Different Methodologies for Indian License Plate Detection

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 09 | Sep 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 924
Different Methodologies for Indian License Plate Detection
Atharvraj Patil1, Jayesh Ingale2
1Department of Electrical Engineering, College of Engineering, Pune, India
2Department of Electrical Engineering, College of Engineering, Pune, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - With the aid of a picture produced by security
cameras, automatic license plate detection systems (ALPDS)
have the potential to automatically track and identify the
vehicle by capturing and recognizing the number plate of
any vehicle. It can be used for a variety of practical
purposes, like recording vehicle numbers at toll gates,
tracking down stolen vehicles using CCTVs, and more.
Detecting the license plate from the input image is the
primary stage in automatic license plate recognition
systems, which are frequently employed for purposes like
effective parking management, updating law enforcement,
systematic toll way management on roads, etc. There are
numerous methods for detecting license plates. This study
focuses on various techniques for reading license plates
from input images of vehicles. Here, three key methods—
edge detection, object detection with YOLOv4 (You Only
Look Once) and WPOD-NET (Warped Planar Object
Detection Network)—are covered. This paper also provides
a description of each method's benefits and downsides.
Key Words: Edge detection, License plate detection,
WPOD-NET, YOLOv4.
1. INTRODUCTION
Massive growth of the vehicular sector has raised the
issues of parking management, traffic control, tracking
stolen vehicles, etc. Also, tracking individual vehicles has
become a tedious task for all organizations. To avoid
issues and for effective parking management, the
Automatic Number Plate Recognition Systems (ANPRS)
are designed. License plate recognition is the base of
ANPRS. It has a wide range of applications which uses
extracted license plate details to create automated
solutions for various problems like access control, tolling,
border control, traffic control, finding stolen cars, efficient
parking management and so on. The primary step in
ANPRS is license plate detection and extraction.
This paper discusses different methods of license plate
extraction where the vehicle’s static image is fed to the
system which further processes the image to extract the
license plate. Each of the techniques discussed in this
paper has various advantages and drawbacks. Like, a
method may require low processing power, but it fails for
different unconstrained scenarios. Another method may
be efficient for real time applications but then it requires
high processing power leading to limitations of hardware.
Thus, further sections describe in detail about various
methods of license plate detection.
2. LITERATURE SURVEY
In [1], the Automatic Number Plate Recognition System
(ANPRS) is divided into 3 main steps namely,
● Vehicle Detection
● License Plate Detection
● Optical Character Recognition (OCR).
There were various algorithms available for the process of
object detection like Fast Region Based Convolutional
Neural Network (RCNN), Single shot multibox detector
(SSD), You Only Look Once (YOLO), etc. After comparing
various algorithms for the process of vehicle detection
they found the YOLO algorithm as the most suited for this
work due to its advantage of quick real-time object
detection and used it for vehicle detection, Warped Planar
Object Detection Network (WPOD-NET), and Optical
Character Recognition (OCR) for License Plate (LP)
detection. They have employed various datasets for
training this network. These datasets consider various
factors, such as the LP angle (frontal and tilted), the
distance from vehicles to the camera (near, intermediate,
and far), and the area where the photographs were taken.
Finally, after training the model on all four datasets
average accuracy obtained was 89.33%. The earlier
YOLOv2 vehicle detection and OCR-NET were developed
and put into use using the DarkNet framework, but the
proposed WPOD-NET is implemented using the
TensorFlow framework. In [4], this project report presents
development on new approaches for the extraction of
license plates. Their proposed algorithm was based on
video acquisition, license plate region extraction, plate
character segmentation, and character recognition. This
project presents a simple straightforward license plate
extraction technique. The method's four main stages—
converting RGB images to grayscale, applying Gaussian
blur, performing morphological operations, and
determining the license plate's precise location—are

