See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338793563

Development of New Anpr Dataset for Automatic Number Plate Detection and
Recognition in North of Iraq

Conference Paper · November 2019

DOI: 10.1109/UBMYK48245.2019.8965512

CITATIONS READS

7 259

3 authors, including:

Naaman Omar Abdulkadir Sengur

Duhok Polytechnic University Firat University
24 PUBLICATIONS 101 CITATIONS 228 PUBLICATIONS 5,667 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Skeleton based efficient fall detection View project

Hermite transform based color texture image segmentation View project

All content following this page was uploaded by Naaman Omar on 31 May 2021.

The user has requested enhancement of the downloaded file.

 Development of New Anpr Dataset for Automatic
Number Plate Detection and Recognition in North of
Iraq
Naaman Omar Yaseen Salim Ganim Saeed Al-Ali Abdulkadir Sengur
Department of Information Technology Department of Information Technology Department of Electrical-Electronics
Duhok Polytechnic University Management Engineering
Duhok, Iraq Duhok Polytechnic University Firat University
[email protected] Duhok, Iraq Elazig, Turkey
[email protected] [email protected]

Abstract- An automatic number plate detection (ANPD) country, in order to investigate and solve the issue of ANPD and
and automatic number plate recognition (ANPR) systems are ANPR [3-7]. This dataset shall be used for three important
robust technologies that are used for detecting and recognizing the purposes. First, it can be used for only vehicle number plate
number plates of vehicles. In this paper, a new dataset, which is
called North Iraq-Vehicle Images (NI-VI) of three provinces
detection. Second, it can be used for only recognition the
(Duhok, Erbil, and Sulaimani) for vehicle images, is presented. number plate, since There are labeled number plates in the
There are 1500 images in this dataset. They were gathered from dataset. Third, it can be used for both detection and recognition
real-time by using handled cameras to form a realistic dataset of together, which means it can be used for ANPR systems.
the vehicle images. The main contribution of this work is the
creation of a new dataset for license plate of vehicles in north Iraq In this paper, initially various license plate image datasets
with Arabic fonts in different and difficult conditions. The dataset are reviewed [8-13]. More specifically, all these datasets are
includes three categories of images: rotated, scaled and translated used for ANPR systems. These datasets contain images under
images. The resolutions of images are 4288 x 2848 and 5184 x 3456. different circumstances of weather conditions. Usually, these
Moreover, some images created for bad weather conditions, such images are related to plate numbers that belong to one country.
as snowy, dusty and low lighting. Some dirty plate images also At the same time, in the other hand, some datasets are employed
considered in the dataset. The purpose of introducing this dataset for automatic number plate detection (ANPD) systems [14-16].
is to provide and produce a realistic dataset for ANPD and as well All these datasets are used only for detection purpose (not
as for ANPR systems. recognition purpose). Moreover, there are some dataset that can
Keywords— Automatic number plate recognition (ANPR) be used for both detection and recognition the number plates in
datasets, Automatic number plate detection (ANPD) datasets, Plate the images. The system of these kinds are called ANPR systems
number recognition datasets, and License number detection [17-22].
datasets. The rest of this paper is organized as follows, second section
is about the datasets of ANPR literature reviews, in which set of
I. INTRODUCTION
datasets for vehicle images are reviewed and investigated all its
In the last decade, traffic control and traffic violation have properties. In the third section, is explained the ground truth of
become one of the most important issues in all countries. The our proposed NI-VI dataset. Moreover, the fourth section is for
flow of the traffic is controlled by many cameras that are spread comparison and discussion of these datasets. Finally, the
around the streets inside and outside of cities [1,2]. As these conclusion of this paper is presented in the last fifth section.
cameras acquire instantaneous image of the vehicles, that are on II. DATASETS OF ANPR LITERATURE REVIEW
the roads, intelligent software are needed to detect vehicles as
well as their license plates. Generally, software needs training There are many datasets created for ANPR systems around
the world. Each of these is concerned to plate numbers that
datasets for recognition of the license plates. Although of using
belong to one country, and usually it’s the researcher’s country.
these cameras, it is still difficult and challengeable to identify In other words, there is some privacy, for each dataset,
owner of the vehicle, who has violate the traffic, at the same connected to the traffic rules of the country about the number
moments using real-time application. It is very necessary to plate features, such as, size, color, font, etc. In this section,
have such real-time dataset that will used to identify the vehicle different datasets of ANPR system are reviewed with more
owner, who has violate the traffic by driving with high speed details about their images and features are presented.
more than the scheduled speed. Therefore, there is an urgent
need to create a dataset for license plate of vehicles in each

