LICENSE PLATE RECOGNITION

A PROJECT REPORT

Submitted in Partial Fulfilment for the degree of

Bachelor of Technology

In COMPUTER SCIENCE AND ENGINEERING

From

MAULANA ABUL KALAM AZAD UNIVERSITY OF

TECHNOLOGY

Submitted By

NAME OF THE STUDENT EXAMINATION ROLL NO.

SAPTARSHI BANIK 11600221038
SREEJEETA DUTTA 11600221054
ADITYA VARDHAN RAI 11600221004
SRIJIT CHAKRABORTY 11600221055
Under the guidance of
MS. MANALI SARKAR
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
MCKV Institute of Engineering
243, G.T. Road (North), Lilah, Howrah-711204
(NAAC Accredited Grade- A Autonomous Institute affiliated to
Maulana Abul Kalam Azad University of Technology, West Bengal)
NOVEMBER, 2024
Appendix 2
i
 DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
MCKV Institute of Engineering
An Autonomous Institute Affiliated to
Maulana Abul Kalam Azad University of Technology, West Bengal

CERTIFICATE

This is to certify that the project report entitled “LICENSE PLATE

RECOGNITION” submitted by SAPTARSHI BANIK, SREEJEETA
DUTTA, ADITYA VARDHAN RAI, SRIJIT CHAKRABORTY for 7th/8th
semester examination has been prepared for the partial fulfilment of the
degree of Bachelor of Technology in Computer Science and Engineering,
awarded by the Maulana Abul Kalam Azad University of Technology, West
Bengal. They have carried out the project work under my supervision.

_____________________________________ _____________________________________

HEAD OF THE DEPARTMENT SUPERVISOR

Mr. Avijit Bose Ms. Manali Sarkar

Assistant Professor Assistant Professor

Dept. of Computer Science and Engg. Dept. of Computer Science and Engg.
MCKV Institute of Engineering MCKV Institute of Engineering
243, G.T. Road (North), Liluah 243, G.T. Road (North), Liluah
Howrah – 711204 Howrah – 711204

Appendix 3

ii
 DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
MCKV Institute of Engineering
An Autonomous Institute Affiliated to
Maulana Abul Kalam Azad University of Technology, West Bengal

CERTIFICATE OF APPROVAL
(B.Tech Degree in Computer Science and Engineering)

This project report is hereby approved as a creditable study of an engineering

subject carried out and presented in a manner satisfactory to warrant its
acceptance as a prerequisite to the degree for which it has been submitted. It is to
be understood that by this approval, the undersigned do not necessarily endorse or
approve any statement made, opinion expressed and conclusion drawn therein but
approve the project report only for the purpose for which it has been submitted

1. _______________
SIGNATURE OF
EXAMINATION 2. _______________
COMMITTEE
MEMBERS FOR 3. _______________
EVALUATION OF 4. _______________
PROJECT REPORT
5. _______________

Date of evaluation:

iii
 Appendix 4

ACKNOWLEDGEMENT
We would like to extend our heartfelt gratitude to all those who have
supported and guided us throughout the development of our group
project on License Plate Recognition.
First and foremost, we are deeply indebted to our project guide, Ms.
Manali Sarkar, for her invaluable guidance, insightful feedback, and
unwavering support. Her expertise and dedication have been
instrumental in shaping the direction and success of this project. We
are also profoundly grateful to our Head of Department, Mr. Avijit
Bose, for his constant support and encouragement, which have been a
source of motivation and inspiration for our team. Our sincere thanks
go to our principal, Dr. Abhijit Lahiri, for providing a conducive
learning environment and the necessary resources to carry out this
project. We would also like to thank all the faculty members of the
Computer Science and Engineering department for their valuable
teachings and continuous support. A special mention goes to our
friends and peers, who have been there to offer support, share ideas,
and provide encouragement during challenging times. Lastly, we
express our deepest appreciation to our families for their love and
support. Thank you for contributing to the successful completion of
our project work for this semester.

