TRIBHUVAN UNIVERSITY

Institute of Science and Technology

Project Report On:

“COUNTERFEIT MONEY DETECTION”

Using SVM Classification with CNN for Matching Percentage Evaluation

Submitted To:

Department of Computer Science and Information Technology

National College of Computer Studies

In partial fulfillment of the requirements of Final Year, 7th Semester for the
Bachelor’s Degree in Computer Science and Information Technology

Submitted By:

Ramila Kumari Shahi (23844/076)

Sanisha Maharjan (23851/076)

Shristi Bajracharya (23853/076)

Under Supervision of:

Mr. Yuba Raj Devkota

 Certificate of Declaration

Award Title: BSc. Computer Science and Information Technology

Declaration Sheet

(Presented in partial fulfillment of the assessment requirements for the above award)

This work or any part thereof has not previously been presented in any form to the
University or to any other institutional body whether for assessment or for other purposes.
We confirm that the intellectual content of the work is the result of our own efforts and of
no other person.

It is acknowledged that the author of any project work shall own the copyright. However,
by submitting such copyright work for assessment, the author grants the University a
perpetual royalty-free license to do all or any of those things referred to in section 16(I) of
the Copyright Designs and Patents Act 1988. (viz. to copy work; to issue copies to the
public; to perform or show or play the work in public; to broadcast the work or to make an
adaptation of the work).

Declared By:

Ramila Kumari Shahi Sanisha Maharjan Shristi Bajracharya

TU Symbol No: 23844/076 TU Symbol No.: 23851/076 TU Symbol No.: 23853/076

Date: 18th March 2024

ii
 National College of Computer Studies

Tribhuvan University

Supervisor Recommendation

I hereby recommend that this project, prepared under my supervision entitled “Counterfeit
Money Detection using SVM classification with CNN for matching percentage
evaluation”, a platform that detects counterfeit money from currency notes and provides
results, in partial fulfillment of the requirements for the degree of B.Sc. in Computer
Science and Information Technology is processed for the evaluation.

.....................................

Mr. Yuba Raj Devkota

Project Supervisor

National College of Computer Studies

Paknajol, Kathmandu

iii
 Certificate of Approval

This is to certify that the project report entitled “Counterfeit Money Detection using SVM
classification with CNN for Matching Percentage Evaluation” is a Bonafede report of
the work carried out by Ms. Ramila Kumari Shahi, Ms. Sanisha Maharjan and Ms.
Shristi Bajracharya under the guidance and supervision for the degree of B.Sc. in
Computer Science and Information Technology at National College of Computer Studies,
Tribhuvan University.

To the best of my knowledge and belief, this work embodies the work of candidates, has
duly been completed, fulfills the requirement of the ordinance relating to the Bachelor
degree of the university and is up to the standard in respect of content, presentation, and
language for being referred to the examiner.

...................................... ......................................
Mr. Yuba Raj Devkota Mr. Rajan Paudel
Project Supervisor Program Coordinator
National College of Computer Studies National College of Computer Studies

...................................... ......................................
External Examiner Internal Examiner
Tribhuvan University National College of Computer Studies

iv
 Acknowledgement

It gives us immense pleasure to express our sincere gratitude with the happiest appreciation
to all those respectable personalities who helped us to make our project work successive as
well as productive.

Our sincere appreciation goes to our mentors, teachers, and guides, without whose
unwavering support and assistance this project would not have come to fruition. We are
deeply thankful for the continuous help and support provided by the college throughout the
development of the project.