978-1-7281-3992-0/19/$31.00 ©2019 IEEE

A. The UFPR-ALPR Dataset [8] B. The GTI Dataset [9]
First, the UFPR-ALPR dataset is created for Brazil. It The GTI dataset was created in highways of Spain (Madrid,
contains images that were fully annotated. These images were Brussels and Turin). The images of this dataset were taken from
taken for 150 vehicles in real-world scenarios, and stored in 150 a video sequences by a camera fixed on the front of the vehicle.
videos. All these videos with frame rate of 30 frames/second. In one hand, there are 3425 positive rear images, which contains
All images, in these videos, were captured from moving (none the number plates within the vehicles, these images are extracted
stationary) vehicle, as well as were taken for moving vehicles, from variety locations of view. In the other hand, there are 3900
which means more realistic, during usual traffic in an urban negative images, which contains the number plates without any
environment. Be informed that this dataset consist of variety of vehicles’ images), these images are taken from way sequences.
vehicle images (cars, motorcycles, and trucks), it consists of There is a small number of images are used from Caltech and
4500 images that have different conditions, such as TU Graz-02 datasets, in order to approximate the total number
backgrounds, lighting, plate positions, plate quality, vehicle of images into 4000 positive and 4000 negative images. Some
types, and with various limitations for plates. Each license of positive images contain vehicle rear completely, while others
number plate has only 30 frame images. Finally, there are 1500 include about half of vehicle rear.
images are acquired by each of three different cameras: Huawei
P9 Lite, iPhone 7 Plus, and GoPro Hero4 Silver. The PNG
format is used to save images, with size of 1920×1080 pixels.
The images of each camera are split into 3 parts: 900 of vehicle
with gray license plates, 300 of vehicle with red plates, and 300
of motorcycles with gray ones. Examples about some images in
the GTI dataset are shown in the Fig. 2. Fig. 2. Image examples of GTI Dataset, for different range of views