__________________________

SAPTARSHI BANIK
__________________________

SREEJEETA DUTTA
__________________________

ADITYA VARDHAN RAI

___________________________
SRIJIT CHAKRABORTY

iv
 Appendix 5

ABSTRACT

The License Plate Recognition (LPR) system is a cutting-edge solution

designed to enhance vehicle identification and streamline traffic management
processes, playing a pivotal role in law enforcement, toll collection, and overall
public safety. Traditionally employed by law enforcement, the technology’s
applications have broadened significantly, now serving as an essential tool in
diverse areas such as security monitoring and traffic regulation. This project
presents an LPR system specifically developed to address the unique challenges
posed by Indian vehicle identification, including the variety of license plate
designs, regional differences in fonts, and inconsistent plate conditions.
Leveraging advanced machine learning algorithms and high-resolution image
processing, this LPR system achieves high accuracy in recognizing license
plates, with real-time processing capabilities that remain robust even under
challenging conditions such as low lighting and high vehicle speeds. Initially
implemented with on-device computer vision and YOLO for object detection, the
system has now transitioned to a cloud-based architecture. This shift enables
real-time multi-user access, reduces end-user hardware demands, and supports
efficient, scalable data processing, enhancing the system’s accessibility and
adaptability across diverse Indian environments. The LPR system not only
integrates seamlessly with existing infrastructure but also provides a scalable,
sustainable solution for smart transportation and improved traffic management
throughout India.
This enhanced LPR solution offers substantial benefits for traffic management
and public safety, as it enables authorities to monitor and regulate traffic more
effectively across diverse Indian environments. By adapting to local conditions
and handling varied license plate designs with high accuracy, the system
empowers smarter transportation infrastructure, contributing to a safer and
more efficient road ecosystem. The scalability and centralized nature of this

v
 LPR system also make it a sustainable choice, capable of growing alongside
India’s dynamic transportation landscape.
Appendix 6

TABLE OF CONTENTS

CHAPTER NO. TITLE PAGE NO.

ABSTRACT v
LIST OF FIGURES vii

1. INTRODUCTION 1

2. LITERATURE REVIEW 5

3. SCHEME AND IMPLEMENTATION 7

4. RESULTS AND DISCUSSIONS 15

5. CONCLUSION 17

6. FUTURE SCOPE 19

REFERENCES 21

vi
 Appendix 8

LIST OF FIGURES

Figure. 3.4.1 Architecture Map 12

Figure. 3.4.2 Working of EasyOCR 12
Figure. 3.5.1 Indian License Plate Format 13
Figure. 4.1 Confusion Matrix 15
Figure. 4.2.1 Performance Metrics 16

vii
 1. Introduction

1.1 Background of the Project

The License Plate Recognition (LPR) system represents a technological leap in vehicle
identification, initially serving law enforcement agencies for traffic enforcement but
quickly expanding to a broader range of applications. From toll collection to surveillance
and traffic management, LPR systems have become indispensable in modern
infrastructure. The primary function of an LPR system is to identify and track vehicles in
real-time, using specialized cameras and algorithms. However, the Indian context
introduces unique challenges, as vehicle license plates vary significantly in design, font,
language, and condition. Regional inconsistencies and differing standards further
complicate accurate recognition.

Recognizing these challenges, the project aims to create a highly accurate, scalable LPR
solution tailored to India’s distinct needs. Initially deployed through local systems using
computer vision and the YOLO (You Only Look Once) detection model, the system has
now transitioned to a cloud-based model. This shift enhances the system’s accessibility,
enabling real-time, multi-user access, and reducing hardware demands for end-users.
Through this transition, the project aspires to provide a seamless, sustainable traffic
management solution that will positively impact public safety and transportation
efficiency across India, in line with the country's rapid infrastructural advancements.

1.2 Problem Statement

India’s vehicle identification landscape presents unique challenges for license plate
recognition systems, primarily due to the diversity in license plate designs, languages,
fonts, and conditions across the country. While traditional LPR systems operate
effectively in more standardized environments, they struggle with the Indian scenario,
where plate variations and wear-and-tear are common. Additionally, lighting conditions,

1
varying speeds, and the sheer volume of traffic can make it difficult for traditional LPR
systems to achieve the required accuracy and reliability.

In the absence of a scalable solution, managing traffic, enforcing laws, and ensuring
security become challenging in India’s dynamic and densely populated environments.
Existing systems may also require significant hardware infrastructure on-site, making
deployment costly and inefficient in many regions. This project aims to tackle these
issues by developing a robust, cloud-based LPR system that can accurately and
consistently identify Indian vehicles across diverse environments. By centralizing the
processing power in the cloud and utilizing advanced machine learning models, the
system will provide a scalable, high-performance solution that adapts to India’s specific
needs, offering improvements in traffic monitoring, law enforcement, and public safety.

1.3 Prerequisites

Implementing the LPR system requires specific prerequisites for both setup and
operation. Essential components include high-resolution cameras capable of capturing
vehicle images under diverse lighting and environmental conditions. Cloud infrastructure
is necessary to centralize data processing and ensure accessibility, along with ample
storage for the high-volume image and video data. Moreover, machine learning
knowledge—especially in YOLO (for detection) and OCR (for character recognition)—is
crucial for achieving high recognition accuracy. The system also demands reliable
internet connectivity to enable real-time data transfer between cameras and the cloud.
Together, these prerequisites establish the foundation for a scalable, efficient LPR system.