based on the Edge Detection algorithm. It uses a character
segmentation algorithm to detect the characters on the
number plate. Templates are created for each of the
alphanumeric characters (from A-Z and 0-9). The mean
squared error method is used for image similarity. Each
segmented character of the license plate is compared with
all the standard template characters and the error is
recorded and the final output is displayed on the monitor.
The proposed system has the following limitations like the
camera should be of good quality to extract correct
characters from license plates. It should have proper
lighting. It doesn’t work in different illumination
conditions and although accuracy is high, mean squared
error leads to low computational results.
The implementation of the proposed system can be
extended for the recognition of number plates of multiple
vehicles in a single image frame. User-friendly android
applications can be developed for traffic surveillance
management systems. Also, character recognition can be
done using various deep learning algorithms as they yield
more accuracy. (Graphical Processing Units) GPUs can be
used to achieve more performance in terms of
computational time.
3. DIFFERENT METHODOLOGIES
3.1 Edge Detection
Digital images have edges, which are notable local changes
in intensity. An edge is a collection of linked pixels that
identifies the boundary between two distinct sections.
Edge detection is the method of image pre-processing
used to recognize points from the input image and in
simple words, it identifies sharp changes in the brightness
of the image. It is a fundamental stage in image processing,
computer vision, and picture pattern recognition. Edge
Detection Operators are of two types:
● Gradient–based operator: First-order derivations
are computed using gradient-based operators in digital
images, such as the Sobel, Prewitt, and Robert operators.
● Gaussian–based operator: Second-order
derivations are computed using Gaussian–based operator
in a digital image like Canny edge detector and Laplacian
of Gaussian.
The input image is firstly converted from RGB to grayscale
to simply reduce complexity: from a 3D pixel value (R, G,
B) to a 1D value. Many tasks do not fare better with 3D
pixels (e.g., edge detection). To reduce the noise, we need
to blur the input image with Gaussian Blur then convert it
to grayscale. Image noise is a distortion in the image that
results from a camera malfunction or from reduced
visibility brought on by shifting weather conditions.
Noises are another term for the random change in pixel
intensity levels. There are many varieties of noise,
including salt & pepper noise and Gaussian noise. For
noise removal, we implemented an iterative bilateral filter.
A bilateral filter is used for smoothening images and
reducing noise, while preserving edges. More effectively
than the median filter, it offers the mechanism for noise
reduction while preserving edges. The process of
binarizing a picture results in an output image that only
has two-pixel values, white and black pixels. The work of
detecting license plates will be simpler if the binarization
step is carried out first since edges will be visible more
clearly in the binary image.
Fig -1: Image Preprocessing on input image
Our image is ready to discover contours after
preprocessing shown in figure 1. The remaining contours
are sent for additional processing once any contours
having an area of less than 30 are eliminated. Every
contour is roughly approximated to form a polygon, and if
a contour is quadrilateral in shape (has four sides), it is
expected to be the license plate. The contours are
then created, and the resulting image is as follows:
Fig -2: Extracted license plate

Results for leveled image and tilted image:
As the image shown in figure 3 is leveled, therefore the
detection and extraction are done properly as depicted in
4.
Fig -3: Detected license plate for levelled image
Fig -4: Extracted license plate for levelled image
For Tilted Image:
As shown in figure 5 and 6, the region of interest detected
is not accurate.
Fig -5: Incorrect license plate detection
Fig -6: Incorrect license plate extraction
This algorithm fails when the image is more tilted. Also, it
works only if the input image is of good quality. As this
algorithm is not efficient enough in real time, we have
tried other algorithms for license plate localization.
3.2 YOLOv4
You only look once (YOLO) is a family of one-stage object
detectors that are fast and accurate. All the YOLO models
are object detection models. Object detection models are
trained to look at an image and search for a subset of
object classes. When found, these object classes are
enclosed in a bounding box and their class is identified.
Object detection models are typically trained and
evaluated on the COCO (Common Objects in Context)
dataset, which contains a broad range of 80 object classes.
Real-time is particularly important for object detection
models that operate on video feeds, such as self driving
cars, etc. The other advantage of real time object detection
models is that they are small and easy to wield by all
developers.
YOLOv4 makes real time detection a priority and conducts
training on a single GPU.
The original YOLO (You Only Look Once) was written in a
custom framework called Darknet. Darknet is a very
flexible research framework written in low-level
languages and has produced a series of the best real-time
object detectors in computer vision.
The general architecture of YOLOv4 includes the
following:
Backbone is the deep learning architecture that basically
acts as a feature extractor. The neck has additional layers
between the head and the backbone. The head, often
referred to as the object detector, essentially locates the
area where the object may be present. Methods such as
"bag of freebies" (BoF) can improve object detecting
accuracy without raising inference costs. Bag of Specials
(BoS) are those plugin modules and post-processing
techniques that only slightly raise the inference cost but

