IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 2, June 2023, pp. 577∼584

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i2.pp577-584 r 577

An image-based convolutional neural network system for

road defects detection
Mohamed Anis Benallal, Mustapha Si Tayeb
Laboratoire SIMPA (Signal-Image-Parole), Département d’informatique, Faculté des Mathématiques et Informatique, Université des
Sciences et de la Technologie d’Oran Mohamed Boudiaf (USTO-MB), Oran, Algeria

Article Info ABSTRACT

Article history: An application of convolutional neural network (CNN) technique for road sur-
face defects detection is presented in this paper. You only look ones (YOLO)
Received Feb 25, 2022
algorithm showed its capabilities as an effective object detection technique in
Revised Sep 27, 2022 many previous works for different problems. Road damages detection and clas-
Accepted Oct 27, 2022 sification is one of the most challenging problems faced by public and private
road management agencies. We present here results for a first attempt on apply-
Keywords: ing YOLO to detect cracks and potholes, the most common defects encountered
in surface roadways. Image database of the Brazilian highways were used to
Convolutional neural network prepare input data, train the model and test it. Despite considering different
Defects detection types of cracks in one class and a less amount of potholes images, results show
Image recognition that the YOLO algorithm performs well with a global rate of 91% of defect de-
Road damages tection. Output results analysis induce us to work on providing a local database
You only look ones for Algerian roadways with a large number of defect images/videos, as well as
producing an automatic road-dedicated defects detector device.

This is an open access article under the CC BY-SA license.

Corresponding Author:
Mohamed Anis Benallal
Laboratoire SIMPA, Département d’informatique, Faculté des Mathématiques et Informatique, Université des
Sciences et de la Technologie d’Oran Mohamed Boudiaf (USTO-MB)
BP 1505, El M’naouer, Oran, Algeria
Email: mohamed.benallal@univ-usto.dz

1. INTRODUCTION
Countries around the world are likely to encounter the current issues with infrastructure management
and maintenance [1]. Traditionally, monitoring the road surface condition was done manually, but this is still
the case in our country. There are many ways to build a road. However, the great majority of our roads are
constructed of an asphalt concrete layer, followed by one or more layers of angular rock, known as the base
course (also called tarmac). Despite being a cost-effective solution, as it provides smooth and durable surface,
roads suffer from a set of problems. The most encountered ones are cracks and potholes. Sunlight, rain, snow,
sleet, and freezing weather, in addition to cars and trucks traffic contribute to the slow process of altering the
road surface.
Since roadways infrastructure is one of the highest value assets owned by a country, maintaining it has
a huge prevalence and a direct effect on the economy and the prosperity of people and businesses in a country,
that is, to extend the life of an asphalt surface, it needs to be repaired as soon as a defect appears. That means
that detecting a damage earlier provides a way of sealing and maintaining a correct road surface quality, and its

Journal homepage: http://ijai.iaescore.com

578 r ISSN: 2252-8938

importance has no longer to be proved.