Special thanks are owed to Mr. Yuba Raj Devkota for his invaluable supervision and
assistance. His guidance and support were crucial during challenging moments, and
without him, this project would not have been possible. We are grateful for his consistent
error corrections, valuable suggestions, and guidance that proved to be instrumental
throughout the project development. Additionally, we extend our thanks to all the teachers
who provided us with the opportunity to undertake this project. Their encouragement and
support were integral to the project's growth and success.

v
 Abstract

Counterfeit money detection is paramount for maintaining financial security. A

comprehensive counterfeit money detection system is presented in this study, integrating
Support Vector Machine (SVM) classification with Convolutional Neural Network (CNN)
analysis for matching percentage evaluation. The system begins with the meticulous
curation of a dataset comprising 472 samples sourced from a reputable repository, ensuring
comprehensive coverage and relevance to the system's objectives. Following dataset
preparation, the SVM classifier undergoes rigorous training to achieve remarkable
accuracy, with an overall accuracy of 94% in distinguishing between genuine and
counterfeit currency instances during testing. Additionally, the CNN analysis evaluates
watermark and security thread similarity with high accuracy rates of 91% and 93%,
respectively. This integrated approach, supported by a detailed workflow encompassing
dataset curation, model training, and testing, underscores the system's effectiveness in
detecting counterfeit currency and contributes significantly to strengthening financial
security measures. The system is deployed using Flask in Python, enabling seamless
integration, and providing an intuitive interface for users.

Keywords: Counterfeit Money Detection, Support Vector Machine, Convolutional

Neural Network, Dataset Curation, Workflow, Accuracy, Financial Security, Flask,
Python

vi
 Table of Contents

Certificate of Declaration ..................................................................................................ii

Supervisor Recommendation .......................................................................................... iii
Certificate of Approval ..................................................................................................... iv
Acknowledgement .............................................................................................................. v
Abstract .............................................................................................................................. vi
List of Abbreviations ........................................................................................................ ix
List of Figures ..................................................................................................................... x
List of Tables ....................................................................................................................xii
Chapter 1: Introduction .................................................................................................... 1
1.1 Introduction ................................................................................................................ 1
1.2 Problem Statement ..................................................................................................... 1
1.3 Objectives ................................................................................................................... 1
1.4 Scope and Limitation ............................................................................................ 2
1.4.1 Project Scope ....................................................................................................... 2
1.4.2 Limitation ............................................................................................................ 2
1.5 Development Methodology ................................................................................... 2
1.6 Report Organization .............................................................................................. 4
Chapter 2: Background Study and Literature Review .................................................. 6
2.1 Background Study ...................................................................................................... 6
2.2 Literature Review ....................................................................................................... 6
Chapter 3: System Analysis ............................................................................................ 11
3.1 System Analysis ....................................................................................................... 11
3.1.1 Requirement Analysis........................................................................................ 11
3.1.2 Feasibility Analysis ........................................................................................ 14
3.1.3 Analysis.......................................................................................................... 16
Chapter 4: System Design ............................................................................................... 20
4.1 Design....................................................................................................................... 20
4.1.1 Database Design ................................................................................................ 20
4.1.2 Form Design ...................................................................................................... 21
4.1.3 Interface Design ................................................................................................. 22
4.2 Algorithm Details ..................................................................................................... 25
Chapter 5: Implementation and Testing........................................................................ 31

vii
 5.1 Implementation......................................................................................................... 31
5.1.1 Tools used. ......................................................................................................... 31
5.1.2 Implementation Details of Algorithm: ............................................................ 36
5.2 Testing ...................................................................................................................... 42
5.2.1 Test Cases for Unit Testing ............................................................................... 42
5.2.2 Test Cases for System Testing........................................................................... 45
5.3 Result Analysis ......................................................................................................... 47
Chapter 6: Conclusion and Future Recommendations ................................................ 56
6.1 Conclusion ................................................................................................................ 56
6.2 Future Recommendation .......................................................................................... 56
References ......................................................................................................................... 57
Appendix ........................................................................................................................... 59

viii
 List of Abbreviations

API Application Programming Interface

ATM Automated Teller Machine
CLAHE Contrast Limited Adaptive Histogram Equalization
CNN Convolutional Neural Network
CSS Cascading Style Sheets
DFD Data Flow Diagram
ER Entity-Relationship
GIMP GNU Image Manipulation Program
HSV Hue, Saturation, Value
HTML Hypertext Markup Language
HTTP Hypertext Transfer Protocol
JSON JavaScript Object Notation
LDA Linear Discriminant Analysis
MATLAB Matrix Laboratory
RGB Red, Green, Blue
ROI Region of Interest
SGD Stochastic Gradient Descent
SML Statistical Machine Learning
SQL Structured Query Language
SVM Support Vector Machine
UI User Interface

ix
 List of Figures

Figure 1. 1 Waterfall Model................................................................................................. 4

