Content uploaded by Sheikh Nooruddin
Author content
All content in this area was uploaded by Sheikh Nooruddin on Apr 07, 2021
Content may be subject to copyright.
2020 23rd International Conference on Computer and Information Technology (ICCIT), 19-21 December, 2020
978-0-7381-2333-2/20/$31.00 ©2020 IEEE
A Bangladeshi License Plate Detection System
Based on Extracted Color Features
Sheikh Nooruddin, Falguni Ahmed Sharna, Sk. Md. Masudul Ahsan
Department of Computer Science and Engineering
Khulna University of Engineering & Technology, Khulna-9203, Bangladesh
nooruddinimad@gmail.com, falguniahmed114@gmail.com, and smahsan@cse.kuet.ac.bd
Abstract— As the number of motorized vehicles is increasing
rapidly in Bangladesh, Automatic License Plate Detection and
Recognition (ALPDR) systems have become a necessity for
proper management of vehicles on roads. The first phase of an
ALPDR system is the detection and localization of number
plates from vehicle images. In this paper, we introduce a dataset
of 630 images that were manually captured. The dataset
represents various real-world scenarios. We propose the use of
color histograms with MinPool and MaxPool features for license
plate detection and localization. The detection system was tested
in multiple color spaces to observe their effect on the detection
phase. The proposed and developed system is very effective and
achieved high levels of correctness in the detection phase
according to different metrics.
Keywords— automatic license plate detection; image
processing, color histograms; horizontal and vertical histograms,
random forests.
I. INTRODUCTION
Automatic License Plate Detection and Recognition
(ALPDR) is the process of detecting a license plate (LP)
region from an image, localizing and cropping the license
plate, and recognizing the contents of the license plate [1].
Due to an increase in both commercial and private vehicles,
automated systems such as automatic traffic control systems,
ALPDR systems have become a necessity for proper
management of vehicles in the road [2]. ALPDR systems
enable law enforcement agencies to properly monitor the
vehicles in the road and swiftly take actions against
responsible parties in case of an accident. Applications of
ALPDR systems include automated parking systems, access
control systems, automated tolling systems, automated traffic
control systems, border control systems, etc. [3]–[5].
According to the records from the Bangladesh Road
Transport Authority (BRTA), a total of 34,19,884 registered
motor vehicles exist in Bangladesh. Among these, there are
around 1,35,081 trucks, 3,35,660 private cars, and 44,374
buses [6]. There is also a huge amount of non-registered
vehicles that are unaccounted for. The sheer number of
vehicles on the roads in Bangladesh requires automated
management systems for proper management, however, very
few such systems exist. Traffic jams, road accidents, hit and
run incidents, etc. are among the numerous effects of a large
number of vehicles and manual control systems [6].
Various notable methods have already been explored for
number plate detection and recognition, such as deep learning
based methods [7]–[9], image processing based methods [10],
[11], fuzzy logic based methods [12], etc. Various ROI
descriptor features such as Histogram of Gradient Orients
(HOG), Features from Binary Robust Independent
Elementary Features (BRIEF), Accelerated Segment Test
(FAST) combined with machine learning techniques have also
been successfully used for license plate detection. Various
automated ALPDR systems have been developed in the
context of Bangladesh [13]–[16]. Deep learning based fall
detection systems do not require implicit feature extraction
techniques. The deep learning models perform feature
extraction, detection, and localization tasks internally. Deep
learning based detection systems are very robust, as larger
models can have millions of parameters and can extract
sophisticated features from images. Image processing based
systems mostly depend on various edge detection and
morphological operations for license plate area detection and
recognition.
Datasets are the most crucial part of any computer vision
system and are mandatory for future incremental
improvements. In this paper, we present the first-ever open-
access Bangladeshi vehicle license plates dataset. We also
propose a novel combination of features namely color
histogram with MinPool and MaxPool for license plate
detection and localization in the context of Bangladesh.
The rest of the paper is organized as follows: Section II
provides information on Bangladeshi license plate regulations
set by BRTA. Several related research works have been
briefly discussed in Section III. Section IV presents the
methodology of the detection system. The implementation
details and experimental results are analyzed in Section V.