1.4 About License Plate Detection

License plate detection is a critical step in the LPR system, involving the precise
identification and extraction of the license plate region from an image of a vehicle. Due
to various challenges—such as plate orientation, lighting, and vehicle speed—this task
requires sophisticated detection models capable of accurately isolating the license plate in

2
real time. The YOLO (You Only Look Once) model, an advanced object detection
algorithm, is particularly well-suited for this purpose due to its speed and accuracy.

YOLO operates by dividing an image into a grid, predicting bounding boxes and class
probabilities for each cell in the grid simultaneously. This "single shot" approach allows
YOLO to process images rapidly, making it ideal for real-time applications like LPR.
YOLO's architecture utilizes a convolutional neural network (CNN) to detect objects
across different scales, which is beneficial in scenarios where license plates vary in size
depending on the vehicle’s distance from the camera. YOLO has undergone several
iterations, with YOLOv5 and YOLOv6 being popular choices for current
implementations due to their increased precision and efficiency.

One of the notable strengths of YOLO in LPR systems is its ability to detect objects in
complex scenes, making it effective for multi-lane environments where vehicles move at
varying speeds and positions. YOLO’s anchor boxes are adaptable, allowing the model to
handle license plates that may be partially obstructed, angled, or positioned at different
heights. Additionally, YOLO’s speed ensures minimal latency in detection, a crucial
factor when handling high-throughput environments like toll plazas or busy intersections.

Training YOLO for license plate detection in the Indian context involves customizing the
dataset to include various Indian plate designs and environmental conditions. Data
augmentation techniques, such as random rotations, blurring, and brightness adjustments,
are often applied to simulate real-world conditions like poor lighting or motion blur. This
training process ensures that YOLO can reliably detect Indian license plates, regardless of
external conditions or regional variations in plate design.

The cloud-based integration of YOLO further enhances its application in LPR systems.
By offloading the computational demand to powerful cloud servers, the system can
handle high-resolution images and process multiple feeds simultaneously, which is vital
for large-scale deployments. Cloud-based YOLO detection also allows for centralized
updates, so improvements in detection accuracy or model efficiency can be rapidly
deployed without requiring hardware upgrades on-site.
3
In summary, YOLO’s balance of speed and accuracy makes it a suitable choice for
license plate detection in the Indian LPR system. Its adaptability to various plate
orientations, combined with the processing power of cloud infrastructure, provides a
scalable, real-time solution for accurate vehicle identification. With YOLO at its core, the
LPR system can achieve the necessary performance to enhance traffic management, law
enforcement, and public safety across India's diverse and dynamic environments.

1.5 About License Plate Recognition

Once the license plate has been detected and isolated, the next phase in the LPR system is
recognition—translating the alphanumeric characters on the plate into machine-readable
text. This process is accomplished using Optical Character Recognition (OCR), a
technology that has advanced significantly in recent years due to deep learning
innovations. In LPR systems, OCR is tasked with interpreting characters under
challenging conditions, such as variable fonts, regional languages, and degraded plate
quality.

OCR functions by processing the segmented license plate area through a sequence of
convolutional layers designed to identify and classify each character. Unlike traditional
OCR, which relies heavily on template matching, modern OCR models are typically
based on deep learning techniques, enabling them to generalize across a wide range of
plate designs and languages. This flexibility is crucial in the Indian context, where plates
may display varying fonts, languages, and even non-standard symbols in some regions.
The OCR model must therefore be robust enough to recognize diverse characters
accurately and consistently.

To train the OCR model effectively, a dataset of Indian license plates in various
conditions—day and night lighting, wear and tear, and different fonts—is required. Data
preprocessing, such as binarization, resizing, and normalization, helps ensure the OCR
model learns to recognize characters under consistent conditions, despite variations in the
raw data. The training process usually leverages architectures like Convolutional
Recurrent Neural Networks (CRNN), which combine CNN layers for feature extraction
4
with Recurrent Neural Network (RNN) layers for sequence prediction, allowing the
model to handle sequential character recognition.

The cloud-based architecture offers distinct advantages for OCR in an LPR system. By
processing OCR on centralized servers, the system gains access to high-performance
hardware, enabling faster and more accurate character recognition, especially for high-
throughput environments. This also enables real-time access and updates to the OCR
model, allowing for improvements to be rapidly deployed across all end users. As license
plate designs or language preferences evolve, the OCR model can be fine-tuned to adapt
to these changes without necessitating new hardware or system-wide modifications.
In conclusion, OCR is a fundamental component of the LPR system, translating image
data into actionable information that supports traffic management, law enforcement, and
security. With deep learning-based OCR, the system can accurately identify and interpret
a wide range of license plate characters in real-time, overcoming the unique challenges
presented by India’s diverse and variable license plate formats. Cloud integration further
enhances OCR’s efficiency, scalability, and adaptability, making the LPR system a
powerful tool for smarter, safer transportation networks across the country.

