2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)

Deep Learning Methods for Automatic Number Plate

Recognition System: A Review
2022 3rd International Conference on Intelligent Engineering and Management (ICIEM) | 978-1-6654-6756-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICIEM54221.2022.9853092

Sushant Poojary Manthan Naik Gaunkar Babani Dessai

Computer Engineering Computer Engineering Computer Engineering
Shri Rayeshwar Institute of Engineering Shri Rayeshwar Institute of Engineering Shri Rayeshwar Institute of Engineering
Information Technology Information Technology Information Technology
Shiroda, Goa Shiroda, Goa Shiroda, Goa
spoojari694@gmail.com Manthannaikgaunkar16@gamil.com dessaibabani@gmail.com

Manju Ayalli Prof. Vijay Kumar Naik Pawar Prof. (Dr.) Shailendra Aswale
Computer Engineering Shri Rayeshwar Institute of Engineering Shri Rayeshwar Institute of Engineering
Shri Rayeshwar Institute of Engineering Information Technology Information Technology
Information Technology Shiroda, Goa Shiroda, Goa
Shiroda, Goa vijaykumar.pawar@ritgoa.ac.in principal.ritgoa@gmail.com
ayallimanju4@gmail.com

Abstract— Automatic number plate recognition(ANPR) is one
of the smart transportation method that provides a
safe medium of transportation and monitoring planning used
by many two wheeler vehicles as well as four wheeler vehicles to
boost traffic controlling, routing, parking system, toll collection,
and governance and insuring highway law. There are various
techniques and algorithms has been developed so that we can
identify and recognize Number plates. This paper shows a
various deep learning approach for recognizing and detecting
number plate. This system is divided into three parts, i.e.
number plate detection, segmentation, and character
recognition. In our proposed system for number plate detection
Yolo method is used, after that some filters are applied and then
characters are segmented. Finally the Convolutional Neural
Network is used to recognize all the segmented characters.
Keywords— Number Plate detection, Number plate recognition
system, Character segmentation, Character recognition, , CNN,
Deep Learning, Yolo algorithm.
I. INTRODUCTION
Rapid road traffic requires a large amount of traffic
monitoring and management skills. In this case, you will not
be able to manually track vehicles moving at high speeds on
the road. Also, human energy and time is wasted. It is having
been operated manually, that will cause a lot of difficulty and
many mistakes. Hence there is a necessary to make an
automated system that assist in tracking vehicles by tracking
their number plate in a more efficiently.
Automatic Number plate recognition (ANPR) became one of
the simplest ways for identifying cars [1]. It can be used in
many countries for many applications like keeping traffic
administration, vehicle tax, car parking, Toll collection etc.
The ANPR algorithm is divided into four categories: Vehicle
image acquiring, number plate detection, character
segmentation and recognition.
First, the camera captures images of a car that contains a
number plate. From it, the number plate is detected and the
extracted number plate contains numbers and alphabets. And
then the various image processing algorithms are used [12] to
analyze the image and find the numbers, alphabets from the
image. After which second part is recognition, in this case
after finding the exact number, this number is searched in the
database for owner information. Thus we can verify whether
a user is authenticated or not.
The automatic number plate recognition (ANPR) approach
can be applied depending on different methodologies.
Although ANPR uses many algorithms but still it is not
successful in a state of precision and performance in real
time. This can be reduced by applying deep learning
techniques. Deep learning is vast area of artificial intelligence
(AI) that uses neural network to learn from vast amounts of
data. This is part of machine learning [4] and uses many
layers to get advanced functionality from raw input. Today,
deep learning is used in almost every real-time application.
Compared to other algorithms, it is very efficient and has
proven to be as error-free as possible.
The methods considered in this project are briefly detailed in
the sections. Section II gives a literature review of the work.
Table giving the efficiency of each particular method in
section II. With a diagram that briefs the working of our
project in section IV and ended the paper in section V.
II. LITERATURE SURVEY
A. Number Plate Detection
In this phase, we cut out the various boundary boxes that
could be treated as number plates. However to determine if a
number plate from various boundary boxes are
acceptable, the deep learning architecture presented in the
CNN model [3] is integrated to filter and identify whether the
number plate is a number plate., The methods are:

