advance to the Neck section

for the purpose of features fusion. The architecture

design of YOLOv7 is based on ELAN (Efficient

Layer Aggregation Network). ELAN strives to

establish a streamlined network by effectively

handling both the shortest and longest gradient paths.

The Neck section utilizes an FPN (Feature Pyramid

Network) network structure for this purpose. The

bottom-up features fusion process combines feature

images of various scales, resulting in a multi-scale

target detection effect (Patel et al., 2022). Ultimately,

the Head comprises output layers responsible for

producing the final detections. In YOLOv7, the head

that generates the final output is referred to as the

Lead Head. Figure 7 illustrates the complete

architecture of the YOLOv7 network structure. In

contrast to earlier versions, YOLOv7 was trained

solely on the MS COCO dataset (Redmon et al.,

2018), without the use of pre-trained backbones.

Figure 7: YOLOv7 architecture (Wang et al., 2023).

3.2.2 Replace ELAN with SwinV2_TDD

Modules

Our initial enhancement involves substituting the first

and last ELAN module within the Backbone block

with SwinV2_TDD (Yang et al., 2023). To enhance

the capture of local information and precisely detect

small areas, Swin Transformer v2 (Liu et al., 2022)

has been upgraded to SwinV2_TDD (Yang et al.,

2023). Since punctuation marks, ligatures and small

characters at multiple levels are often focused in

particular regions, it is essential to pay more attention

to local information. Among the advantages of

integrating SwinV2 TDD are the reduction in size of

feature maps and improving the efficiency of

calculations. This results in quicker feature extraction

while preserving high-quality outcomes and

decreasing memory usage. Secondly, it prevents the

excessive compression of feature maps by restoring

them to their original sizes and enhances their

expressive capability.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

516

3.2.3 Integration of BiFormer Attention

Mechanism

The BiFormer attention mechanism (Zhu et al., 2023)

is a novel visual Transformer model that combines a

two-layer routing attention approach with

Bidirectional Routing Attention (BRA). It operates

by iteratively transferring information to compute

attention weights, thereby achieving dynamic and

query-adaptive sparsity. This capability allows it to

handle intricate feature relationships while

minimizing computational costs. Through multiple

iterations, the two-layer routing attention can

dynamically adjust the attention weights. The

advantages of BiFormer attention mechanism are

multiple. First, the model's capacity to detect essential

features and contextual information within the image

can be enhanced, enabling it to concentrate more

effectively on small scripts. Second, The BiFormer

attention mechanism can capture the relationship

between the target and the background, helping to

produce more distinct feature representations. Figure

8 illustrates the final improvement of the Backbone

block of YOLOv7.

Figure 8: Final improved Backbone of YOLOv7.

3.3 Experiments and Results

It is crucial to note that the effectiveness of the text

detection algorithm using YOLOv7 may vary

depending on several factors involved; features of the

dataset utilized and the training methods employed.

Thus, in our study, the reported precision, recall, and

F-scores differ depending on the improved architecture

of YOLOv7, the inclusion of multilingual text with

Arabic script and the data augmentation methods

applied. The evaluation metrics are the precision, recall

and F-score defined as:

(1)

(2)

(3)

N represents the complete count of images within a

given dataset. |D

| and |G

| are the number of detection

and ground true rectangles in i

image. M

)

and M

( G

) are the matching scores for detection

rectangles D

and ground true rectangle G

The experimental outcomes of the proposed method

are promising as evidenced by the observed values in

Table 2. Moreover, the F-score rankings remain

consistent across the two multilingual datasets, which

confirms the efficacy of the method, providing

additional validation. Table 2 displays the

experimental results tested on SYPHAX dataset, we

can note that the enhanced YOLOv7 obtains the best

result in precision (87.3%) and the F-score (82.41%);

the language of the scripts is distributed equally

between all the images in the database and the

improvement of the architecture of YOLOv7 is well

suited especially to the Arabic language. On the other

hand, the tested on RRC-MLT dataset, note that we

obtain the best result in the recall (79.12%); which also

proves the impact of strengthening feature extraction

even at the script diversity level. From the point of

view of result values, this enhancement is especially

remarkable in recognizing texts of diverse sizes and

various scripts. These findings validate the credibility

of the method introduced in this article. From the point

of view of classification of results, this study is

encouraging and warrants further exploration and

investigation especially in this context of multiple

challenges, which is shown in Table 3 to compare the

results obtained with those of the ICDAR2019

competition (Nayef et al., 2019).