Finally, Section VI concludes the paper.
II. LICENSE PLATES CONVENTION IN BANGLADESH
The registration numbers and the license plates that are
used in vehicles are generally issued by BRTA in Bangladesh.
Fig. 1 represents a typical Bangladeshi license plate.
Bangladeshi license plates generally use Bangla numerals and
characters to form the registration numbers. Currently, 33
letters and 10 numerals of the Bangla language are permitted
for use in the registration plate. The permitted Bangla
characters and numerals can be used to represent all valid
license plate combos. The use of English characters and
numerals are not permitted.
Bangladeshi license plates are generally printed on metal
sheets. Most commercial vehicles such as buses and trucks are
required to present a metal-sheet license plate with green
background and black text for ease of detection and recognition.
Fig. 1. A typical Bangladeshi license plate.
Fig. 2. The input image and extracted features.
The license plates must contain two lines. The standard format
is “City Name -Vehicle Class Letter” in the upper line and
“Vehicle Class Number – Registration Number” in the lower
line. In the illustrated case “গাজীপুর” (Gazipur) is the city
name. “জ” (Ja) is the vehicle class letter. “০৪” (04) is the
vehicle class number and “০০২৬” (0026) is the registration
number. According to BRTA, these have to be written in
Bengali numerals and character, printed on a flat metal plate,
and the total plate dimension must be 524 mm 112 mm. The
plate must be put both in front and in the back of the vehicles.
III. LITERATURE REVIEW
This section provides a descriptive summary of existing
relevant ALPDR systems in the context of Bangladesh.
Siddique et al. [13] used a modified Sobel Edge Detector
for horizontal and vertical edge detection and a size-specific
search based on Bangladesh Road Transport Authority
(BRTA) standards for license plate localization.
Morphological operations and neighborhood connectivity
were used for character segmentation. Horizontal line (মাা),
vertical line, closed loops and upper extension properties were
utilized for character recognition and perspectives and
curvature properties were used for numeral recognition.
However, this system requires a consistent connection to the
national database for license plate approximation. The system
is also less robust as it performs a size-specific search based
on BRTA standards for license plate detection.
Uddin et al. [14] used Sobel edge detection, vertical
histograms, and morphological closing to detect and localize
the license plate region. After performing adaptive
thresholding on the localized license plate region, after noise
removal, connected component analysis and aspect ratio of
bounding boxes were used to segment the characters. After
segmentation, a two-dimensional Gabor filter was used for
feature extraction. The dimensionality of the extracted
features was reduced using the Kernel PCA method and two
separate SVM models for numbers and alphabets were trained
using the features.
A YOLOv3 deep learning model for license plate
detection and ResNet-20-based deep Convolutional Neural
Network (CNN) for character recognition was used by
Abdullah et al. [15]. The system is able to detect the license
plates and recognize the characters in very complex scenarios
and angles as generally state-of-the-art models were used.
The color properties of license plates, namely standard
deviation and mean values of intensity were used by Deb et al.
[16] to detect probable license plate regions from vehicle
images. The images were represented in the HSI color model.
After the morphological closing operation of the probable
license plate regions, connected component analysis was
performed. Then geometric properties such as aspect ratio,
area, and intensity histograms are used to output the final
license plate. This reviewed work does not perform any
license plate recognition task.
IV. METHODOLOGY
In the license plate detection and localization stage,
preprocessing steps such as image rescaling and color space
conversions are performed. After preprocessing, a window of
fixed size is slide across the image, and features are extracted
from the sub-windows. The features are then used for training
a machine learning model. Later, in test images, the trained
model is used for determining whether a window is part of the
LP region or not. The probable windows are then merged to
generate probable license plate regions. The probable LP
regions are finally filtered based on their area and aspect ratio
to detect, localize, and output the license plate.
A. Preprocessing
1) Image scaling: Image resizing or image scaling refers
to changing the size of an image. We used the Bilinear
Interpolation method for scaling the images. All landscape
images in the dataset were resized to (height, width, channel)
= (480, 640, 3) and all portrait images were resized to (height,
width, channel) = (640, 480, 3) for ease of calculation.