19
978-1-6654-6756-8/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)
1. RNN: R-CNN is a CNN detector-based site. R-CNN that
could be applied in object search operations, their
functions first showed that CNN can induce detection of
something much higher in PASCAL VOC records as
such systems based on a simple HOG-like [8].
Appearances. An in-depth verified learning methods are
effective in the area of object discovery.
The R-CNN detector has 4 parts. First part makes
independent regional recommendations for class. The
second part emits a fixed vector of vertical motion for
each area. The third part is a set of particular types of
SVM line sorting in a single image. The final module is
an integrated regressor box for accurate box estimates.
With the details, first of all, to produce regional
suggestions, the authors adopted a preferred search
method. Next, CNN used to obtain a 4096-
dimensional vector from each area suggestion. Since a
completely compacted layer requires the input of vectors
of definite length, the regional motion parameters must
be the similar size. The authors accepted the 227 × 227
fixed pixels as the size of the CNN input. As you know,
objects in different images have different field sizes and
dimensions, this will issue regional suggestions from the
beginning of the various sizes.
Anyways the size of the area of?? Interest or the size of
the aspect ratio, the author combines each and every
pixel of the rotating bound box of necessary size 227 ×
227. And entire CNN parameter is shared between
categories. Every category focus an independent SVM
category that don't share similarity between other SVMs.
2. Fast R-CNN: Three months after the launch of the R-
CNN program, the R-CNN program [8] coordinated the
CNN local program. The R-CNN application uses a
selected search engine to measure RoI, which is faster
and more efficient as a search engine. R-CNN quickly
transformed the RPN newsletter (local planning
network) into an online communication network that
accurately reports local areas with large and large scale.
RPN is rapidly developing a local strategy because it
shares the full features of image conversion and
conversion process in the research networks.
Also, a new way of finding different sizes is that
multiple-sided anchors are used for a reference. Anchors
may ease the process of creating different site plans
without the need for multiple scales for installation or
design. The outer area (map mark) of this last section
shares a flexible layer, which produces a fixed window
size (3 × 3), in the center where the location of each
window is adjacent to the center in the first central image
installation. of k (3 × 3) anchor box. The author describes
three different sizes and three categories of anchor boxes.
Then calculated the distance between the corresponding
box and the limited box.
20
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
3. Faster R-CNN: Later 3 months after the launch of the R-
CNN program, the R-CNN program [8] coordinated the
CNN local program. The R-CNN application uses a
selected search engine to measure RoI, which is faster
and more efficient as a search engine. R-CNN quickly
transformed the RPN newsletter (local planning
network) into an online communication network that
accurately reports local areas with large and large scale.
RPN is rapidly developing a local strategy because it
shares the full features of image conversion and
conversion process in the research network. The
procedure is simplified in Fig.3 (b). Also, a new way of
finding different sizes is that multi-sided anchors are
used as a reference. Anchor can ease the process of
creating different site plans without the need for multiple
scales for installation or design. The outer area (map
mark) of this last section shares a flexible layer, which
produces a fixed window size (3 × 3), in the center where
the location of each window is adjacent to the center in
the first central image installation of k (3 × 3) anchor
box. The author describes three different sizes and three
categories of anchor boxes. Then calculated the distance
between the corresponding box and the limited box
4. Yolo: YOLO (you only see once) one-part R-CNN is
running. The registration key is to take the perfect
webcam image quickly. First, it is due to the fact that
these pipelines estimate less than a hundred boxes per
image while Fast R-CNN uses the imaginary 2000-point
image on the image. Second, the availability of the
YOLO frame as a backlash problem, so some settings
can remove items from the image by quickly entering the
merge box and classroom space. The YOLO network
operates at 45 FPM [10] without major significant
reliability is achieved by multi-dimensional, where P
(product) indicates the end of the box containing the
product and IOU (product intersection) indicates the box.
Each cell imaging box B bind (x, y, w, h) with the
numbers depending on the class location and the C-block
C class in the class. The range of products includes 24
different and Upgrades to the Titan X GPU, Compared
to 0.5frame per second fast RCNN and 7frame per
second RCNN. The YOLO lights begins with splitting
images into V × V grid, with the cell grid responsible for
searching for the central object.
5. Mask R-CNN: The mask R-CNN uses the same two-
stage approach, similar to the first level in the second
stage, matching the square measurements and the offset
drawer, and the R-CNN mask also produces a double
mask for each RoI. This is very different from the latter
systems, where there is a division into mask values. Our
method quickly follows the fast R-CNN, which uses a
separate binding box separator and a droplet that
simplifies many phases of the R’s multi stage pipeline. –
CNN
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)