Object and features detection took a huge place in computer vision these last decades and on of its
most boost is the emergence of convolutional neural networks (CNN), mainly by providing a better perfor-
mance in visualizing and classifying images [2]. They have led to a rapid development of different fields of
studies (identification [3], [4], medical [5], [6], automotive [7], and [8]). In the field of object detection, the
most recent cutting-edge algorithms are faster region based CNN (Faster R-CNN) [3], [9]–[12], single shot
multibox detector (SSD) [11], [13], [14] and you only look once (YOLO) [10], [11], [15]. This last one, as
a deep learning-based model, have been considered to be the most efficient approach to detecting objects and
characteristics such as surface cracks [16], but also a strong and competitive approach for detecting potholes
[11]. In computer vision, identifying objects in images includes detection and localization but also classifi-
cation. The YOLO algorithm was selected due to its excellent accuracy and speed on achieving these tasks.
Although other approaches have demonstrated comparable accuracy on diverse datasets, YOLO exhibits the
highest speed in many object detection applications, such as detecting non-certified work on construction sites
[17], object detection for disaster response and recovery [18], and soldering defect detection [19]. Along with
the recent improvements of GPUs (graphics processing unit), an impressive improvement in accuracy and speed
for real-time applications area is observed this last decade.
Automatic road surface monitoring is one of the most challenging problems to manage for road mon-
itoring agencies. Several works in all over the world have been done so far. Techniques differ in terms of input
data. In fact, there are three main ways which are investigated :
− Vibrations : as it can be sampled using a very affordable equipment, this technique seems to be very
popular, however, being in contact with the damage is a must, for example falling in a pothole to have its
signal [20]–[22].
− laser/ultrasonic : the main goal is characterizing defects and modelling the shape and size of each one.
The main disadvantage of these techniques is the fact that equipment is usually very expensive. The
exceptional use of them remains rare [23], [24], in fact, research involving such equipment is usually
hybridized with images techniques [25]–[27].
− Image analysis and exploitation : which balance between the two above, this is the most common nowa-
days.
Actually, image-based techniques remain to be the most studied way for detecting and monitoring road
defects. The multiplicity of camera types (digital single-lens reflex (DSLR), parking, and smartphone) produces
a gigantic source of data. Moreover, capturing images provides a real and visual idea of the entire scene
(also analogically similar to the traditional way of monitoring roads). For the maintenance and monitoring
of pavement surfaces, automated image analysis methods have been implemented over the years. Images can
bring a lot of useful information. In general, cameras are placed in front top or in the rear of the vehicle.
As they are the most encountered, potholes and cracks are the most studied defects [16], [28]. Other
works tend to work on detecting defects, but more importantly, classifying them as much as possible [1], [29].
In addition to these two main ways of working on images, other works such as [30] focus on identifying defects
by reconstructing them in three-dimensional (3D) models. The most important advantage that this technique
has today is the capacity of handling such data in real-time, especially by the increase of computer processing
capacity in one hand, but also the use of new generation of algorithms such as deep learning in another hand.
Based on our review on the subject, YOLO applications were previously used in different scenarios.
Li et al. [31], ground penetrating radar equipment was used to acquire concealed cracks in asphalt pavement.
The use of YOLOv4 as a single class detector on these radar images shown the best overall performance of
the whole YOLO series. On the one hand, its detection speed is very rapid, and it can still be investigated to
a greater extent utilizing a GPU, thanks to its reduced inference time and higher FPS (frame per second). The
higher level of confidence and robustness, on the other hand, indicates that it has a relatively reliable perfor-
mance in precisely detecting concealed cracks. Zhang et al. [32], YOLOv3 was compared to an improved
version (combined with MobileNets and convolutional block attention module) to tackle bridge surface cracks.
Results showed that the modified architecture of the algorithm can perform at a higher FPS rate but slightly
influence the precision of the model. Du et al. [33], a comparison of the YOLO series was performed against
several pavement distress using a local road dataset. Results showed that the YOLO network can be used to
forecast potential distress location and category. Moreover, comprehensive detection accuracy reached 73.64%,
and the processing speed was just 70% SSD and 9 times faster than Faster R-CNN. This paper presents prelim-

Int J Artif Intell, Vol. 12, No. 2, June 2023 : 577 – 584
Int J Artif Intell ISSN: 2252-8938 r 579

inary results of an experimental application of the YOLO deep-learning technique on tackling the problem of
detecting cracks and potholes based on road images, that is, more investigation should be provided by applying
YOLO on other regions.

2. METHOD
This section presents the used algorithm and how it can be applied on our study. The details about
the used database are also presented to explain how data are used in the algorithm. In fact, raw images of road
surface had been prepared and sampled through different steps.