Red Channel
Green Channel
Blue Channel
Merge
and
Flatten
Histogram in Red Channel
Histogram in Green Channel
Histogram in Blue Channel
Output of MinPool
Output of MaxPool
16 × 16
sliding
window
Histograms
Sub-windows in
Cropped window
Input Image
r0 r1 … rn
g0 g1 … gn
b0 b1 … bn
min0
…
minn
max0
…
maxn
r0 … rn g0 … gn B0 … bn min0
… minn
max0
… maxn
3 × 1
sliding
window
2) Color space conversions: A color space refers to the
specific organization of colors. A color model is an abstract
mathematical model that describes the representation of
colors as tuples of numbers. We used three color models
namely RGB, YCbCr, and LAB to research whether the color
models affect the performance of the feature extraction
method.
B. Feature Extraction
Color histograms represent the distribution of colors in an
image. Color histograms can generally be represented in a 1D
array whose length is equal to 2N, where N corresponds to the
bit depth. As color images in RGB color space contain three
channels: Red, Green, and Blue, it is possible to calculate a
total of three histograms (one histogram per channel). These
color histograms can then be used as feature vectors.
Pooling is a general technique for reducing
dimensionality, computational complexity, and variance. In
the sliding window approach, in the case of MinPool or
minimum pooling, the minimum pixel value among all the
pixel intensities in the sliding window crop is selected.
MinPool can be used to highlight the darker regions in a
window. In the case of MaxPool or maximum pooling, the
maximum pixel intensity value in a sub-window is selected.
MaxPooling can be used to highlight the brighter regions in
images. Eq. 1 and Eq. 2 represents the equations for
minpooling and maxpooling.
In theory, MinPool and MaxPool are essentially identical
to minimum filtering and maximum filtering. We calculated 3
color histograms (1 for each channel) and used a filter of size
(height, width) = (3, 1) to calculate the MinPool and MaxPool
features from each sliding window. The extracted features
were then concatenated to generate the feature vector for the
color histogram with MinPool and MaxPool based feature
extraction method. While using a window size of 16 16, a
total of 984 features are generated. The histogram bin size was
set to 256. Fig. 2 represents the sliding window based feature
extraction process.
I'
u, v
← min
I
u+i, v+j
(
i,j
)
ϵ R}
I'
u, v
← max
I
u+i, v+j
(
i,j
)
ϵ R}
C. Model Training
We trained a Random Forest machine learning classifier
on the extracted features. Random forests are ensemble
learning methods that are generated by constructing numerous
decision trees each working with a random subsample of the
entire dataset. The final output is generated by taking the mode
of all the classes of the individual trees in case of
classification. Information gain was used to decide which
feature to use as a basis for splitting in the decision trees. The
best feature is the one that provides the most information gain.
Information gain is based on entropy and is calculated as
follows where H(T) represents the entropy. The equations of
information gain and entropy are presented in Eq. 3 and Eq. 4.
IG
(
T,a
)
=H
(
T
)
–H
(
T|a
)
H(T) = IE (p1, p2, …, pN)
Fig. 3. The detection and localization process.
Fig. 4. Sample images from our dataset.
Fig. 5. Performance metrics of the developed models in different color
spaces.
Random forests can overcome the overfitting tendency of
decision trees by limiting the maximum number of trees used
or the depth of the trees and by providing random subsamples
of the dataset to each tree.