detection is achieved by using multiple boxes or filters of

6. R-FCN: Region-based Fully Convolutional
Networks(R-FCN) is a form of local object acquisition.
Unlike previous regional recipients, such as Fast / Fast
RCNN, using hundreds of low-cost network locations in
each location, R-FCN is fully convolution with the
common values in the whole picture. To achieve this, R-
FCN uses sensitivity maps to solve the problem
between image classification and translation-variance in
object detection.
7. SSD: Single Shot detector like YOLO takes just one
photo to see many things in a photo with the help of
several multibox. This is very fast and the accuracy of the
object detection algorithm. A quick comparison between
the speed and accuracy of the models to find the
difference of VOC2007. The high accuracy of SSD
different sizes, as well as visual field measurements.
Apply these filters to cards with multiple symbols from
the last level of the network. It helps to perform
detection in many sizes.
8. BlitzNet: For many applications, such as auto-drive, it is
important to understand the real-time event. In this page,
we introduce an advanced architecture called BlitzNet,
which allows real-time computer use, object integration
and semantic separation. In addition to the calculation
success of a single multi-tasking network, we show that
object detection and semantic separation benefit each
other with accuracy. The VOC and COCO dataset test
results provide a modern benchmark for discovery and
differentiation between real-time systems.

TABLE I. SUMMARY OF THE NUMBER PLATE DETECTION

Reference
using Methods Accuracy Advantages Limitations Future Scope
Algorithm
R-CNN (Regional The very complexity of time that
Achieve end-to-end
[5] Convolution Neural 66.0% makes RCNN unsuitable for real-
structure in real time.
Network) world applications.
However, with the introduction of The problem of generating
Fast R-CNN (Regional the layer of RoI pooling, time inaccurate region suggestions
[5] Convolution Neural 66.9% complexity of FastRCNN has caused by the inability to train
Network) decreased somewhat compared to selective search algorithms exists
RCNN. in FastRCNN as well as in RCNN.
Because of the efficiency of faster
RCNNs It performs the function of To make an end to end
Faster R-CNN (Regional Requires many passes through a
generating accurate region system and achieve more
[7] Convolution Neural 69.9% single image to extract all the
proposals and significantly performance for
Network) objects.
reducing the time complexity of processing.
RCNNs and fast RCNNs.
Discover CNN's new
Yolo (you only look The network is able to generalize
[22] 98.52% Struggles to detect small objects. approach to optimizing
once) the image better.
detection.
simple and effective
Addresses the challenge of
classification method to
[23] MASK R-CNN 97.1% inaccurate bounding boxes and
optimize and improve
masks.
network.
The BlitzNet pipeline processes
[24] BlitzNet 83.6% multiple types of data at the same
time.
The model is further
trained with more datasets
under different
conditions, and the
SSD is a technique for creating a developed LPR is
SSD (Single Shot bounding box for each segment combined with
[22] 96.94%
detector) that is likely to be the same object motorcycle and helmet
in the image. wearing systems to
enhance traffic safety in
conjunction with
surveillance video camera
systems increase.
This method requires a large
R-FCN(Region-based number of legitimate vehicle
Decrease the detection time and
[25] Fully Convolutional 87.48% samples and cannot capture and
improving vehicle detection rate.
Networks) recognize small targets in distant
scenes.