Table 2: Experimental result of the proposed method.

Dataset Precision

(%)

Recall

(%)

F score (%)

SYPHAX

87.3 78.05 82.41

RRC-MLT

83.75 79.12

81.36

Table 3: Results compared to the ICDAR 2019 competition

(challenge for task-1: multi-lingual text detection) (Nayef

et al., 2019).

RRC-MT Dataset Precision

(%)

Recall

(%)

F score

(%)

Tencent-DPPR

Team

87.52 80.05 83.61

Proposed method

83.75 79.12 81.36

MM-MaskRCNN

84.73 70.21 76.79

MLT2019 ETD

78.71 54.44 64.36

Cyberspace

69.48 35.61 47.09

3.4 Discussion

The ‘Backbone’ is one of the most crucial

components in YOLOv7, significantly influencing

A Reﬁned Multilingual Scene Text Detector Based on YOLOv7

517

the performance of the entire detection model. By

improving this block by replacing ELAN modules

and incorporating the BiFormer attention mechanism,

the model can more effectively utilize spatial and

semantic information in the image, thereby enhancing

the representation of text during feature extraction.

This enhances the model's generalization ability and

robustness while decreasing its reliance on other

stages. Therefore, these modifications to the

YOLOv7 backbone can enhance the model’s

detection accuracy for small targets and improve the

feature extraction stage, effectively leveraging key

features and contextual information in the image.

This design choice can enhance text detection

performance and offer more precise feature

representation for subsequent processing stages,

resulting in improved detection outcomes (see Figure

9).

Figure 9: Text detection samples based on the proposed

method.

4 CONCLUSION

This research paper introduces a novel method for

multilingual text detection in the wild, utilizing the

YOLOv7 model optimized through specific data

augmentation techniques. The proposed method

improves the YOLOv7 model by enhancing the

features extraction capabilities of the backbone

network and takes into account the characteristics of

Arabic language scripts rarely treated in the context

of multilingual detection. The effectiveness of this

method is showcased through experiments performed

on two multilingual datasets. The obtained results are

promising, indicating that the enhanced YOLOv7

model has the potential to further advance by

incorporating targeted enhancements to improve text

detection performance across various multilingual

scripts. This may involve modifications to the

network architecture and refining training strategies.

As YOLOv7 is a relatively recent development, it

holds significant promise for future research and

exploration in the field of text detection.

REFERENCES

Amrouche, A., Bentrcia, Y., Hezil, N., Abed, A., Boubakeur,

K. N., & Ghribi, K. (2022, November). Detection and

localization of arabic text in natural scene images. In

2022 First International Conference on Computer

Communications and Intelligent Systems (I3CIS) (pp.

72-76). IEEE.

Azevedo, P., & Santos, V. (2024). Comparative analysis of

multiple YOLO-based target detectors and trackers for

ADAS in edge devices. Robotics and Autonomous

Systems, 171, 104558.

Bai, X., Yang, M., Lyu, P., Xu, Y., & Luo, J. (2018).

Integrating scene text and visual appearance for fine-

grained image classification. IEEE Access, 6, 66322-

66335.

Boujemaa, K. S., Akallouch, M., Berrada, I., Fardousse, K.,

& Bouhoute, A. (2021). ATTICA: a dataset for arabic

text-based traffic panels detection. IEEE Access, 9,

93937-93947.

Boukthir, K., Qahtani, A. M., Almutiry, O., Dhahri, H., &

Alimi, A. M. (2022). Reduced annotation based on deep

active learning for arabic text detection in natural scene

images. Pattern Recognition Letters, 157, 42-48.

Cui, Y., Guo, D., Yuan, H., Gu, H., & Tang, H. (2024).

Enhanced YOLO Network for Improving the

Efficiency of Traffic Sign Detection. Applied Sciences,

14(2), 555.