D. Detection and Localization
The sliding window approach used in the feature
extraction method is used in the detection and localization
phase. The features are first extracted from the cropped
images from the windows, then standardized and sent to the
92
94
96
98
100
RGB LAB YCbCr
Percentage (%)
Color Models
Accuracy
Precision
Recall
F1 Score
13401366
23020321
03532675
36751271
61320362
35721071
1 3 4 0 1 3 6 6
2 3 0 2 0 3 2 1
0 3 5 3 2 6 7 5
3 6 7 5 1 2 7 1
6 1 3 2 0 3 6 2
3 5 7 2 1 0 7 1
13401366
23020321
03532675
36751271
61320362
35721071
Merging All
Detected
Areas
Filter Based
on Aspect
Ratio and
Area
Image with All
Detected Regions
Image with
Merged Regions
Final Output with
Localized License Plate
(a)
(b)
(c)
1 3 4 0 1 3 6 6
2 3 0 2 0 3 2 1
0 3 5 3 2 6 7 5
3 6 7 5 1 2 7 1
6 1 3 2 0 3 6 2
3 5 7 2 1 0 7 1
0 3 2 1
2 6 7 5
1 2 7 1
0 3 6 2
Input Image
Extracted
Features
Trained
Model
Classification
Draw a Box
if Area is LP
Region
13401366
23020321
03532675
36751271
61320362
35721071
Cropped
Sub-window Multiple LP Regions
X
1
, Y
1
X4, Y4
X
2
, Y
2
X3, Y3
min(X1, X3), min(Y1, Y3)
max(X
2
, X
4
), max(Y
2
, Y
4
)
Back Back-Left
Back-Right
Front
Front-Left
Front-Right
pre-trained random forest classifier for classification. If the
window is classified as part of a license plate, the co-ordinates
are recorded and a bounding box is drawn. After the entire
image has been classified, the overlapping bounding boxes are
merged to get the probable license plate region. The probable
license plate regions are then filtered based on area and aspect
ratio. The one with the largest area and aspect ratio within 1.5
to 2.5 is output as the license plate region.
Fig. 3 illustrates the detection and localization process.
Fig. 3(a) shows the candidate license plate region detection
process. The yellow box represents the sliding window. The
areas identified by the classifier as part of a probable LP
region is represented by Red bounding boxes. Fig. 3(b) shows
the final license plate selection process. Overlapping initial
bounding boxes are combined using the formula presented in
Fig. 3(c). If the coordinates of the top left corner and bottom
right corner are known for each of the two overlapping
bounding boxes, coordinates for a larger bounding box can be
calculated that contains the object detected by the initial small
bounding boxes. Probable LP regions are generated by
merging the red bounding boxes. Probable license plate
regions are presented by blue bounding boxes. After filtering
the probable LP regions based on the Aspect Ratio and area,
the final license plate region is selected. The final license plate
region is represented by the green bounding box.
V. EXPERIMENTAL RESULTS
The results and robustness of the automatic license plate
detection system are presented in this section. The system was
executed on a computer with 2.8GHz dual-core Intel Core i7
(Turbo Boost up to 3.3GHz) processor with 4MB shared L3
cache, and Intel Iris Graphics, and 16GB 1600MHz DDR3L
onboard memory running Ubuntu 14.04.5 LTS. Python was
used to develop the system and scikit-learn was used for the
machine learning models.
A. Dataset
We have created a custom dataset containing manually
captured images that represent various real-world scenarios.
The dataset contains a total of 630 images. The portrait and
landscape images had a resolution of 3120 × 4160 pixels and
4160 × 3120 pixels, respectively. As the original images have
very high resolution, the license plate regions can be cropped
to build a high-quality dataset for the recognition task.
However, as image processing based methods are
computationally extensive, the images were resized to 640 ×
480 pixels in case of landscape images. The portrait images
were resized to 480 × 640 pixels.
Blurring and depth effects were not used while capturing
the images. Thus, a lot of the images contain multiple objects
in the background. However, it was made sure that there is
only one license plate visible per image. All the images were
captured with express permission from the owners of the
vehicles.
A smartphone containing 13MP Sony Exmor RS sensor
with f/2.0 aperture, and pixel size of 1.12μm was used as the
capturing device. The captured images were divided into six
categories based on the position of the vehicle relative to the
camera. These categories are: Back, Back-Left, Back-Right,
Front, Front-Left, Front-Right. For example, in the case of
Front-Right, the front side of the vehicle is in the right side of
the camera. Fig. 4 presents some representative images from
each type in the dataset.