B. Character Recognition

21

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)
Character recognition is a final stage in this system There are
currently many method are applied to the character
recognition, such as syntactic, statistical and neural
networks[14] to recognize the characters on an number plate
we separate the characters and numbers using a CNN model
with 37 classes. There are 36 entry classes which are classes
of digits (0 - 9) and 26 uppercase characters (A - Z) and
another non-character class are also considered. The methods
as follow:
1. Artificial Neural Network (ANN): Artificial Neural
Network (ANN) is also known as neural network is
mathematical word, containing interconnected artificial
neurons. In ANN two pieces Access neural network with
topology 180-180- 36. Procedural adjustment procedures
are performed 128 ms. In Multilayered Perceptron
(MLP), the ANN mode is used for stratification of
character. Inside there is a way to enter options, hide
layer to read much complex associations and output layer
for giving decision. Feeding back to the Back-
Propagation (BP) algorithm and used to train ANN [14].
BP neural network-based system are used with a working
time of 0.06s, ANN is used reduce obscurity between
the same characters.
2. Template Matching: Template matching is important for
comparing colors with same size characters. It is also
used for routine facial examination and medical imaging.
It is divided into two parts which are feature-based
matching and template-based matching. The method-
based approach works well when the diagram has a
robust character that limit the functionality depending on
the model to run. In this method is used to obtain the
receiving rate of 85%. In it, many marks are removed and
subtraction is as important as training. Standard
algorithms are used to calculate all images of the same
size. A figure of 95.7% was seen in 1176 images. The
source achieved 99.5%, 98.6%, and 97.8% success rate
of the total [15], or, at the address found. This method is
described in for measuring race interactions.
3. Back Propagation Neural Network (BPN): Back
Propagation Neural (BPN) Network which has
mathematical fundament, the BPN network is used for
training multi-layer ANN. The purpose of BPN network
is to train the neural network. The patterns used as input
are balanced among them. BPN method is categorized
into two portions. The training segment and the
recognition segment.
4. Feed Forward Back Propagation: Feed Forward Back
propagation is the feature of the Back Propagation (BP)
algorithm is the detected error. The error received from
the neural network will occur with its output. The
difference between the initial value and the actual value
is equal to the error obtained. On the output, the BP
algorithm lessens error.
22
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
5. Convolutional Neural Network (CNN): A neural
network is a set of algorithms designed to recognize
structure. CNN has many applications in Natural
Language Processing (NPL), image and video
recognition, and recommendation engines. Excellent for
Neural Network (NN) inspired by biological images and
recognition. The convolution layer is the most important
layer in this network. Used to perform convolution
operations. The next layer is the pooling layer. This large
layer for the trainable parameters .The trainable
parameters of this layer are large. This layer helps to
reduce the size of the image. The actual size of the image
is 128x128 pixels. The image size is reduced to 28x28,
reducing the time required for training the neural
network. The range of 94% accuracy is obtained after
training the neural network model.
6. Deep Neural Network (DNN): Deep Neural Network
(DNN) is a neural network with three or more layers.
These layers are connected to other nodes called
interconnect nodes. These layers communicate with each
other through hidden layers. The actual processing is
done over a weighted connection. These hidden layers
are connected to the output layer.
7. K-Nearest Neighbour (KNN): The KNN algorithm is
one of the simple classification algorithms and one of the
most commonly used for learning algorithms. KNN is
not parametric because it does not contain explicit
assumptions about the relationship between the predictor
and the dependent variable. The characteristics of kNN
explain that it is a lazy learning algorithm because it
makes predictions based on local information and does
not build an explicit model. Its purpose is to predict the
classification of new sample points using a database
where the data points are divided into several classes.
This method retrieves k training samples whose
attributes are relatively similar (closest) to the test
sample. Therefore, the test samples are categorized based
on the class specification of the closest training sample.
8. Support Vector Machine (SVM): Support Vector
Machine (SVM) is used in character recognition, and it
is a supervised learning technique mostly used in
regression and classification. Each successful candidate's
recognition will be handled by SVM, which will test
against a library of trained license plate number models.
The accuracy is calculated by classifying the SVM
outputs based on the maximum value. SVM, however,
must be trained using valid number plate sample
database prior to use. The 2 types of SVM classifier are
OVA (One Against All) and OVO (One Against One).
TABLE II. SUMMARY OF THE CHARACTER RECOGNITION
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)