2. Literature Review

1. Early Works and Traditional Methods

- Jiang et al. (2002): Early works in LPR focused on traditional image processing
techniques. Jiang et al. proposed a method that used edge detection, morphological
operations, and template matching for license plate localization and recognition. These
methods were effective but often struggled with varying lighting conditions and complex
backgrounds.

2. Machine Learning-Based Approaches

- Anagnostopoulos et al. (2006): This comprehensive survey highlighted various machine
learning approaches for LPR, including Support Vector Machines (SVMs) and Neural

5
Networks. These methods showed improved accuracy over traditional techniques by
learning from data, but they still required significant feature engineering.

3. Convolutional Neural Networks (CNNs)

- Silva and Jung (2017): Silva and Jung utilized CNNs for end-to-end license plate
detection and recognition. Their approach demonstrated significant improvements in
accuracy and robustness by leveraging deep learning to automatically learn features from
raw images, reducing the need for manual feature extraction.

4. Real-Time LPR Systems

- Zhang et al. (2019): This work focused on real-time LPR systems using the YOLO (You
Only Look Once) framework. YOLO's ability to perform object detection quickly and
accurately made it suitable for real-time applications. Zhang et al. showed that YOLO
could effectively handle various challenges such as different lighting conditions and plate
orientations.

5. Advanced Deep Learning Models

- Li et al. (2020): Li and colleagues introduced a model using a combination of CNNs
and Recurrent Neural Networks (RNNs) for license plate detection and character
recognition. The use of RNNs allowed the model to capture sequential dependencies in
the license plate characters, improving recognition accuracy.

6. Multi-Scale and Attention Mechanisms

- Wang et al. (2021): Wang et al. incorporated multi-scale feature extraction and attention
mechanisms into their LPR model. This approach improved detection and recognition
performance, especially for license plates of different sizes and under varying conditions.
Real-time systems, such as those based on the YOLO framework, have made significant
strides in enabling practical and efficient LPR applications.

3. Scheme and Implementation of the Project

6
This section provides a detailed breakdown of the project’s methodology, from initial
motivations to the technical setup and implementation steps, highlighting how each
component contributes to solving the unique challenges of LPR in India.

3.1 Motivation: India’s diverse vehicle license plate designs, inconsistent font styles, and
environmental conditions present unique challenges for LPR technology. This project
aims to address these challenges by developing a system that can recognize license plates
in a wide variety of formats and conditions, integrating seamlessly into both urban and
rural infrastructure. The LPR system’s ability to reliably operate under varying lighting
and motion conditions will make it a valuable tool for traffic enforcement, security
monitoring, and toll collection in India.

3.2 Dataset Preparation: A dataset comprising 2000 images of Indian license plates was
initially chosen and annotated by the project team. Later-on an addition of 600 more
images having variations in brightness, angle, clarity, etc was done to the dataset making
the total to 2600 images. The data includes a broad representation of different plate
styles, fonts, and environmental conditions, such as varying lighting and weather. To
ensure compatibility with machine learning models, all images were annotated in YOLO8
format, enabling efficient use in object detection models.

3.3 Experimental Setup:

3.3.1 Hardware Details

• Name of the Hardware: Samsung RAM 16GB
Purpose of the Hardware: To manage substantial volumes of data throughout the
model training process and to provide ample memory capacity, ensuring efficient
processing and storage of the extensive datasets utilized during training.
Specification: DDR4 3200MHz
Model/ Built: M471A1K43EB1
Price (Approx): Rs. 2099

•Name of the Hardware: NvidiaGeForce GTX 1650 Graphics Card

7
 Purpose of the Hardware: Manages graphical computations essential for training
the model and efficiently handles the complex parallel processing tasks inherent
in machine learning training procedures. It effectively accelerates computation-
intensive operations, significantly reducing training time and enhancing overall
performance.
Specification: GDDR6 4GB 300MHz
Model/ Built: GTX1650
Price (Approx): Rs. 12200

3.3.2 Software Details

•Name of the Package/ Tool: Visual Studio Code
Purpose of use: Serves as the preferred integrated development environment
(IDE) for programming tasks.
License and Availability: Visual Studio Code (VS Code) is free and open-source
software. The source code for VS Code is available under the MIT License, which
allows users to freely use, modify, and distribute the software.
Source Link: https://github.com/microsoft/vscode

•Name of the Package/ Tool: YOLO( You Only Look Once)