TABLE I. HYPER PARAMETERS FOR THE RANDOM FOREST
Parameter Name Value
bootstrap True
class_weight “balanced_subsample”
criterion “gini”
max_depth 50
max_features “sqrt”
min_samples_leaf 5
min_samples_split 12
n_estimators 100
TABLE II. PERFORMANCE MEASURES (IN %) OF DETECTION AND
LOCALIZATION
Color
Space
Performance
Metrics IoU = 30% IoU = 50%
RGB
Accuracy 89.9 80.1
Precision 98.0 97.9
Recall 91.5 81.6
F1 Score 94.6 88.6
LAB
Accuracy 87.8 76.2
Precision 96.5 95.6
Recall 90.3 78.2
F1 Score 93.2 85.7
YCbCr
Accuracy 89.3 76.5
Precision 95.7 94.9
Recall 92.7 79.3
F1 Score 94.2 86.3
TABLE III. CATEGORICAL PERFORMANCE METRICS IN RGB COLOR
SPACE
Category
Metrics (IoU = 50%)
Accuracy Precision Recall
Back 86.6 97.5 88.6
Back-Left 82.8 92.3 92.3
Back-Right 60.0 90.0 60.0
Front 83.3 97.6 87.0
Front-Left 91.3 100.0 90.9
Front-Right 77.7 100.0 76.4
TABLE IV. EXPERIMENTAL RESULTS ON PKU DATASET FOR LICENSE
PLATE DETECTION
Methods
Subset
G1 G2 G3 G4 G5 Avg.
Zhou et al. [17] 95.4 97.8 94.2 81.2 82.0 90.2
Li et al. [18] 98.8 98.4 95.8 81.1 83.3 91.5
Yuan et al. [19] 98.7 98.4 97.7 96.2 97.3 97.6
Li et al. [8] 99.8 99.8 99.8 100 99.3 99.8
Selmi et al. [20] 99.5 99.4 99.4 99.6 99.1 99.4
Our System 90.5 91.1 98.6 76.5 87.9 88.9
(a) (b) (c)
Fig. 6. Sample detection and localization process.
(a) (b) (c)
Fig. 7. Some missed detection and localization samples.
B. License plate detection and localization
We used a Random Forest classifier model with the
number of decision trees set to 100 and the maximum number
of splits for each tree set to 50. This optimal specification was
found after testing several combinations using the Grid Search
technique. The features were extracted from images in three
different color spaces: RGB, LAB, and YCbCr. Separate
random forest models were trained for each color space. The
performance metrics for the trained models are illustrated in
Fig. 5. In all the cases, hyperparameters were tuned to
optimize the model performance. The optimal hyper-
parameter names and their values are presented in Table I.
Almost all of the models achieved accuracy, precision, recall,
and f1 score of 99.7%, 99.6%, 99.9%, 99.7% on average. So,
it is evident that the used color space does not affect the
proposed feature extraction method for detection and
localization. It is also evident that the features of the two
classes are highly separable and unique.
The percentile localization result of the detection and
localization phase with respect to Intersection over Union
(IoU) is provided in Table II. It is evident that the proposed
feature extraction method for license plate detection and
localization performs similarly in different color spaces.
However, the best performance was achieved in the RGB
color space. When IoU was set to 30%, the accuracy,
precision, recall, and f1 score of the detection and localization
system is 89.9%, 98%, 91.5%, and 94.6%, respectively. When
IoU was set to 50%, the accuracy, precision, recall, and f1
score of the detection and localization system is 80.1%,
97.9%, 81.6%, and 88.6%, respectively. Fig. 6 illustrates a
complete detection and localization output from an example
image. The respective license plate region is zoomed and
presented inset in all the images. Fig. 6(a) represents the input
image. In Fig. 6(b), the windows that were detected as part of
a license plate region by the classifier are represented using
blue bounding boxes. The green bounding box was generated
by merging all the blue bounding boxes. As there was only one
probable region detected by the system, the region bounded by
the green bounding box is output as the license plate region.
In Fig. 6(c), the output from the system is presented using the
green bounding box while the ground truth is marked using the
red bounding box. The calculated IoU, in this case, is 0.93.