Reference Success
using Methods Rate Types of Categories
Algorithm (%)
ANN(Artificial English letters and
[17] 98.17%
Neural Network ) digit
English letters and
[13] Template matching 96.7%
digit
Chinese letters,
Back Propagation
[21] 93.5% English letters and
Neural Network
digit
Feed Forward Back English letters
[20] 95.69%
propagation and digit
CNN(Convolutional English letters Figure 2: vehicle Image
[8] 92.7%
Neural Network) and digit
DNN (Deep Neural 2. Number Plate Detection
[16] 90.6% English letters
Network)
KNN (K Nearest English letters YOLO method is used for localization of number plate, with
[2] 96.22%
Neighbours) and digit this we can separate number plate from the vehicle. This
[12]
SVM (Support
98.05%
English letters process removes the other complex pre-processing like edge
Vector Machine) and digit and contour detection for number plate detection [7]. The
number plates may be recognized under unexpected or
III. METHODOLOGY inappropriate conditions like the angle of number plate
placement, brightness, etc.
Our main motive is to make a framework which detects and
recognizes number plate in a given image or video. Our
approach is designed to focus on finding and reading number
Plate in tough environments. The proposed approach and it is
divided into different stages as shown in fig.1. In this we
proposes the number plate detection and recognition system
which is based upon the YOLO that works with efficiency
under a variety of conditions. YOLO is system which can be
trained to run a regression task evaluate both objects
bounding the box and an object class. This technique is Figure 3: Number plate detected from vehicle using YOLO
surprisingly fast, it continuously forms an image at 45 frames
per second.

Figure 4: Detected number plate result after using YOLO processing

3. Character Segmentation

Segmentation is one of the most important techniques for

programmatically recognizing tags.[4]. To ensure legitimate
segmentation, first processing should be performed. First, the
detected number plate result in fig.4 is converted from BGR
into Grayscale. Then it maximizes the contrast to improve
and clarify better the image. Various filter are then applied.
After this process the characters can be easily segmented
from the detected number plate using contour detection. This
Figure 1: flow chart process shown in figures:
1. Image Acquisition

The input can be both a image or video. Video is a sequence

of image. Before start with detection of number plate, the
image sample for input should be checked if it is suitable for Figure 4: BGR into Grayscale
further process .The image sample used for input is shown in
Fig.2 and this image sample will be processed.