greatly increase object detection accuracy. The general
architecture of YOLOv4 is depicted in figure 7.
Fig -7: General architecture of YOLOv4(One stage
detector)
Selection of architecture for the entire network is a very
tedious task as there can be a lot of possibilities. Thus,
depending upon some factors like number of
convolutional layers, number of parameters and additional
characteristics which include bag of specials, freebies, etc,
the final architecture of YOLOv4 can be presented as:
CSPDarkNet53 (Head) - SSP + PANet (Neck) -
YOLOv3(head)
The main idea behind the YOLOv4 object detector is that it
divides the image into evenly sized grids, and then
predicts the boundary boxes and corresponding
probability of each grid. Owing to this, YOLOv4 uses CIoU
as loss function as it considers distance between center
points and overlapping area, and it uses Mish activation
function.
As discussed, the YOLOv4 network has the ability of
detecting 80 classes, but for our real time application
detection of unnecessary classes leads to reduction of
efficiency and increases time of operation. Thus, in order
to deal with this problem, we decided to train a custom
YOLOv4 detector for a single class i.e., license plate.
Further in order to create a custom detector the primary
requirement is a good dataset of images and labels so that
the detector can be efficiently trained to detect objects.
We used the images from Google Open Images and created
a dataset of 5000 images and are split into training and
validation dataset in 80:20 ratio. All the images in the
training dataset were labeled with the bounding box
coordinates and further these labels are converted in
YOLOv4 format. The parameters for the training process
were set to optimal values found by fine tuning it:
batch=64, subdivisions=16, learning_rate=0.001,
max_batches = 6000, policy = steps, steps=4800, 5400.
Finally, we received an accuracy of 91%. The results of
YOLOv4 custom detector are shown in figure 8.
Fig -8: Result of YOLOv4 custom detector
Fig -9: Extracted license plate
Even though the YOLOv4 custom object detector has an
advantage of high-speed processing making it apt for real
time applications, it sometimes fails in unrestricted
situations when the LP may be significantly altered as a
result of oblique perspectives.
3.3 WPOD-NET
Warped Planar Object Detection Network (WPOD-NET) is
a novel Convolutional Neural Network (CNN) capable of
detecting and rectifying multiple distorted license plates
in a single image. The methodology of this network is
complex and is described further. The network is initially
fed by the vehicle detection's output that has been
downsized. An 8-channel feature map with object/non-
object probability and affine transformation parameters is
the result of the feed forwarding. Let's first consider an
imaginary square with a constant size surrounding a cell's
center in order to extract the warped LP (m, n). A portion
of the regressed parameters are then utilized to create an
affine matrix that converts the hypothetical square into an
LP region if the object probability for this cell is greater
than a specific detection threshold. There are a total of 21
convolutional layers in the proposed design. All
convolutional filters have a constant size of 3x3. Except for
the detection block, the entire network employs ReLU
activations. There are four max pooling layers with a size