Some example images where the system failed to detect the
license plates are presented in Fig. 7. The images are from the
Back-Right, Front-Right, and Back-Left category,
respectively Lighting and the camera angles are one of the
most important concerns in an automatic license plate
detection system. In 7(a), due to the lighting angle and the
deep color of the license plate, the license plate region is not
easily discernable from the deep colored background. In 7(b)
and 7(c), the output of the system is presented in a red
bounding box and the ground truth is presented in a green
bounding box. The calculated IoU is presented at the top of
each respective image. In fig. 7(b), the system failed to detect
the license plate due to the sharp camera angle. However, it
was able to detect a very small portion of the license plate. In
Fig. 7(c), the system failed to detect the license plate due to
the discoloration of the license plate. The black coating of the
license plate characters has peeled off and thus there are
insufficient color features to identify potential license plate
regions. The calculated IoU in are 0.0, 0.04, and 0.19,
respectively. The percentile performance metrics on the 6
separate categories of the dataset in RGB color space are
presented in Table III. The performance was calculated with
IoU set to 50%. In all the cases, precision is consistently high,
meaning the system is highly capable of detecting true license
plates.
The performance and robustness of the detection and
localization systems were evaluated on the PKU (Peking
University) benchmark dataset [19]. The PKU dataset contains
a total of 3977 images of Chinese license plates. The photos
were captured under varying environmental and lighting
conditions such as daytime, nighttime, nighttime with
headlights on, daylight with sunshine glare, daytime with
reflective glare, etc. The dataset is divided into 5 subsets: G1-
G5. While the images in G1-G4 subsets contain one license
plate per image, the images in the G5 dataset contains multiple
license plates per image. The resolution of the images in the
G1-G3 is 1082 × 728. The images in the G4 and G5 subsets
have resolutions of 1600 × 1236 and 1600 × 1200. The images
were resized to 640 × 480 pixels before applying our detection
system. The comparison results of our method with other state
of the art with other existing state-of-the-art license plate
detection methods are presented in Table IV.
VI. CONCLUSION
Automatic License Plate Detection and Recognition
(ALPDR) systems are one of the applications of Computer
Vision that has revolutionized intelligent traffic control
systems. Many developed countries use ALPDR systems for a
multitude of purposes, from automatic centralized tolling, to
law enforcement. In this paper, we have presented the first-
ever open-access dataset containing Bangladeshi vehicles with
license plates to facilitate further research. We have also
proposed a novel feature extraction method for vehicle
number plate detection and localization in the context of
Bangladesh. The proposed system is very robust and can
detect license plates regardless of their shape. Although our
system performs very well in various scenarios, it has some
limitations. One such limitation is that the selected window
size for the detection phase is dependent on the distance of the
capturing camera from the vehicle. If the vehicle is too close,
the window size might be too small for detection and might
miss license plate areas. Although there are set standards for
Bangladeshi license plates, a lot of the vehicles use non-
standard license plates such as license plates printed on paper
instead of the color standard, single-line license plate instead
of the multi-line standard, etc. In the future, shape features
might be added with the existing features to get a robust
detection system. Real-time ALPDR systems from video
sequences in the context of Bangladesh might be developed.
The Optical Character Recognition (OCR) system based on
this dataset is currently under development.
REFERENCES
[1] C. N. E. Anagnostopoulos, I. E. Anagnostopoulos, I. D. Psoroulas, V.
Loumos, and E. Kayafas, “License plate recognition from still images
and video sequences: A survey,” IEEE Trans. Intell. Transp. Syst.,
2008.
[2] S. Du, M. Ibrahim, M. Shehata, and W. Badawy, “Automatic license
plate recognition (ALPR): A state-of-the-art review,” IEEE
Transactions on Circuits and Systems for Video Technology. 2013.
[3] R. Polishetty, M. Roopaei, and P. Rad, “A next-generation secure
cloud-based deep learning License Plate Recognition for smart cities,”
in Proceedings - 2016 15th IEEE International Conference on Machine
Learning and Applications, ICMLA 2016, 2017.