23

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)
Figure 5: Blurred image with a Gaussian filter of detected number plate
Figure 6: Binary conversion of detected number plate
Figure 7: Results of contour detection
4. Character Recognition
The final step in the number plate detection is character
recognition. There are currently so many method applied for
character recognition, which can be the syntactic, statistical
and neural networks [14][15]. In this we will be using
Convolutional Neural Network (CNN) for recognize the
characters on number plate. We separate the characters and
numbers using a CNN model with 37 classes. There are 36
entry classes which are classes of digits (0 - 9) and 26
uppercase characters (A - Z) and another non-character class
are also considered.
Figure 7: Recognition of the number plate
Final Output in character recognition of the number plate
shown in Fig.7.
IV. CONCLUSION
Automatic Number Plate Recognition (ANPR) system is a
significant execution of artificial intelligence solutions for
numerous aspect. It is generally helpful for security purposes,
keeping vehicle record, toll collection, improved traffic
monitoring, better parking system, vehicle tracking, etc.
Many studies have been conducted on automatic number
plate detection as well as character recognition. In reality, a
variety of researchers have been done on many methods and
techniques for this process. From the survey we can
understand that there are different methods and techniques
used and it has advantages and disadvantages, and the
effectiveness of technique is different from each other. Each
country has its own number plate numbering system,
different number plate sizes and colors, and character
language. The above survey can give you that Valuable
understanding and instructions for the approach use. We have
proposed a deep learning technique represented by the CNN
model for both number plate detection and character
24
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.
recognition for ensuring a decent function of automatic
Number Plate Recognition system.
REFERENCES
[1] Selmi, Zied; Ben Halima, Mohamed; Alimi, Adel M. (2017). [IEEE
2017 14th IAPR International Conference on Document Analysis and
Recognition (ICDAR) - Kyoto, Japan (2017.11.9-2017.11.15)] 2017
14th IAPR International Conference on Document Analysis and
Recognition (ICDAR) - Deep Learning System for Automatic License
Plate Detection and Recognition.
[2] Varma P, Ravi Kiran; Ganta, Srikanth; B, Hari Krishna; Svsrk, Praveen
(2020). A Novel Method for Indian Vehicle Registration Number Plate
Detection and Recognition using Image Processing Techniques.
Procedia Computer Science.
[3] Xie, Lele; Ahmad, Tasweer; Jin, Lianwen; Liu, Yuliang; Zhang, Sheng
(2018). A New CNN-Based Method for Multi-Directional Car License
Plate Detection. IEEE Transactions on Intelligent Transportation
Systems.
[4] Hossen, M. S., Akter, M., & Saifuddin, K. M. (2021). Automatic Digit
and Alphabet Recognition Based Online Toll Collection System. 2021
International Conference on Information and Communication
Technology for Sustainable Development (ICICT4SD).
[5] Jamtsho, Yonten; Riyamongkol, Panomkhawn; Waranusast,
Rattapoom (2019). Real-time Bhutanese license plate localization
using YOLO. ICT Express.
[6] Mondal, Madhusree; Mondal, Parmita; Saha, Nilendu; Chattopadhyay,
Paramita (2017). [IEEE 2017 IEEE Calcutta Conference (CALCON) -
Kolkata, India (2017.12.2-2017.12.3)] 2017 IEEE Calcutta Conference
(CALCON) - Automatic number plate recognition using CNN based
self-synthesized feature learning.
[7] Hashmi, S. N., Kumar, K., Khandelwal, S., Lochan, D., & Mittal, S.
(2019). Real Time License Plate Recognition from Video Streams
using Deep Learning. International Journal of Information Retrieval
Research.
[8] Abdullah, Sohaib; Mahedi Hasan, Md; Muhammad Saiful Islam,
Sheikh (2018). [IEEE 2018 International Conference on Bangla Speech
and Language Processing (ICBSLP) - Sylhet, Bangladesh (2018.9.21-
2018.9.22)] 2018 International Conference on Bangla Speech and
Language Processing (ICBSLP) - YOLO-Based Three-Stage Network
for Bangla License Plate Recognition in Dhaka Metropolitan City.
[9] Cheng-Hung Lin, Yong-Sin Lin, and Wei-Chen Liu, An Efficient
License [IEEE International Conference on Applied System Innovation
2018] Plate Recognition System Using Convolution Neural Networks
[10] Hendry, Chen, Rung-Ching (2019). Automatic License Plate
Recognition Via Sliding-Window Darknet-Yolo Deep Learning. Image
and Vision Computing.
[11] Gnanaprakash, V., Kanthimathi, N., & Saranya, N. (2021). Automatic
number plate recognition using deep learning. IOP Conference Series:
Materials Science and Engineering.
[12] Hafiz Ahamed , Ishraq Alam , Md. Manirul Islam [International
Conference on Engineering Research and Education School of Applied
sciences & Technology, SUST, Sylhet 2019] SVM Based Real Time
Hand-Written Digit Recognition System.
[13] Avinash Bhujbal, Deepak ane [INTERNATIONAL JOURNAL OF
SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 8, ISSUE
12, DECEMBER 2019] A Survey on Deep Learning Approaches for
Vehicle and Number Plate Detection.
[14] Jiao, Licheng; Zhang, Fan; Liu, Fang; Yang, Shuyuan; Li, Lingling;
Feng, Zhixi; Qu, Rong (2019). A Survey of Deep Learning-based
Object Detection. IEEE.
[15] Tian Ying , Lu Xin , Li Wanxiang [ The 30th Chinese Control and
Decision Conference (2018 CCDC)] License Plate Detection and
Localization in Complex Scenes Based on Deep Learning.
[16] Kessentini, Yousri; Besbes, Mohamed Dhia; Ammar, Sourour;
Chabbouh, Achraf (2019). A Two-Stage Deep Neural Network for
Multi-norm License Plate Detection and Recognition. Expert Systems
with Applications.
[17] Hinduja, Desai; Dheebhika, R.; Jacob, T. Prem (2019). [IEEE 2019
International Conference on Communication and Signal Processing
 2022 3rd International Conference on Intelligent Engineering and Management (ICIEM)