Purpose of use: Utilized for the implementation of machine learning (ML)
algorithms and training of models.
License and Availability: YOLOv8 is Open Sourced. It is available on GitHub and
is released under the GPL-3.0 license.
Source Link: https://github.com/ultralytics/ultralytics

•Name of the Package/ Tool: OpenCV (Open Source Computer Vision Library)
Purpose of use: OpenCV is a comprehensive computer vision and machine
learning software library. It includes a wide range of functions for image
processing, video analysis, and real-time computer vision applications. It can be
used for various purposes such as object detection, facial recognition, image
stitching, camera calibration, 3D reconstruction

8
License and Availability: OpenCV is free and open-source software. The source
code for VS Code is available under the Apache2 License, which allows users to
freely use, modify, and distribute the software.
Source Link: https://github.com/opencv/opencv

•Name of the Package/ Tool: CVAT (Computer Vision Annotation Tool)

Purpose of use: CVAT is a web-based tool designed for annotating digital images
and videos. It is widely used for creating high-quality, annotated datasets for
computer vision tasks, including object detection, image segmentation, and
tracking.
License and Availability: CVAT is Open Sourced. It is available on GitHub and is
released under the MIT license.
Source Link: https://github.com/cvat-ai/cvat

•Name of the Package/Tool: AWS SageMaker

Purpose of Use: AWS SageMaker is a fully managed machine learning service
that enables developers and data scientists to build, train, and deploy machine
learning models at scale. It streamlines the end-to-end workflow for machine
learning applications.
License and Availability: AWS SageMaker is a commercial service provided by
Amazon Web Services. While the service itself is not open-source, it supports
integration with open-source tools and libraries for ML. Users pay for the
resources and features used.
Source Link: https://aws.amazon.com/sagemaker/

•Name of the Package/Tool: AWS S3

Purpose of Use: Amazon Simple Storage Service (AWS S3) provides scalable,
high-speed, and low-cost object storage. It is commonly used for data storage,
backup, archival, data lakes, and as a storage solution for applications.
License and Availability: AWS S3 is a commercial service provided by Amazon
Web Services, not open-source. It operates on a pay-as-you-go model, where
users pay for the amount of storage used and data transfer out of AWS.
9
 Source Link: https://aws.amazon.com/s3/

•Name of the Package/Tool: AWS Bedrock

Purpose of Use: AWS Bedrock is a fully managed service that allows developers
to build and scale generative AI applications. It provides access to foundation
models from multiple model providers, enabling rapid development of AI-driven
applications.
License and Availability: AWS Bedrock is a commercial service under Amazon
Web Services, offered on a pay-as-you-go basis. It is not open-source, but
integrates with open-source ML libraries and frameworks where applicable.
Source Link: https://aws.amazon.com/bedrock/

3.4 Methodology: This project involved multiple stages, including data collection,
annotation, pre-processing, and training, to develop a License Plate Recognition (LPR)
system tailored to Indian conditions. The first phase, data collection, was conducted
manually, capturing diverse images of Indian license plates in various environmental
conditions. Since India has a wide range of license plate designs, fonts, and regional
variations, it was essential to gather a comprehensive dataset that included these
differences. We ensured data diversity by collecting images from different lighting
conditions, angles, vehicle speeds, and distances, which would help the model generalize
effectively in real-world scenarios.

Following data collection, we proceeded with data annotation and pre-processing using
CVAT.ai, an open-source tool for annotating image and video data. CVAT.ai enabled us to
label each license plate accurately, defining bounding boxes around the plates and noting
other attributes necessary for effective model training. Precise annotation was critical, as
object detection models rely on these labels to learn and distinguish target objects. During
this phase, we also performed data pre-processing, addressing any inconsistencies or
noise within the dataset. Pre-processing included steps such as image resizing, color
normalization, and filtering out low-quality images. These steps ensured that the dataset

10
maintained a standard quality, helping the model learn robust features and making
training more efficient.

In the data processing and reading stage, we used the cv2 library (OpenCV) to handle and
manipulate the image data effectively. cv2 facilitated operations like reading images from
files, resizing, augmenting, and converting them to suitable formats for training. This step
was essential as it transformed raw image data into structured input suitable for the object
detection model. Data augmentation, which included techniques such as random
rotations, flips, and brightness adjustments, was also applied to make the model resilient
to variations it may encounter in real-world scenarios. Each pre-processed image was
then saved in Amazon S3, which served as our storage repository throughout the project.