[4] P. Sadhukhan, “An IoT-based E-parking system for smart cities,” in
2017 International Conference on Advances in Computing,
Communications and Informatics, ICACCI 2017, 2017.
[5] Y. Joshi, P. Gharate, C. Ahire, N. Alai, and S. Sonavane, “Smart
parking management system using RFID and OCR,” in International
Conference on Energy Systems and Applications, ICESA 2015, 2016.
[6] “Bangladeshi motor vehicle registration plates statistics.” [Online].
Available: https://www.ceicdata.com/en/bangladesh/motor-vehicle-
registered/motor-vehicle-registered-bangladesh-total%0D%0A.
[Accessed: 01-Feb-2020].
[7] Z. Selmi, M. Ben Halima, and A. M. Alimi, “Deep Learning System
for Automatic License Plate Detection and Recognition,” in
Proceedings of the International Conference on Document Analysis
and Recognition, ICDAR, 2018.
[8] H. Li, P. Wang, and C. Shen, “Toward End-to-End Car License Plate
Detection and Recognition with Deep Neural Networks,” IEEE Trans.
Intell. Transp. Syst., 2019.
[9] M. Mondal, P. Mondal, N. Saha, and P. Chattopadhyay, “Automatic
number plate recognition using CNN based self synthesized feature
learning,” in 2017 IEEE Calcutta Conference, CALCON 2017 -
Proceedings, 2018.
[10] N. Saleem, H. Muazzam, H. M. Tahir, and U. Farooq, “Automatic
license plate recognition using extracted features,” in 2016 4th
International Symposium on Computational and Business Intelligence,
ISCBI 2016, 2016.
[11] J. H. Kim, S. K. Kim, S. H. Lee, T. M. Lee, and J. Lim, “License plate
detection and recognition algorithm for vehicle black box,” in 2017
International Automatic Control Conference, CACS 2017, 2018.
[12] R. A. Bedruz, E. Sybingco, A. R. Quiros, A. C. Uy, R. R. Vicerra, and
E. Dadios, “Fuzzy logic based vehicular plate character recognition
system using image segmentation and scale-invariant feature
transform,” in IEEE Region 10 Annual International Conference,
Proceedings/TENCON, 2017.
[13] N. A. Siddique, A. Iqbal, F. Mahmud, and M. S. Rahman,
“Development of an automatic vehicle license plate detection and
recognition system for Bangladesh,” in 2012 International Conference
on Informatics, Electronics & Vision (ICIEV), 2012, pp. 688–693.
[14] M. A. Uddin, J. B. Joolee, and S. A. Chowdhury, “Bangladeshi Vehicle
Digital License Plate Recognition for Metropolitan Cities Using
Support Vector Machine,” Proc. Int. Conf. Adv. Inf. Commun.
Technol., 2016.
[15] S. Abdullah, M. Mahedi Hasan, and S. Muhammad Saiful Islam,
“YOLO-Based Three-Stage Network for Bangla License Plate
Recognition in Dhaka Metropolitan City,” in 2018 International
Conference on Bangla Speech and Language Processing, ICBSLP
2018, 2018.
[16] K. Deb, M. K. Hossen, M. I. Khan, and M. R. Alam, “Bangladeshi
Vehicle License Plate Detection method based on HSI color model and
geometrical properties,” in Proceedings - 2012 7th International
Forum on Strategic Technology, IFOST 2012, 2012.
[17] W. Zhou, H. Li, Y. Lu, and Q. Tian, “Principal visual word discovery
for automatic license plate detection,” IEEE Trans. Image Process.,
2012.
[18] B. Li, B. Tian, Y. Li, and D. Wen, “Component-based license plate
detection using conditional random field model,” IEEE Trans. Intell.
Transp. Syst., 2013.
[19] Y. Yuan, W. Zou, Y. Zhao, X. Wang, X. Hu, and N. Komodakis, “A
Robust and Efficient Approach to License Plate Detection,” IEEE
Trans. Image Process., 2017.
[20] Z. Selmi, M. Ben Halima, U. Pal, and M. A. Alimi, “DELP-DAR
system for license plate detection and recognition,” Pattern Recognit.
Lett., 2020.