Figure 3. 1 Use Case ......................................................................................................... 11

Figure 3. 2 ER-Diagram ..................................................................................................... 16
Figure 3. 3 DFD-0 Diagram ............................................................................................... 18
Figure 3. 4 DFD-1 Diagram ............................................................................................... 19

Figure 4. 1 Database Design .............................................................................................. 20

Figure 4. 2 Form Design .................................................................................................... 21
Figure 4. 3 Home Design ................................................................................................... 22
Figure 4. 4 Login Design ................................................................................................... 22
Figure 4. 5 Sign-Up Design ............................................................................................... 23
Figure 4. 6 Select Image .................................................................................................... 23
Figure 4. 7 Result ............................................................................................................... 24

Figure 5. 1 Custom SVM ................................................................................................... 37

Figure 5. 2 CustomStandardScalar .................................................................................... 38
Figure 5. 3 Model Saving................................................................................................... 38
Figure 5. 4 Watermark Detection....................................................................................... 39
Figure 5. 5 Security Thread Detection ............................................................................... 40
Figure 5. 6 Custom CNN model ........................................................................................ 41
Figure 5. 7 Training Custom CNN .................................................................................... 42
Figure 5. 8 Fake image....................................................................................................... 46
Figure 5. 9 Real Image ....................................................................................................... 46
Figure 5. 10 SVM Training and Testing Accuracy............................................................ 48
Figure 5. 11 SVM Training and Testing Loss ................................................................... 48
Figure 5. 12 Watermark Training and Validation Accuracy ............................................. 49
Figure 5. 13 Watermark Training and Validation Loss ..................................................... 49
Figure 5. 14 Security Thread Training and Validation Accuracy ...................................... 50
Figure 5. 15 Security Thread Training and Validation Loss ............................................. 50
Figure 5. 16 SVM: Accuracy for Training Data Table ...................................................... 51
Figure 5. 17 SVM: Accuracy for Testing Data Table ........................................................ 52
Figure 5. 18 CNN (Watermark): Accuracy for Training Data Table ................................ 52

x
Figure 5. 19 CNN (Watermark): Accuracy for Validation Data Table ............................. 53
Figure 5. 20 CNN (Security Thread): Accuracy for Training Data Table ......................... 53
Figure 5. 21 CNN (Security Thread): Accuracy for Validation Data Table ...................... 53
Figure 5. 22 SVM: Confusion Matrix ................................................................................ 54
Figure 5. 23 CNN (Watermark): Confusion Matrix .......................................................... 55
Figure 5. 24 CNN (Security Thread): Confusion Matrix ................................................... 55

Figure 7. 1 Home Page....................................................................................................... 59

Figure 7. 2 Login Page ....................................................................................................... 60
Figure 7. 3 Signup Page ..................................................................................................... 60
Figure 7. 4 Image with 0.7 threshold ................................................................................. 61
Figure 7. 5 Matched Percentage......................................................................................... 61
Figure 7. 6 Image with 0.9 Threshold ................................................................................ 62
Figure 7. 7 Fake image input ............................................................................................. 62

xi
 List of Tables
Table 3. 1 Schedule Plan .................................................................................................... 15

Table 5. 1 Test case For Registration................................................................................. 43

Table 5. 2 Test Case for User Signup ................................................................................ 43
Table 5. 3 Test for Feature Extraction ............................................................................... 44
Table 5. 4 Test Case for Model Accuracy ......................................................................... 45
Table 5. 5 System Testing .................................................................................................. 45

xii
 Chapter 1: Introduction

1.1 Introduction

Counterfeit Money or Fake money is any currency (paper note or coin) produced outside
the legal sanction of a state or government to imitate the currency to deceive its recipient.
The business of counterfeiting money is as old as the history of money itself. Today, the
finest counterfeit banknotes are called super dollar/super currency because of high quality
and imitation of real currency.