2.1. YOLO algorithm

First published in 2015 by a team of researchers from the University of Washington whose head of
this project Joseph Redmon [34], YOLO is an automatic detection algorithm which unlike other methods uses
a totally different approach. It represents a CNN for object detection which can be used in real time, and as its
name indicates it looks at the processed image only once. In fact, this makes the method extremely faster at the
testing phase compared to other approaches [35].
Due to the fact that YOLO is a supervised method, the training dataset must first be manually con-
structed and contain bounding box information for road defects. Parameters used to describe the bounding
box information are : center X and Y, width and height. To perform this task, we used the Yololabel software
(an image-labeling tools for object detection), more details are presented in section 2.2.. Figure 1 explains the
main four steps that YOLO performs to provide the bounding boxes according to the detected objects in the
image. In this sequence of images, we can see that the image has been split into several squares (top-left) which
generate several random anchor boxes (top-right).

Figure 1. The YOLO flowchart algorithm

The intersection over union (IoU) is then calculated to group together the anchor boxes that detect the
same object (we can see this through the groups of colours in the bottom-right). YOLO uses IoU to provide
a box that surrounds the whole targeting object. Basically, it divides the overlap area between the boxes by

An image-based convolutional neural network system for road defects detection (Mohamed Anis Benallal)
580 r ISSN: 2252-8938

the union of those areas. Then finally, thanks to the non-max suppression algorithm included in the YOLO
algorithm, we go from thousands of anchor boxes to only 5 in this example (one for cracks and four for
potholes in the bottom-left).

2.2. Dataset preparation

To perform our application of the YOLO algorithm, different databases are available online. We
choose to use the one provided in [36]. Note that there is no local available database at the time of this paper.
However, we are working on providing a new one, especially dedicated to Algerian roadways for future works.
The used dataset includes pictures of defects on Brazilian asphalt roads. This dataset was developed
using pictures provided by DNIT (National Department of Transport Infrastructure, Brazil. Available through
the access to information law - protocol 50650.003556 / 2017-28). Capturing process was done on highways
in the state of Espirito Santo, Rio Grande do Sul and the Federal District, between 2014 and 2017. To use this
database, we went through each images and followed these steps :
− Check every one and select images, which contain a set of the studied defects, a total of 665 was obtained.
− Use the YoloLabel software to manually annotate the database. This step is the most time-consuming part
of the database preparation. It mainly consists of selecting the appropriate class and drawing bounding
boxes around the appropriate defect. The software then provides a text file for each image that contains
information about boxes positions and dimensions.
The final step is to upload all files (database and needed YOLO files) into a Google drive repository
to be easily used in the processing. Figure 2 presents examples of both classes, which are cracks in Figure 2(a)
and potholes in Figure 2(b). Note that these examples represent the sampled defect as it is injected into the
YOLO algorithm for the training process (raw images along with text files explained above). As shown, we
can assume that the shape, the orientation, the dimension, and even the size are quite different from one defect
to another in both classes.

(a) (b)

Figure 2. Examples of the studied defects (YOLO algorithm inputs); (a) cracks example and (b) potholes
examples

3. RESULTS AND DISCUSSION

All the experiment presented here have been done using Colab, the Google Colaboratory platform that
offers the possibility of executing Python code directly in the browser. This platform, which is intended for
machine learning training and research, enables us to train machine learning models directly in the cloud. As
a first attempt, we would like to answer the question of how does the YOLO algorithm work on classifying
the two main encountered road defects (cracks and potholes)? For this purpose, we prepared the database
by selecting bounding boxes for these two classes. Note that we deliberately choose to put multiple types
of cracks (horizontal/vertical/diagonal (an unconnected crack which generally takes a horizontal, vertical, or
diagonal line across a pavement, crocodile (interconnected or interlocked fissures that look like a crocodile’s
skin or a collection of tiny polygons), transverse (rupture across the pavement) [37]) to understand how it
affects the learning process. Potholes are defined as a unique class described as irregularly shaped holes of
various sizes in the pavement [37].