(ICCSP) - Chennai, India (2019.4.4-2019.4.6)] 2019 International

Conference on Communication and Signal Processing (ICCSP) -
Enhanced Character Recognition using Deep Neural Network-A
Survey.
[18] Wei, Tan Chiang; Sheikh, U. U.; Rahman, Ab Al-Hadi Ab
(2018). [IEEE 2018 IEEE 14th International Colloquium on Signal
Processing & Its Applications (CSPA) - Malaysia (2018.3.9-
2018.3.10)] 2018 IEEE 14th International Colloquium on Signal
Processing & Its Applications (CSPA) - Improved optical character
recognition with deep neural network.
[19] Yousaf, Adeel; Khan, Muhammad Junaid; Khan, Muhammad Jaleed;
Javed, Nizwa; Ibrahim, Haroon; Khurshid, Khurram; Khurshid,
Khawar (2019). [IEEE 2019 2nd International Conference on
Computing, Mathematics and Engineering Technologies (iCoMET)
Size Invariant Handwritten Character Recognition using Single Layer
Feedforward Backpropagation Neural Networks.
[20] Tuba, Milan; Akashe, Shyam; Joshi, Amit (2020). Information and
Communication Technology for Sustainable Development Volume
933(Proceedings of ICT4SD 2018) || Template Matching for
Automatic Number Plate RecognitionSystem with Optical Character
Recognition.
[21] Singh, Ashima; Desai, Swapnil (2016). [IEEE 2016 International
Conference on Inventive Computation Technologies (ICICT) -
Coimbatore, India (2016.8.26-2016.8.27)] 2016 International
Conference on Inventive Computation Technologies (ICICT) - Optical
character recognition using template matching and back propagation
algorithm.
[22] Puarungroj, Wichai; Boonsirisumpun, Narong (2018). Thai License
Plate Recognition Based on Deep Learning. Procedia Computer
Science.
[23] Wu, Minghu; Yue, Hanhui; Wang, Juan; Huang, Yongxi; Liu, Min;
Jiang, Yuhan; Ke, Cong; Zeng, Cheng (2020). Object detection based
on RGC mask R-CNN. IET Image Processing,
[24] Dvornik, Nikita; Shmelkov, Konstantin; Mairal, Julien; Schmid,
Cordelia (2017). [IEEE 2017 IEEE International Conference on
Computer Vision (ICCV) 2017 IEEE International Conference on
Computer Vision (ICCV) - BlitzNet: A Real-Time Deep Network for
Scene Understanding.
[25] Zhigang, Zhou; Huan, Lei; Pengcheng, Ding; Guangbing, Zhou; Nan,
Wang; Wei-Kun, Zhou (2018). [IEEE 2018 Chinese Control And
Decision Conference (CCDC) - Shenyang, China (2018.6.9-
2018.6.11)] 2018 Chinese Control And Decision Conference (CCDC)
- Vehicle target detection based on R-FCN.

25

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on December 24,2022 at 17:43:19 UTC from IEEE Xplore. Restrictions apply.