Out of many security features of the bank note such as watermark, see- through register,
security thread, emboss, visually impaired feature, our fake money detection system will
use watermark feature and security thread as its main tool to identify the fake currency. By
employing image processing techniques, the system focuses on detecting the watermark
and security thread in genuine currency notes to distinguish them from fake ones. This
approach aims to offer a dependable and effective solution for identifying counterfeit
banknotes, emphasizing the importance of watermark and security thread analysis in the
detection process.

1.2 Problem Statement

In Nepal, counterfeit money fraud is on rise. The past three-month data (from Nepal News)
shows that over 1 million worth counterfeit notes were confiscated, many more are still
hidden and unidentified. The Bank and financial institutions, being a mediator to circulate
the money around the country, is the main targeted area for the people possessing the fake
currency. So, development of such a system and implementation in banks and financial
institutions will decrease the flow of the fake currency in the market. This system will help
to identify the fake currency beforehand and block the flow of the fake notes.

1.3 Objectives

The main objectives of this project are as follows:

• To detect the watermark engraved in notes to find the genuine ones.

• To detect the security thread engraved in notes.

1
1.4 Scope and Limitation

1.4.1 Project Scope

This section of our project explains the scope of our project, how we plan to deliver it and
make it effective. It is a system with an added functionality of currency detection.

1. This system focuses on the improvement and implementation of fake currency

detection.
2. It determines whether notes are genuine or counterfeit.

1.4.2 Limitation

• The picture must be stored on your device.

• Inadequate lighting hampers accurately identify features.

• Banknotes that show signs of tear or aging can occasionally be mistakenly identified
as counterfeit.

1.5 Development Methodology

In the project, Waterfall method has been used as the development model. This project is a
structured and systematic approach, making it suitable when requirements are well-defined,
and changes are expected to be minimal. Each phase must be completed before moving on
to the next, ensuring a clear progression in the development process. modify this paragraph.

Waterfall model follows the following processes:

1. Requirement Analysis:
In this phase, the requirements for counterfeit money detection are thoroughly
defined. This includes specifying the types of currency to be detected and analyzing
the security features that need to be identified.

2. System Design:
Building upon the gathered requirements, the system design phase entails creating
a comprehensive blueprint or architecture for the counterfeit money detection

2
 system. This encompasses defining modules, data structures, interfaces, and
algorithms crucial for accurate detection.

3. Implementation (Coding) Phase:

The actual coding or programming of the counterfeit money detection system
occurs in this phase. Developers utilize the design specifications to write code for
each module or component aimed at effectively identifying counterfeit currency.

4. Testing Phase:
The testing phase is dedicated to ensuring that the individual components or
modules function correctly when integrated into the entire counterfeit money
detection system. Various testing methods, including unit testing and integration
testing, are employed to validate the system's functionality.

5. Validation (System Testing) Phase:

The entire counterfeit money detection system undergoes thorough testing to verify
its compliance with the specified requirements. This phase focuses on evaluating
the overall functionality and performance to ensure accurate and reliable counterfeit
detection.

6. Deployment (Implementation) Phase:

Upon successful completion of testing and validation, the counterfeit money
detection system is deployed to the operational environment. This often involves
installing the system on designated hardware and making it accessible to end-users.

7. Maintenance Phase:
The maintenance phase involves providing ongoing support and updates to the
counterfeit money detection system. Any identified issues or bugs are addressed
promptly, and updates or enhancements may be implemented based on user
feedback or evolving counterfeit threats.

3
 Figure 1. 1 Waterfall Model

Reason for Selection of Waterfall model as the development methodology

Since the project requirements are well-defined, stable and not expected to undergo
significant changes. The Waterfall Model was selected for its ability to provide a controlled
and predictable development process, catering to the specific characteristics and
requirements of the counterfeit money detection project.