Int J Artif Intell, Vol. 12, No. 2, June 2023 : 577 – 584
Int J Artif Intell ISSN: 2252-8938 r 581

Working with YOLO implies defining a set of parameters to perform the learning process. In fact, it
is implemented based on the darknet, an open source neural networks. Note that we used the v4 version of the
YOLO algorithm [38]. Actually , this version remains the best choice to balance between accuracy and custom
configuration. Having two classes, we need to set the configuration file according to, that is: line classes to
2, max batches to the minimum of 6,000, line steps to 80% and 90% of max batches and filters to 21. All of
these parameters work in a network sized at 416 by 416. Due to Colab limitations (in its free version), a total of
5,700 iterations were performed. To test the performance of our model, and in addition to the selected images
for training, we also selected 650 images from the used database in which 450 visually contain similar learned
defects and the rest of 200 images which contain no learned defects (to test false positive cases) or no defects
at all. Note that these 650 images have not been used in any step of the learning process. Indeed, we consider
that doing so ensures a real and true test to the generalization capacity of the trained network.
To provide this step, we injected all these testing images to the model and then checked manually all
the resulting images one by one. For every defect (cracks in all types and potholes), we went through counting
the correctness of the detection. Note that we can have many defects in the same image. Figure 3 presents
examples of a visual representation for the obtained results. As we can see in its sub-figures, boxes provided
by the YOLO algorithm make it capable of: i) avoiding any detection where there is no defect (Figure 3(a)), ii)
Detecting multiple defects of the same class (Figure 3(b) for cracks and Figure 3(c) for potholes), and detecting
multiple defects of the different classes (Figure 3(d)).

(a) (b)

(c) (d)

Figure 3. Examples of the algorithm result detections; (a) no detection, (b) cracks detection, (c) potholes
detection, and (d) both crack and pothole detection

Table 1 shows that the testing step, considering the testing dataset as different from the training one
and with a comparable number of images, brings us a general detection rate of 91% with approximately 90%
and 93% for cracks and potholes respectively. Furthermore, we got an insignificant number of false positive
cases for both classes. These results show that, even with taking into consideration multiple types of cracks
in the same class, this gives an acceptable detection rate. However, this could be certainly improved with
creating multiple classes which anyone correspond to a specific type of cracks. For potholes, even if results are

An image-based convolutional neural network system for road defects detection (Mohamed Anis Benallal)
582 r ISSN: 2252-8938

similar to those of cracks, we consider that we should have a larger number of images to be more consistent.
Nevertheless, that lead us to work on other databases, but also create and publish a local one(since local roads
significantly suffer from poor road conditions).

Table 1. Model testing results

Detected Not detected False positive
Number % Number % Number
Cracks 335 90% 38 10% 2
Potholes 95 93% 7 7% 4
In total 430 91% 45 9% 6

4. CONCLUSION
We presented in the paper an experimental test on detecting the two main classes of road defects
(cracks and potholes). The observed lack of studies involving YOLO as a multiple and dedicated road-damages
detector induce us to answer the question of, to what extent, the YOLO can be an algorithm to consider using
in road surface damages detection and classification. An overall rate of 91% (90% and 93% for cracks and
potholes respectively) was obtained using a Brazilian road images database. As a first attempt, we consider
these results favourable to more investigations. Furthermore, this comforts us in using the YOLO algorithm as
model-based technique to provide a more powerful and versatile system to detect better cracks and potholes,
but also work on other kinds of defects. Therefore, this lead us to focus on two main aspects to improve our
out-coming results: i) capture videos/images, in order to create a local database of Algerian road defects (work
in progress at the time of this paper to provide more images of the studied problem but also to build an effective
local detector), and ii) take into consideration the differences between defect features of the same class. In fact,
our hypothesis about the observed results of cracks detection rate is mainly due to the choice of combining
multiple cracks types, in terms of directions and shape. Overall, results are very motivating to explore the
model optimisation in terms of input data, meta-parametrization, and model training.