The GTI dataset is also proposed the percentages of 2000 of

The images have resolution of 360x256 pixels sequences negative and positive images of different region of positions,
with resolution of 64x64 pixels. This dataset is included images 400 images are taken for each of these weather conditions:
that captured from different point of views. Based on distance, sunny, cloudy, and medium (neither very sunny nor cloudy).
the images are divided into two groups: middle, and far distance Another 400 images are taken for poor illumination such as
ranges. The middle range is also divided, based on the view of down or dusk weather condition. For light raining weather
the vehicle, into three subgroups: left, center, and right. As a condition, 200 image are used. The researchers include 100
result of these division, four independent regions are created, images for bad resolution camera pictures. There are 50 images
each region involves 1,000 images of a certain view, the Fig. 2, for industrial light, which are taken in the tunnels, are involved.
below, shows some image examples of GTI dataset. The dataset Finally, there is 50 images out of the total 2000 images, the
is proposed different situations such as weather condition, and researchers did not mention about their weather conditions.
lighting.
C. The Markus Weber dataset [10]
Markus Weber Cars dataset are taken by Markus Weber in
California Institute of Technology's parking. It is not a very wide
dataset since it includes only 126 images of resolution 896x592
pixels and all images are saved in JPG format. This dataset
involves only images that taken from rear, and only for salon
vehicles without including any track and bus vehicles. All
images of this dataset acquired under the same conditions, which
Fig. 1. Image examples of UFPR-ALPR Dataset is only sunny days. Moreover, the dataset is not including any
images that captured at night, low lighting, rain and shadow
The Brazilian license number plates are diverse in (size and weathers. In addition to all these obstacles, there is no any tilt or
color) based on the kind of vehicles and its category. Vehicle’ rotation and clear translation in the dataset. The Fig.3. shows
LPs have 40cm _ 13cm of size, But motorcycles LPs have the some image examples of Markus dataset.
size of 20cm _ 17cm. Also the color of cars LPs are variety
according to vehicle type, for examples the LPs for private
vehicles have gray color, while transportation vehicles, buses
and taxis have red LPs. Also for other types of cars like older
and official there are other LPs color used.
The images in the dataset divided in to three groups: 40% of Fig. 3. Image examples of Markus Weber Dataset
the images for learning, 40% of the images for training and 20%
of the images for validation. Each images has some explanatory D. The Baza-Slika Dataset [11]
notes in a file text, like by which camera the image taken, the The Baza-Slika is a dataset of vehicle images created using
position on the vehicle and types cars or motorcycle, due to the Olympus Camedia C-2040Zoom digital camera. In this
dataset contains both of them, also manufacturer, model and database, there are more than 500 images of the resolution used
year, position of the LP and the position of its characters can 640×460 for only rear views of cars. It includes three categories
annotated by the image. of vehicles: cars, trucks, and busses.
 The images in this dataset are acquired from all over Croatia
country through variety of lighting and weather conditions, for
example there are seven folders of sunny, cloudy, sunshine,
rainy, twilight and night light weather images. As well, the
images are taken in diverse time of day, such as morning, after
noon, evening and night with variant of qualities, brightness or
Fig. 6. Image examples of AOLP Dataset
contrasts, as shown as in the Fig. 4.
The images, in the AOLP dataset, are taken in different
lighting conditions, such at night time, day time, outdoor and
indoor. Also, the illumination of these images covers indoor,
outdoor, daytime, nighttime, and various weather conditions.
III. GROUND TRUTH OF NI-VI DATASET
Fig. 4. Image examples of Baza-Slika Dataset This section provides details and information related to
vehicle images taken in different conditions, which are found in
E. The SLVDS-iLPR Dataset [12] the present dataset. The aim is to improve ground truth of this
This vehicle image dataset is created for Stop-Line Violation dataset. Our new dataset is used for all different vehicle images
Detection System (SLVDS) dataset and is achieved for Indian
of North of Iraq, so it is called NI-VI. This dataset with Arabic
Traffic Management system (ITMS). The images are collected
using surveillance cameras of traffic monitoring of most metro font (text and numbers) has been created to help researchers to
cities in India. apply their methods in automatic number plate detection and
The dataset consists of 4717 vehicle images, each image recognition systems to increase performance of these systems.
with resolution of 704x576 pixels. These images are taken from Images in the proposed dataset are taken from real time by using
more than 30,000 snapshots of SLVDS. Some of these images two handled (unfixed) digital cameras from variety positions
are captured during the days, while others in the night and for and angles. It involves images captured by using Canon 60D,
different seasons of the year, which mean it covered different EFS 18-55mm and Nikon DX, AF-S NIKKOR 18-105mm
weather condition in India. Some examples about the vehicles’ cameras of resolution 4288 x 2848 and 5184 x 3456
images of this dataset is shown below in the Fig. 5.
respectively.
Moreover, the (NI-VI) dataset comprises 1500 images taken
from real time in different condition and variety weather
situations such as day and night lighting with various
backgrounds such as sunny, cloudy, snowy, foggy, dusty, and
inside and outside cities. Furthermore, even some images of
Fig. 5. Image examples of SLVDS-ITMS Dataset
vehicles of dirty number plates are included in the dataset.
F. The AOLP Dataset [13] Moreover, the images are captured in different times and places.
The Application Oriented license Plate (AOLP) database has Some of these images are taken under low or extra light source,
been created by in the Artificial Vision lab, NTUST. The images other are taken under weak or strong sunlight. Our NI-VI dataset
in this dataset is classified into three categorization groups: also involves images of different types of vehicles like Trucks,
Access Control (AC), traffic Law Enforcement (LE), and Road
Buses and Salon with different colors for foreground and
Patrol (RP).
background colors. This differences is due that plate numbers
The total number of images in this dataset include 2049 are different. Some samples of images in NI-VI dataset are
images with resolution of 320 x 240. In the first AC group, there
shown below in in Fig. 7.
are 681 images of moving, parking (stop), steady passing
conditions. The distance between camera and number plate is
equal 5m, the plate width is between 0.2 and 0.25 compare with
the image’s width. The second LE group has 757 images, which
are taken by camera stand on road side. The images are for
vehicles that violate traffic laws. The last RP group involves 611
images that captured from different points and variant distances
using camera, which is handheld on a moving vehicle. The Fig.
6, below, shows some examples about the vehicles’ images of
this AOLP dataset.