1.6 Report Organization

The report on ‘Counterfeit Money Detection’ is structured into 6 chapter, which include:

Chapter 1: Introduction

The chapter initiates with an introduction to the comprehensive project, systematically

partitioned into six distinct subchapters. Within this section, the subchapter addresses
components, including the problem statement, objectives, scope, limitations, and the
chosen development methodology employed in the project.

Chapter 2: Background study and literature Review

It includes the study of fundamental theories, general concept and literature review i.e.,
existing system, people’s preference etc. of the present Counterfeit money detection.

4
Chapter 3: System Analysis

This chapter includes requirement analysis and feasibility of the system. It also consists of
ER-diagram and DFD-diagram.

Chapter 4: System Design

It consists of the overall design of the system and the algorithm used.

Chapter 5: Implementation and Testing

It provides the detail tool we used while developing the system. It also includes required
testing with different test as per requirement.

Chapter 6: Conclusion and Future Recommendation

It summarizes the key findings, insights, and outcomes derived from the research or project
and future suggestions for potential actions, improvements, or areas of further exploration
based on the conclusions drawn.

5
 Chapter 2: Background Study and Literature Review

2.1 Background Study

The study of fake money detection involves a multidimensional approach encompassing

physical conditions, image processing techniques, and feature extraction methods. Various
research papers delve into the development of systems for detecting counterfeit currency,
particularly focusing on Indian currency notes. These systems utilize computer vision-
based approaches, image processing, and feature extraction to distinguish between genuine
and fake currency. Techniques like Radiometric and Geometric corrections are employed
to rectify spectral errors and distortions in currency images.[1] The use of MATLAB for
feature extraction, image segmentation, and decision-making processes is highlighted in
these studies. Additionally, advancements in technology, such as the application of
Generative Adversarial Networks (GANs) for counterfeit money detection, showcase the
continuous evolution of methods to combat counterfeiting in financial systems.

2.2 Literature Review

ANALYSIS OF COUNTERFIET CURRENCYB DETECTION TECHINQUES FOR

CLASSIFICATION MODEL (Akansha Upadhyaya, Dr. Vinod Shokeen): [2] Counterfeit
currency is one of the threats which creates vice to nation’s economy and hence impacts
the growth worldwide. Producing forged currency or fabricating fake features in the
currency considered to be a crime. Currency crime comes under the criminal law and
known to be as Economical crime. Over the past few years many researchers have proposed
various techniques to identify and detect forged currency. A serious problem has been come
up with variety of solutions in terms of hardware related techniques, Image processing and
machine learning methods. Advancements in printing and scanning technology, trading of
material are some of the problems in germinating counterfeit currency. The study presents
various fake currency detection techniques proposed by various researchers. The review
highlighted the methodology implemented on characteristics feature with success rate of
each method to detect counterfeited currency. Moreover, the study includes the analysis of
widely acceptable statistical classification techniques for currency authentication. The
comparative analysis of Logistic Regression and Linear Discriminant Analysis (LDA) was
performed to realize the better model for currency authentication. It has been found that

6
classification Model using Logistic regression shows better accuracy of 99identifying most
feasible technique to be implemented based on the accuracy rate REVIEW ON
DETECTION OF

FAKECURRENCY USING IMAGE PROCESSING TECHNIQUES (Dr. S.V.

Virakthamath, Kshama Tallur [3]: Many countries are affected by the matter of fake notes.
Indian is one among them. With improved technology, anyone can print fake notes. These
notes are produced without legal sanction of the state and continued production of such
kinds of notes can degrade countries economy. When such counterfeited notes are produced
and circulated, it becomes impossible for ordinary citizens to distinguish whether the
money is real or fake because they differentiate according to physical appearance. The
biggest challenge for many countries like India is the detection of fake currency. Even if
banks and other big organizations have automatic machines designed to identify counterfeit
currency notes, ordinary people can hardly differentiate between them. Nowadays
recognition of fake currency has become a challenging issue for many researchers. The
identification involves many steps like edge detection, feature extraction, image
segmentation, image acquisition, gray scale conversion, and comparison of images. This
paper provides some related works of paper currency recognition and has explained the
spread of various currency recognition systems. Choosing the right feature would improve
overall system performance. The goal of this work is to review previous papers and
literature, identify the benefits and disadvantages of every method.