REFERENCES
[1] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata, “Road damage detection and classification using deep
neural networks with smartphone images,” Computer-Aided Civil and Infrastructure Engineering, vol. 33, no. 12, pp.
1127–1141, 2018, doi: 10.1111/mice.12387.
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,”
Communications of the ACM, vol. 60, no. 6, pp. 84–90, May 2017, doi: 10.1145/3065386.
[3] I. Hachchane, A. Badri, A. Sahel, I. Elmourabit, and Y. Ruichek, “Image and video face retrieval with query image
using convolutional neural network features,” IAES International Journal of Artificial Intelligence (IJ-AI), vol. 11,
no. 1, pp. 102-109, Mar. 2022, doi: 10.11591/ijai.v11.i1.pp102-109.
[4] N. A-Alam, M. Ahsan, M. A. Based, J. Haider, and M. Kowalski, “An intelligent system for automatic fingerprint
identification using feature fusion by Gabor filter and deep learning,” Computers and Electrical Engineering, vol. 95,
p. 107387, Oct. 2021, doi: 10.1016/j.compeleceng.2021.107387.
[5] K. S. Ananda Kumar, A. Y. Prasad, and J. Metan, “A hybrid deep CNN-Cov-19-Res-Net transfer learning architype
for an enhanced brain tumor detection and classification scheme in medical image processing,” Biomedical Signal
Processing and Control, vol. 76, p. 103631, Jul. 2022, doi: 10.1016/j.bspc.2022.103631.
[6] S. A. Shadab, M. A. Ansari, N. Singh, A. Verma, P. Tripathi, et R. Mehrotra, “Detection of cancer from histopathol-
ogy medical image data using ML with CNN ResNet-50 architecture,” in Computational Intelligence in Healthcare
Applications, Elsevier, 2022, pp. 237-254, doi: 10.1016/B978-0-323-99031-8.00007-7.
[7] M. Sha and A. Boukerche, “Performance evaluation of CNN-based pedestrian detectors for autonomous vehicles,”
Ad Hoc Networks, vol. 128, p. 102784, Apr. 2022, doi: 10.1016/j.adhoc.2022.102784.
[8] S. Narmadha and V. Vijayakumar, “Spatio-temporal vehicle traffic flow prediction using multivariate CNN and LSTM
model,” Materials Today: Proceedings, May 2021, doi: 10.1016/j.matpr.2021.04.249.
[9] W. S. Aji, K. H. Bin Ghazali, and S. A. Akbar, “Oil palm unstripped bunch detector using modified faster regional
convolutional neural network,” IAES International Journal of Artificial Intelligence (IJ-AI), vol. 11, no. 1, pp. 189-
200, Mar. 2022, doi: 10.11591/ijai.v11.i1.pp189-200.
[10] M. A. Anwer, S. M. Shareef, and A. M. Ali, “Accident vehicle types classification: a comparative study between
different deep learning models,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol.
21, no. 3, pp. 1474-1484, Mar. 2021, doi: 10.11591/ijeecs.v21.i3.pp1474-1484.

Int J Artif Intell, Vol. 12, No. 2, June 2023 : 577 – 584
Int J Artif Intell ISSN: 2252-8938 r 583