Fig. 7. Image examples of NI-VI Dataset

 In order to cover all aspects for 2D transformation in names of whole images in our dataset. Some samples about these
computer graphics, translation, scaling, and rotation are used in cropped images is shown below in Fig. 9.
all different cases of the images in the dataset. Therefore, the
images, within this dataset, are divided into three categories
rotation, scale and translation. The fig. 7. Shows the three
categories of all images in the NI-VI dataset. In the rotation
category, the images are taken in both left and right direction.
The angle slope between cars and cameras, inside the images, is
±20 . Totally, the rotation includes 400 images in four folders,
such as 100 images for left angle with near distance, 100 images
for left angle with far distance, 100 images for right angle with
near distance and the last 100 images for right angle with far Fig. 9. Cropped image examples of NI-VI Dataset
distance. For the scaling category, it involves 300 images based
IV. COMPARISON AND DISCUSSION
on the distance between the captured cars and the cameras, such
The researchers, of the ANPD and ANPR approaches, are
as 100 images are taken from near distance, 100 images from
more likely interested to work with the number plates of their
mid and 100 images from far distance. In the third category for countries. Therefore, it is very urgent to design the system in
translation, the images are divided into two sub-categories: global matter to detect and recognize any number plate
corners and sides. In one hand, 400 images are captured where regardless of its belonging country, since the cars can move
the number plates are translated into the corners of the images, among different countries. Thus, a dataset is created for north of
Iraq for vehicles plates with Arabic text and numbers because
such as 100 images for left down corner, 100 images for left top, there is no any dataset has been created for such kind of Arabic
100 images for right down and 100 images for right corner. In fonts.
the other hand, 400 images are captured where the number plates
A comparison among some reviewed datasets is achieved in
are translated into the sides of the images, such as100 images for this paper, as shown as in the Table I. This table compares these
left side, 100 images for right, 100 images for top and the last datasets based on some factors and attributes, such as country,
100 image for the bottom side. All categories and sub-categories year, and total number of images, resolution, and other
of images inside our NI-VI dataset is shown below in the Fig. 8. conditions. Some of these datasets are old with bad resolutions,
which there is a big lack of details inside the images. Nowadays,
with the improvement happened in the cameras technologies
would help to provide high resolution images but with more
storage capacities. Furthermore, the weather conditions are not
fully covered in some of these datasets, which mean that the
dataset is far away from being as a realistic example of the real
life. All the datasets that showed in the comparison table are
representing the researchers’ countries and as shown there is no
any country with Arabic fonts. Thus, our suggested dataset is
covering the lack of Arabic text and numbers in the vehicle
images.
V. CONCLUSION AND FUTURE WORKS
In this paper, a new dataset (NI-VI) of north of Iraq vehicle