Next, we moved on to model training using the YOLOv8 model, leveraging transfer
learning to fine-tune the model on our Indian license plate dataset. Transfer learning
allowed us to take advantage of a pre-trained YOLOv8 model that had already learned
generic object detection features, which we adapted to detect Indian license plates
specifically. YOLOv8 was chosen due to its high speed and accuracy in object detection,
as well as its capability to handle real-time applications. AWS SageMaker provided a
scalable platform for training the model, allowing us to utilize the necessary
computational resources without requiring local infrastructure. SageMaker’s integration
with S3 also enabled seamless access to the dataset, and its support for distributed
training helped speed up the process.

11
 Fig 3.4.1: Architecture Map

The trained YOLOv8 model demonstrated strong performance in license plate detection,
achieving high accuracy in bounding box predictions across a range of images. Once
detection was achieved, we integrated character recognition capabilities using EasyOCR,
an open-source OCR tool designed for recognizing alphanumeric characters. EasyOCR
enabled us to extract characters from the detected license plates accurately, a crucial step
for converting visual information into text for further processing. By combining
YOLOv8’s detection accuracy with EasyOCR’s character recognition capabilities, the
LPR system could not only identify license plates but also read them, supporting a full
recognition pipeline.

Fig 3.4.2: Working of EasyOCR

In summary, this project integrated various technologies and tools to develop a robust
LPR system. By leveraging manual data collection, precise annotation with CVAT.ai,

12
efficient data processing with cv2, transfer learning with YOLOv8, and character
recognition with EasyOCR, we achieved a solution capable of accurate and real-time
license plate recognition under diverse conditions. The use of AWS SageMaker for model
training and S3 for data storage streamlined the workflow, providing scalability and
reliability, making this LPR system highly adaptable for practical applications.

3.5 Usage: The Indian vehicle registration plate is structured to follow a standardized
format that provides essential information about the vehicle’s origin, making it easier for
authorities to identify, track, and manage vehicles. The High-Security Registration Plate
(HSRP) system is an advanced version designed with various security features to prevent
tampering, duplication, and theft.

Figure 3.5 1: Indian License Plate Format

State Code (XX): The first two letters on the registration plate indicate the state or Union
Territory where the vehicle is registered. Each Indian state or Union Territory has a
unique code—for instance, "MH" represents Maharashtra, "DL" stands for Delhi, "KA"
signifies Karnataka, and "TN" denotes Tamil Nadu. This two-letter state code serves as
an immediate reference to the vehicle's place of origin, allowing authorities and
individuals to recognize the state of registration quickly. For instance, in cases involving
13
interstate movement or traffic violations, identifying the state code helps in jurisdictional
processing. Additionally, it aids in data collection and analysis related to the regional
distribution of vehicles, enabling state authorities to assess traffic density, regulate
vehicle registrations, and implement region-specific policies.

District Code (88): Following the state code, the next two digits on the plate indicate the
Regional Transport Office (RTO) or district where the vehicle was registered. Each RTO
within a state is assigned a unique district code, allowing for more localized
identification. This district code helps pinpoint the exact location of the RTO that
registered the vehicle, which is valuable for tracking purposes and local law enforcement.
For example, if a vehicle is involved in an incident in a particular district, the district
code can assist authorities in quickly identifying the vehicle's origin. This feature also
supports administrative processes by helping RTOs monitor the number of vehicles in
specific regions, manage vehicle-related issues like tax collection, and improve traffic
management.

Unique Alphanumeric Code (XY8888): The last section of the registration plate consists
of an alphanumeric code that is unique to each vehicle. This part is made up of letters and
numbers and is assigned sequentially by the RTO. The alphanumeric code is the most
distinctive aspect of the plate, setting one vehicle apart from all others registered in the
same state and district. This unique identifier ensures that every registered vehicle has a
distinct identification, making it essential for tracking and distinguishing vehicles. The
alphanumeric code aids law enforcement in precise identification, especially in cases of
vehicle theft, traffic violations, or accident investigations, by helping locate the registered
owner and their details quickly.

Together, these components—state code, district code, and unique alphanumeric code—
form a standardized format that ensures uniformity across all registration plates in India.
This structure not only aids in quick and easy identification but also plays a critical role
in security and data management. The HSRP system includes additional security features,
such as tamper-resistant snap locks, which prevent the easy removal of plates, reducing

14
the risk of misuse or theft. This system supports law enforcement in efficiently tracking
vehicles, handling traffic enforcement, and enhancing crime control.

4. Results and Discussions

This section presents the performance evaluation and analysis of the License Plate
Recognition (LPR) system, which was developed using the YOLOv8 framework, combined
with OpenCV and Ultralytics libraries for detection and the popular EasyOCR for recognition
of the number plates. The results are discussed in terms of key performance metrics,
challenges encountered, and potential areas for future improvement. The confusion matrix is
attached for further references.