[11] K. Gayathri and S. Thangavelu, “Novel deep learning model for vehicle and pothole detection,” Indonesian Jour-
nal of Electrical Engineering and Computer Science (IJEECS), vol. 23, no. 3, pp. 1576-1582, Sep. 2021, doi:
10.11591/ijeecs.v23.i3.pp1576-1582.
[12] J. Shi, Y. Chang, C. Xu, F. Khan, G. Chen, and C. Li, “Real-time leak detection using an infrared cam-
era and Faster R-CNN technique,” Computers & Chemical Engineering, vol. 135, p. 106780, Apr. 2020, doi:
10.1016/j.compchemeng.2020.106780.
[13] D. Biswas, H. Su, C. Wang, A. Stevanovic, and W. Wang, “An automatic traffic density estimation using single shot
detection (SSD) and MobileNet-SSD,” Physics and Chemistry of the Earth, Parts A/B/C, vol. 110, p. 176-184, Apr.
2019, doi: 10.1016/j.pce.2018.12.001.
[14] W. Liu et al., “SSD: single shot multibox detector,” in European conference on computer vision (ECCV 2016),
Springer International Publishing, 2016, pp. 21–37. doi: 10.1007/978-3-319-46448-0 2.
[15] J. Azimjonov and A. Özmen, “A real-time vehicle detection and a novel vehicle tracking systems for estimating
and monitoring traffic flow on highways,” Advanced Engineering Informatics, vol. 50, p. 101393, Oct. 2021, doi:
10.1016/j.aei.2021.101393.
[16] Y.-J. Cha, W. Choi, and O. Büyüköztürk, “Deep learning-based crack damage detection using convolutional neu-
ral networks,” Computer-Aided Civil and Infrastructure Engineering, vol. 32, no. 5, pp. 361–378, May 2017, doi:
10.1111/mice.12263.
[17] Q. Fang et al., “A deep learning-based method for detecting non-certified work on construction sites,” Advanced
Engineering Informatics, vol. 35, pp. 56–68, Jan. 2018, doi: 10.1016/j.aei.2018.01.001.
[18] Y. Pi, N. D. Nath, and A. H. Behzadan, “Convolutional neural networks for object detection in aerial imagery
for disaster response and recovery,” Advanced Engineering Informatics, vol. 43, p. 101009, Jan. 2020, doi:
10.1016/j.aei.2019.101009.
[19] W. Dai, A. Mujeeb, M. Erdt, and A. Sourin, “Soldering defect detection in automatic optical inspection,” Advanced
Engineering Informatics, vol. 43, p. 101004, Jan. 2020, doi: 10.1016/j.aei.2019.101004.
[20] H. Bello-Salau, A. M. Aibinu, A. J. Onumanyi, E. N. Onwuka, J. J. Dukiya, and H. Ohize, “New road anomaly
detection and characterization algorithm for autonomous vehicles,” Applied Computing and Informatics, vol. 16, no.
1/2, pp. 223–239, May 2018, doi: 10.1016/j.aci.2018.05.002.
[21] P. M. Harikrishnan and V. P. Gopi, “Vehicle vibration signal processing for road surface monitoring,” IEEE Sensors
Journal, vol. 17, no. 16, pp. 5192–5197, Aug. 2017, doi: 10.1109/JSEN.2017.2719865.
[22] T. Nomura and Y. Shiraishi, “A method for estimating road surface conditions with a smartphone,” International
Journal of Informatics Society, vol. 7, no. 1, pp. 29–36, 2015.
[23] R. Madli, S. Hebbar, P. Pattar, and V. Golla, “Automatic detection and notification of potholes and humps on roads to
aid drivers,” IEEE Sensors Journal, vol. 15, no. 8, pp. 4313–4318, Aug. 2015, doi: 10.1109/JSEN.2015.2417579.
[24] X. Yu and E. Salari, “Pavement pothole detection and severity measurement using laser imaging,” in 2011 IEEE
International Conference on Electro/Information Technology, May 2011, pp. 1–5. doi: 10.1109/EIT.2011.5978573.
[25] H. Wu et al., “Road pothole extraction and safety evaluation by integration of point cloud and images de-
rived from mobile mapping sensors,” Advanced Engineering Informatics, vol. 42, p. 100936, Oct. 2019, doi:
10.1016/j.aei.2019.100936.
[26] M. Gavilán et al., “Adaptive road crack detection system by pavement classification,” Sensors, vol. 11, no. 10, pp.
9628–9657, Oct. 2011, doi: 10.3390/s111009628.
[27] T. S. Nguyen, M. Avila, and S. Begot, “Automatic detection and classification of defect on road pavement using
anisotropy measure,” in 2009 17th European Signal Processing Conference, Aug. 2009, pp. 617–621.
[28] M. H. Yousaf, K. Azhar, F. Murtaza, and F. Hussain, “Visual analysis of asphalt pavement for detec-
tion and localization of potholes,” Advanced Engineering Informatics, vol. 38, pp. 527–537, Oct. 2018, doi:
10.1016/j.aei.2018.09.002.
[29] S. C. Radopoulou and I. Brilakis, “Automated detection of multiple pavement defects,” Journal of Computing in Civil
Engineering, vol. 31, no. 2, p. 4016057, Mar. 2017, doi: 10.1061/(asce)cp.1943-5487.0000623..
[30] D. Kassem and C. Arce-Lopera, “Road-condition monitoring and classification for smart cities,” in Advances in Arti-
ficial Intelligence, Software and Systems Engineering. AHFE 2020. Advances in Intelligent Systems and Computing,
Jul. 2020, pp. 437–441. doi: 10.1007/978-3-030-51328-3 60.
[31] S. Li et al., “Detection of concealed cracks from ground penetrating radar images based on deep learning algorithm,”
Construction and Building Materials, vol. 273, p. 121949, Mar. 2021, doi: 10.1016/j.conbuildmat.2020.121949.
[32] Y. Zhang, J. Huang, and F. Cai, “On bridge surface crack detection based on an improved YOLO v3 algorithm,”
IFAC-PapersOnLine, vol. 53, no. 2, pp. 8205–8210, Jan. 2020, doi: 10.1016/j.ifacol.2020.12.1994.
[33] Y. Du, N. Pan, Z. Xu, F. Deng, Y. Shen, and H. Kang, “Pavement distress detection and classification based on
YOLO network,” International Journal of Pavement Engineering, vol. 22, no. 13, pp. 1659–1672, Nov 2021, doi:
10.1080/10298436.2020.1714047.
[34] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: unified, real-time object detection,” in
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2016, vol.