image was systematically presented. All 1500 images in this
dataset was gathered from real time by using handled cameras.
The dataset includes three categories of images: rotation, scale
and translation of resolution 4288 x 2848 and 5184 x 3456. The
purpose of introducing dataset is to provide data for testing
ANPD and ANPR algorithms by researchers to increase
methods performance.
Fig. 8. Block diagram of NI-VI Dataset. The NI-VI dataset was systematically presented in this
paper. The importance of this dataset, it is required for ANPD
Furthermore, the presented dataset also involved 1500 and ANPR approaches and especially for north of Iraq. This
cropped plate numbers from each whole image manually. Each dataset would help and assist the researches in these approaches.
of these images contains three main parts: (1) Top part involves All works in the ANPD and ANPR approaches, in north of Iraq,
license numbers. (2) The left of bottom part include provinces should to be tested and experimented by some dataset that
name. (3) The right of bottom part is used for country name. The contains standardized images of vehicles and related to north
names of these cropped license plate numbers are matched with
Iraq location. The limitation in this work is that this dataset Dataset Attributes
related to vehicle license plates of only north Iraq, but it can be Name Country Yea Images Res. Conditions
improved easily in the future work to include whole Iraq r
Different
country. conditions,
night light,
TABLE I. COMPARISON OF DATASETS FOR NUMBER PLATE OF VEHICLES day, sunny,
Dataset Attributes snow, fogy,
Name Country Yea Images Res. Conditions dirty,
r 4288 shadow,
Different x cloudy,
backgrounds, Our 2848 rainy, front,
North of
192 lighting dataset 2019 1500 and rear,
Iraq
0 conditions, (NI-VI) 5184 rotation,
UFPR- x scale,
Brazil 2018 4500 x rear number
ALPR 3456 translation,
108 plate
0 positions, unfixed
and cars distance,
type. two-handle
Day camera,
morning, different
evening, angles.
night, sunny,
rainy,
704
SLVDS cloudy, fog,
India 2014 4717 x
-iLPR shadow, low
576
illumination,
blurriness, REFERENCES
various tilt
angles and [1] K. Raghunandan et al., “Riesz fractional based model for enhancing
distances. license plate detection and recognition,” IEEE Trans. Circuits Syst.
Different Video Technol., vol. 28, no. 9, pp. 2276–2288, 2017.
lighting, day, [2] L. Hu and Q. Ni, “IoT-driven automated object detection algorithm for
320
night, indoor urban surveillance systems in smart cities,” IEEE Internet Things J., vol.
AOLP Taiwan 2013 2049 x
and outdoor 5, no. 2, pp. 747–754, 2017.
240
illuminations [3] N. Saleem, H. Muazzam, H. Tahir, and U. Farooq, “Automatic license
. plate recognition using extracted features,” presented at the 2016 4th
Video International Symposium on Computational and Business Intelligence
sequences, (ISCBI), 2016, pp. 221–225.
sunny,
360 [4] S. Du, M. Ibrahim, M. Shehata, and W. Badawy, “Automatic license
cloudy, poor
GTI Spain 2012 3425 x plate recognition (ALPR): A state-of-the-art review,” IEEE Trans.
illumination,
256 Circuits Syst. Video Technol., vol. 23, no. 2, pp. 311–325, 2012.
light rain,
artificial [5] S. Du, M. Ibrahim, M. Shehata, and W. Badawy, “Automatic license
lights. plate recognition (ALPR): A state-of-the-art review,” IEEE Trans.
Parking cars, Circuits Syst. Video Technol., vol. 23, no. 2, pp. 311–325, 2012.
Marku 896 sunny days, [6] M. S. Al-Shemarry, Y. Li, and S. Abdulla, “An efficient texture