Fig 4.1: Confusion Matrix

15
4.1. Model Performance Metrics
The LPR system was assessed using several performance indicators, including precision,
recall, mean Average Precision (mAP), and inference time:
- Precision: The system achieved a precision of 88%, indicating a strong ability to correctly
identify license plates with minimal false positives.
- Recall: A recall rate of 95% was observed, showcasing the model’s effectiveness in
detecting a high proportion of actual license plates.
- mAP: The overall mean Average Precision was 91%, reflecting balanced precision and
recall across various test cases.
- Accuracy: The measured accuracy of the model was slightly above 92% proving its
effectiveness in real-life scenarios.
- Inference Time: The model processed frames at an average speed of 30 ms per frame on a
standard GPU, ensuring suitability for real-time operations.

4.2. Analysis of Results

The system performed consistently well under optimal conditions, such as daytime with
sufficient lighting and moderate traffic. It accurately detected and recognized license plates
across multiple vehicle types, including sedans, SUVs, and light trucks. The system
maintained high accuracy even when license plates were partially obscured or varied in size.

Fig 4.2.1: Performance Metrics

16
4.3. Challenges and Limitations:
- Lighting Conditions: Performance decreased under low-light or nighttime conditions, where
recall dropped to 85%, highlighting a need for enhanced image preprocessing or specialized
training datasets.
- Complex Backgrounds: The presence of cluttered or complex backgrounds occasionally led
to false positives, which impacted precision.
- Angles and Orientation: Plates captured at steep angles (greater than 45 degrees) were more
challenging for the model, resulting in a decrease in detection accuracy. Additional training
with augmented data representing diverse angles could improve this aspect.

5. Conclusion
In conclusion, this project has successfully developed a License Plate Recognition (LPR)
system that effectively addresses the unique demands of Indian vehicle identification.
The system demonstrates high accuracy and adaptability, handling the diversity of plate
designs, fonts, and regional inconsistencies found across India. By integrating advanced
machine learning algorithms, particularly the YOLO object detection model, the LPR
system provides robust, real-time license plate detection. The cloud implementation of
YOLO allows end users to access the system directly from the cloud, enhancing the
scalability and accessibility of the technology while minimizing local hardware
requirements. This cloud-based approach not only supports higher data processing
volumes and multi-user access but also makes it easier to maintain and update the system
as needed.

An important advancement in this project is the incorporation of Optical Character

Recognition (OCR) capabilities. The OCR module extracts and digitizes alphanumeric
characters from detected plates, enabling the system to interpret diverse fonts, text
arrangements, and even handle variations in plate conditions due to wear or weather.
OCR enhances the utility of the LPR system by providing digital conversion of license
plate information that can be integrated directly with traffic management databases and
law enforcement systems. The combination of YOLO for detection and OCR for
character recognition creates a comprehensive system that can be deployed in real-time

17
for various applications, including toll collection, traffic law enforcement, and parking
management.

Furthermore, the transition of the LPR system to a cloud-based model is a key

enhancement. The cloud platform enables the system to achieve rapid data processing and
scalability, as well as easier integration with pre-existing infrastructure. This transition
means that end users, including government agencies and private organizations, can
access a centralized, powerful LPR solution without the need for costly hardware
installations. Cloud implementation also simplifies the process of deploying updates,
such as adjustments to the OCR module for new plate designs or modifications to the
detection algorithm, ensuring the system remains adaptable to evolving requirements.

Overall, the LPR system developed in this project serves as a valuable tool in advancing
traffic management and public safety in India. It achieves the necessary balance of
accuracy, efficiency, and adaptability, addressing both technical and operational needs.
The integration of cloud-based YOLO detection, high-resolution image processing, and
OCR represents a significant leap in the capabilities of LPR systems for the Indian
context, setting a foundation for further enhancements and broader deployment in India’s
evolving transportation ecosystem.

An important advancement in this project is the incorporation of Optical Character

Recognition (OCR) capabilities. The OCR module extracts and digitizes alphanumeric
characters from detected plates, enabling the system to interpret diverse fonts, text
arrangements, and even handle variations in plate conditions due to wear or weather.
OCR enhances the utility of the LPR system by providing digital conversion of license
plate information that can be integrated directly with traffic management databases and
law enforcement systems. The combination of YOLO for detection and OCR for
character recognition creates a comprehensive system that can be deployed in real-time
for various applications, including toll collection, traffic law enforcement, and parking
management.

18
Furthermore, the transition of the LPR system to a cloud-based model is a key
enhancement. The cloud platform enables the system to achieve rapid data processing and
scalability, as well as easier integration with pre-existing infrastructure. This transition
means that end users, including government agencies and private organizations, can
access a centralized, powerful LPR solution without the need for costly hardware
installations. Cloud implementation also simplifies the process of deploying updates,
such as adjustments to the OCR module for new plate designs or modifications to the
detection algorithm, ensuring the system remains adaptable to evolving requirements.