Chu, C. H., & Liu, S. (2024). Virtual Footwear Try-On in

Augmented Reality Using Deep Learning Models.

Journal of Computing and Information Science in

Engineering, 24(3), 031002.

Chen, J., Liu, H., Zhang, Y., Zhang, D., Ouyang, H., &

Chen, X. (2022). A Multiscale Lightweight and

Efficient Model Based on YOLOv7: Applied to Citrus

Orchard. Plants, 11(23), 3260.

Diwan, T., Anirudh, G., & Tembhurne, J. V. (2023). Object

detection using YOLO: Challenges, architectural

successors, datasets and applications. Multimedia Tools

and Applications, 82(6), 9243-9275.

Dong, J., Wang, N., Fang, H., Lu, H., Ma, D., & Hu, H.

(2024). Automatic augmentation and segmentation

system for three-dimensional point cloud of pavement

potholes by fusion convolution and transformer.

Advanced Engineering Informatics, 60, 102378.

Hendrawan, A., Gernowo, R., Nurhayati, O. D., & Dewi, C.

(2024). A Novel YOLO-ARIA Approach for Real-

Time Vehicle Detection and Classification in Urban

Traffic. International Journal of Intelligent Engineering

& Systems, 17(1).

He, W., Zhang, C., Dai, J., Liu, L., Wang, T., Liu, X., ... &

Liang, X. (2024). A statistical deformation model-based

data augmentation method for volumetric medical image

segmentation. Medical Image Analysis, 91, 102984.

Julia, R., Prince, S., & Bini, D. (2024). Medical image

analysis of masses in mammography using deep

learning model for early diagnosis of cancer tissues. In

Computational Intelligence and Modelling Techniques

for Disease Detection in Mammogram Images (pp. 75-

89). Academic Press.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

518

Joshi, G., Natsuaki, R., & Hirose, A. (2023). Neural

Network Fusion Processing and Inverse Mapping to

Combine Multisensor Satellite Data and Analyze the

Prominent Features. IEEE Journal of Selected Topics in

Applied Earth Observations and Remote Sensing, 16,

2819-2840.

Khan, N., Ahmad, R., Ullah, K., Muhammad, S., Hussain,

I., Khan, A., ... & Mohamed, H. G. (2023). Robust

Arabic and Pashto Text Detection in Camera-Captured

Documents Using Deep Learning Techniques. IEEE

Access, 11, 135788-135796.

Kunekar, P., Narule, Y., Mahajan, R., Mandlapure, S.,

Mehendale, E., & Meshram, Y. (2024). Traffic

Management System Using YOLO Algorithm.

Engineering Proceedings, 59(1), 210.

Karakaya, M., Ersoy, S., Feyzioğlu, A., & Ersoy, S. (2024).

Reading Gokturkish text with the Yolo object detection

algorithm. Journal of Mechatronics and Artificial

Intelligence in Engineering.

Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., ... & Guo,

B. (2022). Swin transformer v2: Scaling up capacity

and resolution. In Proceedings of the IEEE/CVF

conference on computer vision and pattern

recognition (pp. 12009-12019).

Liu, S., Wang, Y., Yu, Q., Liu, H., & Peng, Z. (2022).

CEAM-YOLOv7: Improved YOLOv7 Based on

Channel Expansion and Attention Mechanism for

Driver Distraction Behavior Detection. IEEE Access,

10, 129116-129124.

Mangal, A., Garg, H., & Bhatnagar, C. (2024). A Robust

Co-saliency Object Detection Model by Applying

CLAHE and Otsu Segmentation Method. International

Journal of Intelligent Systems and Applications in

Engineering, 12(1s), 481-490.

Nourali, K., & Dolkhani, E. (2024). Scene text visual

question answering by using YOLO and STN.

International Journal of Speech Technology, 1-8.

Nayef, N., Yin, F., Bizid, I., Choi, H., Feng, Y., Karatzas, D.,

... & Ogier, J. M. (2017, November). Icdar2017 robust

reading challenge on multi-lingual scene text detection

and script identification-rrc-mlt. In 2017 14th IAPR

international conference on document analysis and

recognition (ICDAR) (Vol. 1, pp. 1454-1459). IEEE.