of 2x2 and a stride of 2, which lowers the input
dimensionality by a factor of 16. Two parallel
convolutional layers make up the detection block's last
layer. The architecture of WPOD-NET can be seen in figure
10.
Fig -10: Architecture of WPOD-NET
A dataset of 196 photos is utilized to train the WPOD-NET,
comprising 105 images from the Cars Dataset, 40 images
from the SSIG Dataset (training subset), and 51 images
from the AOLP (Application-oriented License
Plate) dataset (LE subset). Four of the corners of the LP in
each image are manually annotated. Most of the LPs in the
Cars Dataset's chosen photos are European, but there are
also a lot of American and other types of LPs. Brazilian and
Taiwanese LPs are depicted in images from SSIG and
AOLP, respectively. Data augmentation is done on the
dataset because it is minimal in size.
Various augmentations are done on images such as
rectification, centering, scaling, rotation, mirroring and
translation. Due to hardware limitations, we studied
WPOD-NET architecture and decided to use pre-trained
weights trained on the dataset mentioned above. The
average accuracy obtained on all the dataset is 91%.
According to the WPOD_NET results, the suggested
strategy performs significantly better than previous
methods in complex datasets with LPs collected at
severely oblique viewpoints while maintaining good
results in more controlled datasets.
4. CONCLUSIONS
The above discussed methods work well in different
scenarios. As per discussed, edge detection algorithm
requires low computation power, but may not work well
in unconstrained scenarios. While YOLOv4 is known for
real time applications but struggles to detect small sized
objects on the other hand WPOD-NET works well with
small, oblique and distorted license plate images and in
unconstrained scenarios but requires high computation
power. Hence considering the requirement of application,
one can opt for any of these algorithms.
REFERENCES
[1] S´ergio Montazzolli Silva, Cl´audio Rosito Jung,
“License Plate Detection and Recognition in Unconstrained
Scenarios”, Institute of Informatics - Federal University of
Rio Grande do Sul Porto Alegre, Brazil, ECCV 2018.
[2] Cheng-Hun Lin, Yong-Sin Lin, Wei-Chen Lui, “An
Efficient License Plate Recognition System Using
Convolutional Neural Networks”, IEEE ICASI 2018
[3] Zhenbo Xu, Wei Yang, Ajin Meng, Nanxue Lu, Huan
Huang, Changchun Ying, and Liusheng Huang, “Towards
End-to-End License Plate Detection and Recognition: A
Large Dataset and Baseline”, School of Computer Science
and Technology, University of Science and Technology of
China, Hefei, China (2018).
[4] Arpitha D, Archana T, Rajashree Y G, Priya G,
“Automatic License Plate Recognition for Real-Time
Videos”, Visvesvaraya Technological University Belagavi,
Karnataka.
[5] V Gnanaprakash, N Kanthimathi, N Saranya,
“Automatic number plate recognition using deep
learning”, Bannari Amman Institute of Technology,
Department of Electronics and Communication
Engineering, Sathyamangalam, Erode, Tamilnadu, India,
ICCSSS 2020.
[6] Muhammad Tahir Qadri, Muhammad Asif, “Automatic
Number Plate Recognition System for Vehicle
Identification Using Optical Character Recognition”, 2009
International Conference on Education Technology and
Computer.
[7] Joseph Redmon, Ali Farhadi, “YOLO9000: Better,
Faster, Stronger”, CVPR 2017.
[8] Prof.Amit Kukreja, Swati Bhandari, Sayali Bhatkar,
Jyoti Chavda, Smita Lad, “Indian Vehicle Number Plate
Detection Using Image Processing”, Apr -2017.
[9] Ibtissam Slimani, Abdelmoghit Zaarane, Wahban Al
Okaishi, Issam Atouf, Abdellatif Hamdoun, An automated
license plate detection and recognition system based on
wavelet decomposition and CNN,Array, Volume 8,2020.
[10] M. Jogin, Mohana, M. S. Madhulika, G. D. Divya, R. K.
Meghana and S. Apoorva, "Feature Extraction using
Convolution Neural Networks (CNN) and Deep Learning,"

2018 3rd IEEE International Conference on Recent Trends
in Electronics, Information and Communication
Technology (RTEICT), 2018.
[11] Lubna, N. Mufti, and S. A. A. Shah, “Automatic Number
Plate Recognition:A Detailed Survey of Relevant
Algorithms,” Sensors, vol. 21, no. 9, p. 3028, Apr. 2021.
[12] J.-S. Chou and C.-H. Liu, “Automated Sensing System
for Real-Time Recognition of Trucks in River Dredging
Areas Using Computer Vision and Convolutional Deep
Learning,” Sensors, vol. 21, no. 2, p. 555, Jan. 2021.

Different Methodologies for Indian License Plate Detection

More Related Content

Similar to Different Methodologies for Indian License Plate Detection (20)

More from IRJET Journal (20)

Recently uploaded (20)

Different Methodologies for Indian License Plate Detection