A HYBRID FAKE BANK NOTE DETECTION MODEL USING OCR, FACE

RECOGNITION AND HOUGH FEATURES (Abida zarin, Jia Uddin) [4]: Currency
duplication remains an emergent concern among nations due to the advancement of printing
and scanning technology. Many note detection systems are present in banks but they are
very costly and often inaccurate. The fields of image processing, neural network and
machine vision have potential to significantly overcome this issue. This paper proposes a
model comprised of Optical Character recognition (OCR), Face Recognition and Hough
transformation algorithm. The microprinting, watermark, and ultraviolet lines features of
Bangladeshi notes are extracted for testing of genuine notes. The experimental results of
the proposed model give accuracy as high as 93.33mobile application. Moreover, the
obtained results are compared with the output from individual algorithm of OCR, Face
Recognition and Hough transformation, to show that the proposed algorithm gives the
highest accuracy.

7
AUTOMATIC RECOGNITION OF FAKE INDIAN CURRENCY NOTE (Sonali R.
Darade, Prof. G.R. Gadveer) [5]: In this paper, the automatic system is designed for
identification of Indian currency notes and check whether it is fake or original. The
automatic system is very useful in banking systems and other field also. In India increase
in the counterfeit currency notes of 100, 500 and 1000 rupees. As increase in the technology
like scanning, color printing and duplicating, because of that there is increase in counterfeit
problem. In this paper, recognition of fake Indian currency notes is done by using image
processing technique. In this paper, recognition of fake Indian currency notes is done by
using image processing technique. In this technique first the image acquisition is done and
applies preprocessing to the image. In pre-processing crop, smooth and adjust then convert
the image into grey color after conversion apply the image segmentation then extract
features and reduce, finally comparing image.

EVALUATON OF MACHINE LEARNING ALGORITHMS FOR DETECTION OF

FAKE BANK CURRENCY (Anuj Yadav, Tarun Jain) [6]: The one important asset of our
country is Bank currency and to create discrepancies of money miscreants introduce the
fake notes which resembles to original note in the financial market. During demonetization
time it is seen that so much of fake currency is floating in the market. In general, for a
human being, it is very difficult to identify forged note from the genuine not instead of
various parameters designed for identification as many features of forged note are like
original one. To discriminate between fake bank currency and original note is a challenging
task. So, there must be an automated system that will be available in banks or in ATM
machines. To design such an automated system there is need to design an efficient
algorithm which can predict whether the banknote is genuine or forged bank currency as
fake notes are with high precision. In this paper six supervised machine learning algorithms
are applied on dataset available on UCI machine learning repository for detection of Bank
currency authentication. To implement this, we have applied Support Vector machine,
Random Forest, Logistic Regression, Naive Bayes, Decision Tree, K- Nearest Neighbor by
considering three train test ratio 80:20, 70:30 and 60:40 and measured their performance
on the basis various quantitative analysis parameter like Precision, Accuracy, Recall, MCC,
F1-Score and others. And some of SML algorithm are giving 100 train test ratios.

FAKE CURRENCY DETECTION USING IMAGE PROCESSING AND MACHINE

LEARNING (S. Gothe, K. Naik, V. Joshi) [7]: With the increase in the technology, the
convenience and ease of people to carry out various task is increasing on a large scale. But

8
with the advancement in technology, the amount of crime carried out due to wrong use of
these technologies is also increasing on a large scale. Similar thing 7applies to the currency
notes being handled by us on day to day basis. Fake currency notes are made so accurate
that finding out which one is real note and which one is fake becoming increasingly
difficult. Fake currency note is the imitation of the authentic currency note for wrong
purposes and without the permission of the state and central government. Hence with the
advancement in the development of fake currency notes, the detection mechanism needs to
be developed as well to reduce its flow in the market. Fake currency notes have led to
reduction in the value of money and also loss on economic and social front. In the paper
we are using Image Processing and Machine Learning to check the authenticity of the
currency note. An android application will help people to easily detect the fake note as
many people today carry smart phones and hence android application is not a difficult thing
for people to handle. This led us to satisfy our purpose of our application being helpful to
common citizen of India. Key words Fake Currency Detection, Counterfeit Detection