Overall, the LPR system developed in this project serves as a valuable tool in advancing
traffic management and public safety in India. It achieves the necessary balance of
accuracy, efficiency, and adaptability, addressing both technical and operational needs.
The integration of cloud-based YOLO detection, high-resolution image processing, and
OCR represents a significant leap in the capabilities of LPR systems for the Indian
context, setting a foundation for further enhancements and broader deployment in India’s
evolving transportation ecosystem.

6. Future Scope
The future scope of this LPR project offers significant opportunities through cloud-based
deployment and mobile application development, expanding its accessibility and utility.
One primary direction is to implement the LPR system as an API on Amazon Web
Services (AWS) using SageMaker, creating a scalable and robust solution. AWS
SageMaker supports efficient model training, deployment, and management, making it
ideal for hosting the LPR model. By deploying the LPR system as an API on AWS,
various users—such as law enforcement agencies, transportation authorities, and private
organizations—can integrate LPR functionality directly into their systems without
needing specialized hardware or complex installations.

Through SageMaker, the LPR model can be continuously managed, updated, and
optimized, ensuring that users have access to the latest and most efficient version. Its
scalability allows the API to handle variable data loads and simultaneous requests,
making it suitable for real-time applications like toll management, traffic monitoring, and
19
secure facility access. AWS’s strong security framework further enhances data integrity
and confidentiality, supporting sensitive applications in public safety.

Parallel to the API, developing a mobile application, especially for Android, presents an
exciting opportunity for expanding the LPR system’s reach. This mobile app will allow
field personnel, such as traffic police and toll operators, to use the system directly from
their devices. Utilizing the device’s camera and GPS, the app will capture images, send
them to the cloud-based LPR model, and receive real-time recognition data, providing a
solution that works across diverse environments, including rural areas.

This API and mobile app setup allows the LPR system to scale with future needs, as
updates to handle new plate designs or improved processing can be deployed directly on
the cloud, requiring minimal changes to the mobile app. Together, these developments
enable a versatile, accessible LPR system that supports high-accuracy vehicle
identification, contributing to a smarter, more connected transportation ecosystem.

20
References

[1] A. Silva, L. Oliveira. (2020, February). “Automatic license plate recognition systems:
A review and comparative study.” IEEE Transactions on Intelligent Transportation
Systems [Online]. vol. 21, issue 2. Available:
https://ieeexplore.ieee.org/document/9310202

[2] R. Sharma, M. Kumar. (2024, October). “A robust approach for licence plate
detection using deep learning.” Journal of Advanced Computer Vision and Applications
[Online]. vol. 22, issue 4. Available:
https://www.researchgate.net/publication/378296867_A_Robust_Approach_for_Licence_
Plate_Detection_Using_Deep_Learning

[3] X. Li, Y. Zhang. (2020, March). “An improved vehicle detection algorithm based on
YOLOv3.” Journal of Intelligent Transportation Systems [Online]. vol. 18, issue 2.
Available:
https://www.researchgate.net/publication/340228558_An_Improved_Vehicle_Detection_
Algorithm_Based_on_YOLOV3

[4] P. Gupta, R. Mehta. (2022, December). “An end-to-end automated license plate
recognition system using YOLO-based vehicle and license plate detection with vehicle
classification.” International Journal of Computer Vision Systems [Online]. vol. 20, issue
6. Available: https://www.researchgate.net/publication/366051654_An_End-to-
End_Automated_License_Plate_Recognition_System_Using_YOLO_Based_Vehicle_an
d_License_Plate_Detection_with_Vehicle_Classification

[5] L. Ahmed, S. Rao. (2024, September). “License plate recognition-based automatic

gate opening system.” Journal of Automation and Control Engineering [Online]. vol. 25,
issue 3. Available:
https://www.researchgate.net/publication/383333866_License_Plate_Recognition-
Based_Automatic_Gate_Opening_System

[6] H. Lee, J. Park. (2023, March). “Real-time license plate recognition using deep
learning techniques: A comprehensive approach.” Sensors [Online]. vol. 23, issue 6.
Available: https://www.mdpi.com/1424-8220/23/6/2938

[7] M. Johnson, T. Patel. (2021, April). “Automatic license plate recognition using optical
character recognition.” International Journal of Machine Vision and Applications
[Online]. vol. 17, issue 4. Available:
https://www.researchgate.net/publication/350714073_Automatic_License_Plate_Recogni
tion_Using_Optical_Character_Recognition

21