An image-based convolutional neural network system for road defects detection (Mohamed Anis Benallal)
584 r ISSN: 2252-8938

2016-Decem, pp. 779–788, doi: 10.1109/CVPR.2016.91.

[35] M.-T. Cao, Q.-V. Tran, N.-M. Nguyen, and K.-T. Chang, “Survey on performance of deep learning models for detect-
ing road damages using multiple dashcam image resources,” Advanced Engineering Informatics, vol. 46, p. 101182,
Oct. 2020, doi: 10.1016/j.aei.2020.101182.
[36] B. T. Passos, M. J. Cassaniga, A. M. da R. Fernandes, K. B. Medeiros, and E. Comunello, “Cracks and potholes in
road images,” Mendeley data, vol. 4, Jul. 2020, doi: 10.17632/t576ydh9v8.4.
[37] Research & Development Division, “Guidance notes - catalogue of road defects (CORD),” Research Development
division of the Highway Department, no. January, pp. 1-53, 2013.
[38] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: optimal speed and accuracy of object detection,” arXiv
preprint arXiv:2004.10934, 2020, doi: 10.48550/ARXIV.2004.10934.

BIOGRAPHIES OF AUTHORS

Mohamed Anis Benallal is a lecturer at the Department of Computer Science at the

University of Sciences and Technology of Oran Mohamed Boudiaf (USTO-MB), Algeria. He re-
ceived his PhD in Computer Science at the University of Perpignan Via Domitia, France in 2016.
His research interests are related to machine learning and artificial intelligence, remote sensing, data
mining, estimation and interpolation of parameters, natural language processing, environment, and
internet of things. He has published research papers international journals, co-author of conference
proceedings as well as a chapter of book edited in Springer International Publishing. Besides, he is
also one of the founder and a trainer in a non-profit foundation for student transversal trainings. He
can be contacted at email: mohamed.benallal@univ-usto.dz.

Mustapha Si Tayeb is a lecturer at the Department of Computer Science at the University

of Sciences and Technology of Oran Mohamed Boudiaf (USTO-MB), Algeria. He received his PhD
in Computer Science at the same university in 2018. His research interests include issues related
to artificial intelligence, remote sensing, image processing, data collection, analysis and classifica-
tion, internet of things, and pattern extraction. He is author of a set of research studies published
at national and international conference, and international journals. He can be contacted at email:
mustapha.sitayeb@univ-usto.dz.

Int J Artif Intell, Vol. 12, No. 2, June 2023 : 577 – 584