SUPPORT VECTOR MACHINE IN THE ANTICIPATION OF CURRENCY

MARKETS. Int. J. Eng. Technol. (UAE) 7(2), 66–68 (V. Lalithendra Nadh, G. Syam
Prasad) [8]: Various researchers have done an expansive research within the domain of
stock market anticipation. The majority of the anticipated models is confronting some
pivotal troubles because of the likelihood of the market. Numerous normal models are
accurate when the data is linear. In any case, the expectation in view of nonlinear data could
be a testing movement. From past twenty years with the progression of innovation and the
artificial intelligence, including machine learning approaches like a Support Vector
Machine it becomes conceivable to estimate in light of nonlinear data. Modern researchers
are combining GA (Genetic Algorithm) with SVM to achieve highly precise outcomes.
This analysis compares the SVM and ESVM with other conventional models and other
machine learning methods in the domain of currency market prediction. Finally, the
consequence of SVM when compared with different models it is demonstrated that SVM
is the premier for foreseeing.

REAL TIME FAKE CURRENCY NOTE DETECTION USING DEEP LEARNING. Int.
J. Eng.Adv. Technol. (IJEAT) 9(1S5) (M. Laavanya, V. Vijayaraghavan). ISSN: 2249-
8958 [9]: Great technological advancement in printing and scanning industry made

9
counterfeiting problem to grow more vigorously. As a result, counterfeit currency affects
the economy and reduces the value of original money. Thus it is most needed to detect the
fake currency. Most of the former methods are based on hardware and image processing
techniques. Finding counterfeit currencies with the methods is less efficient and time
consuming. To overcome the above problem, we have proposed the detection of counterfeit
currency using a deep convolution neural network. Our work identifies the fake currency
by examining the currency images. The transfer learned convolutional neural network is
trained with two thousand, five hundred, two hundred and fifty Indian currency note
datasets to learn the feature map of the currencies. Once the feature map is learnt the
network is ready for identifying the fake currency in real time. The proposed approach
efficiently identifies the forgery currencies of 2000, 500, 200, and 50 with less time
consumption.

FAKE CURRENCY DETECTION USING IMAGE PROCESSING,ICSET-

2017.IOPConf.Ser.Mater.Sci.Eng. 263, 052047 (T.Agasti,G. Burand, P.Wade, P.Chitra)
[10]: The advancement of color printing technology has increased the rate of fake currency
note printing and duplicating the notes on a very large scale. Few years back, the printing
could be done in a print house, but now anyone can print a currency note with maximum
accuracy using a simple laser printer. As a result the issue of fake notes instead of the
genuine ones has been increased very largely. India has been unfortunately cursed with the
problems like corruption and black money. And counterfeit of currency notes is also a big
problem to it. This leads to design of a system that detects the fake currency note in a less
time and in a more efficient manner. The proposed system gives an approach to verify the
Indian currency notes. Verification of currency note is done by the concepts of image
processing. This article describes extraction of various features of Indian currency notes.
MATLAB software is used to extract the features of the note. The proposed system has got
advantages like simplicity and high performance speed. The result will predict whether the
currency note is fake or not.