Naveed, H., Anwar, S., Hayat, M., Javed, K., & Mian, A.

(2024). Survey: Image mixing and deleting for data

augmentation. Engineering Applications of Artificial

Intelligence, 131, 107791.

Nayef, N., Patel, Y., Busta, M., Chowdhury, P. N.,

Karatzas, D., Khlif, W., ... & Ogier, J. M. (2019,

September). Icdar2019 robust reading challenge on

multi-lingual scene text detection and recognition—rrc-

mlt-2019. In 2019 International conference on

document analysis and recognition (ICDAR) (pp. 1582-

1587). IEEE.

Patel, K., Bhatt, C., & Mazzeo, P. L. (2022). Improved Ship

Detection Algorithm from Satellite Images Using

YOLOv7 and Graph Neural Network. Algorithms,

15(12), 473.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016).

You only look once: Unified, real-time object detection.

In Proceedings of the IEEE conference on computer

vision and pattern recognition (pp. 779-788).

Ravi, N., & El-Sharkawy, M. (2022). Real-Time Embedded

Implementation of Improved Object Detector for

Resource-Constrained Devices. Journal of Low Power

Electronics and Applications, 12(2), 21.

Ramajo-Ballester, Á., Moreno, J. M. A., & de la Escalera

Hueso, A. (2024). Dual license plate recognition and

visual features encoding for vehicle identification.

Robotics and Autonomous Systems, 172, 104608.

Redmon, J., & Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv preprint arXiv:1804.02767.

Saha, S., Chakraborty, N., Kundu, S., Paul, S., Mollah, A.

F., Basu, S., & Sarkar, R. (2020). Multi-lingual scene

text detection and language identification. Pattern

Recognition Letters, 138, 16-22.

Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on

image data augmentation for deep learning. Journal of

big data, 6(1), 1-48.

Schaefer, S., McPhail, T., & Warren, J. (2006). Image

deformation using moving least squares. In ACM

SIGGRAPH 2006 Papers (pp. 533-540).

Turki, H., Elleuch, M., Kherallah, M., & Damak, A. (2023a,

September). Arabic-Latin Scene Text Detection based

on YOLO Models. In 2023 International Conference

on Innovations in Intelligent Systems and Applications

(INISTA) (pp. 1-6). IEEE.

Turki, H., Elleuch, M., & Kherallah, M. (2023b,

November). Multi-lingual Scene Text Detection

Containing the Arabic Scripts Using an Optimal then

Enhanced YOLO Model. In International Conference

on Model and Data Engineering (pp. 47-61). Cham:

Springer Nature Switzerland.

Turki, H., Elleuch, M., Kherallah, M. (2023c). SYPHAX

Dataset. IEEE ataport. https://dx.doi.org/10.21227/ydqd-

2443.

Turki, H., Halima, M. B., & Alimi, A. M. (2017,

November). Text detection based on MSER and CNN

features. In 2017 14th IAPR international conference

on document analysis and recognition (ICDAR) (Vol.

1, pp. 949-954). IEEE.

Tiendrebeogo, A. (2024). Identification of plants from the

convolutional neural network. Multimedia Tools and

Applications, 1-11.

Wang, C. Y., Bochkovskiy, A., & Liao, H. Y. M. (2023).

YOLOv7: Trainable bag-of-freebies sets new state-of-

the-art for real-time object detectors. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition (pp. 7464-7475).

Yang, Y., & Kang, H. (2023). An enhanced detection

method of PCB defect based on improved

YOLOv7. Electronics, 12(9), 2120.

Zhu, L., Wang,X., Ke, Z., Zhang, W., & Lau, R. W. (2023).

Biformer: Vision transformer with bi-level routing

attention. In Proceedings of the IEEE/CVF conference

on computer vision and pattern recognition (pp. 10323-

10333).

Zheng, J., Wu, H., Zhang, H., Wang, Z., & Xu, W. (2022).

Insulator-Defect Detection Algorithm Based on

Improved YOLOv7. Sensors, 22(22), 8801.

A Reﬁned Multilingual Scene Text Detector Based on YOLOv7

519