s USA 2003 126 x and rear descriptor for the detection of license plates from vehicle images in
Weber 592 center of difficult conditions,” IEEE Trans. Intell. Transp. Syst., 2019.
small cars. [7] O. I. Al-Sanjary, A. A. Ahmed, and G. Sulong, “Development of a video
Sunny, tampering dataset for forensic investigation,” Forensic Sci. Int., vol. 266,
cloudy, pp. 565–572, 2016.
sunshine,
[8] R. Laroca et al., “A robust real-time automatic license plate recognition
rainy,
640 based on the YOLO detector,” presented at the 2018 International Joint
Baza- twilight and
Croatia 2001 500 x Conference on Neural Networks (IJCNN), 2018, pp. 1–10.
Slika night light,
460 [9] J. Arróspide, L. Salgado, and M. Nieto, “Video analysis-based vehicle
rear view left
and right detection and tracking using an MCMC sampling framework,”
rotation, near EURASIP J. Adv. Signal Process., vol. 2012, no. 1, p. 2, 2012.
scaling. [10] M. Oliveira and V. Santos, “Automatic detection of cars in real roads
using haar-like features,” Dep. Mech. Eng. Univ. Aveiro, vol. 3810,
2008.
[11] “Projekt ‘License Plates.’” [Online]. Available:
http://www.zemris.fer.hr/projects/LicensePlates/english/. [Accessed:
17-Jul-2019].
[12] S. Saha, S. Basu, and M. Nasipuri, “iLPR: an Indian license plate
recognition system,” Multimed. Tools Appl., vol. 74, no. 23, pp. 10621–
10656, 2015.
[13] G.-S. Hsu, J.-C. Chen, and Y.-Z. Chung, “Application-oriented license [19] S. M. Silva and C. R. Jung, “Real-time brazilian license plate detection
plate recognition,” IEEE Trans. Veh. Technol., vol. 62, no. 2, pp. 552– and recognition using deep convolutional neural networks,” presented at
561, 2012. the 2017 30th SIBGRAPI Conference on Graphics, Patterns and Images
[14] L. Xie, T. Ahmad, L. Jin, Y. Liu, and S. Zhang, “A new CNN-based (SIBGRAPI), 2017, pp. 55–62.
method for multi-directional car license plate detection,” IEEE Trans. [20] S. Kaur and S. Kaur, “An efficient approach for number plate extraction
Intell. Transp. Syst., vol. 19, no. 2, pp. 507–517, 2018. from vehicles image under image processing,” Int. J. Comput. Sci. Inf.
[15] S. He, Y. Yuan, C. Fu, X. Hu, and Y. Zhao, “Robust license plate Technol., vol. 5, no. 3, pp. 2954–2959, 2014.
detection using profile-based filter,” presented at the 2018 Tenth [21] M. R. Asif, Q. Chun, S. Hussain, M. S. Fareed, and S. Khan,
International Conference on Advanced Computational Intelligence “Multinational vehicle license plate detection in complex backgrounds,”
(ICACI), 2018, pp. 794–800. J. Vis. Commun. Image Represent., vol. 46, pp. 176–186, 2017.
[16] S. Yu, B. Li, Q. Zhang, C. Liu, and M. Q.-H. Meng, “A novel license [22] N. More and B. Tidke, “License plate identification using artificial
plate location method based on wavelet transform and EMD analysis,” neural network and wavelet transformed feature selection,” presented at
Pattern Recognit., vol. 48, no. 1, pp. 114–125, 2015. the 2015 International Conference on Pervasive Computing (ICPC),
[17] M. S. Al-Shemarry, Y. Li, and S. Abdulla, “Ensemble of adaboost 2015, pp. 1–5.
cascades of 3L-LBPs classifiers for license plates detection with low
quality images,” Expert Syst. Appl., vol. 92, pp. 216–235, 2018.
[18] K. Deb, H.-U. Chae, and K.-H. Jo, “Vehicle License Plate Detection
Method Based on Sliding Concentric Windows and Histogram.,” JCP,
vol. 4, no. 8, pp. 771–777, 2009.

View publication stats