10
 Chapter 3: System Analysis

3.1 System Analysis

The system analysis of the counterfeit money detection system involves defining
objectives, gathering requirements, selecting appropriate technologies, implementing
feature extraction techniques, training CNN models, and designing a user-friendly interface
to accurately detect counterfeit currency notes and enhance financial security measures.[11]

3.1.1 Requirement Analysis

The requirement analysis for the counterfeit money detection system encompasses
both functional and non-functional aspects essential for accurate and efficient
detection of counterfeit currency.
i. Functional Requirement
Use Case Diagram:

Figure 3. 1 Use Case

Use Cases:
a. Login:
• Actor: User

11
 • Description: The user logs into the system using valid credentials (username
and password).
b. Registration:
• Actor: User
• Description: New users can register for the system by providing necessary
information and creating a username and password.
c. Upload:
• Actor: User
• Description: Users can upload image of currency note for analysis. The
uploaded image is then processed by the system.
d. Counterfeit Check:
• Actor: System
• Description: The system analyzes the uploaded image to determine if the
currency notes are genuine or counterfeit. This involves using image
processing and counterfeit detection algorithms.
e. Display Result:
• Actor: System
• Description: After the counterfeit check is performed, the system displays
the results to the user. It indicates whether the uploaded currency note is
genuine or potentially counterfeit. The system also displays the matching
percentage of watermark and security thread of the uploaded currency note.

ii. Non-Functional Requirement

The non-functional requirements of a counterfeit money detection system encompass

aspects beyond the system's specific functionalities. Some non-functional requirements
that could be considered for such a system area as follows:

• Performance:
The system should be able to process currency images quickly and accurately
to detect counterfeit notes efficiently.

12
 It should have a high level of accuracy in distinguishing between genuine and
fake currency.
• Reliability:
The system should be reliable, ensuring consistent and accurate detection
results.
It should be robust enough to handle variations in currency notes and
environmental conditions.
• Usability:
The system should have a user-friendly interface that is easy to navigate for
users checking the authenticity of currency notes.
It should provide clear and understandable results to users.
• Security:
The system should maintain the confidentiality and integrity of the data
processed, especially when dealing with sensitive financial information.
It should have measures in place to prevent unauthorized access to the system.
• Scalability:
The system should be scalable to accommodate an increasing number of users
and currency checks without compromising performance.
It should be able to handle a growing dataset of currency images for training
and testing.
• Maintainability:
The system should be easy to maintain and update with new counterfeit
detection techniques or features.
It should have clear documentation and modular design for ease of
maintenance.

These non-functional requirements are essential for ensuring the overall effectiveness,
usability, and reliability of a counterfeit money detection system.

13
3.1.2 Feasibility Analysis

i Technical Feasibility
The technical feasibility of the project is high, as the required hardware and
software resources are widely available and accessible. Requirements of our system
can be categorized as:

Hardware Requirements:
• A computer with a minimum of 4 GB of RAM and a multi-core processor
• Sufficient RAM for storing and processing image data during the
counterfeit detection process.
• Storage devices for maintaining a database of known currency features and
patterns for comparison during counterfeit detection.

Software Requirements:
• Operating System: Windows 10, Linux, or MacOS
• Web Browser: Chrome, Firefox, or Safari
• Python 3.6 or above: This is required to run the Flask web framework.
• Flask web framework: This is required to build the web application and API for
the system.
• ReactJS: This is required to build the frontend of the web application and to
create a responsive user interface.
• Git and GitHub: These are required for version control and collaboration among
the development team.
In addition to the above requirements, some optional software may be used for
development such as virtual environments like Anaconda to manage dependencies.

ii Operational Feasibility
The operational feasibility of the project is moderate, as the project requires the
development of a reliable and accurate counterfeit money classifier. The operational
feasibility of implementing a counterfeit money detection system hinge on its
seamless integration with existing processes, compatibility with current
infrastructure, and user acceptance. Regulatory compliance, scalability, reliability,
and minimal operational impact are also key considerations. The system's

14
 maintenance requirements, risk management, and customer experience implications
further contribute to determining its operational viability. Ultimately, a thorough
evaluation of these factors is essential to ascertain the practicality and effectiveness
of a counterfeit money detection system.

iii Economic Feasibility

The economic feasibility of the project is high, as the project does not require any
significant upfront investment or ongoing operational costs. The required hardware
and software resources are typically available on most computers and can be
acquired at a relatively low cost. In addition, the development can use open-source
software and libraries to reduce the cost of development.

iv Schedule Feasibility
The time given for the completion of this project was a whole semester. So, the
project had enough time for completion. Since the project has some machine
learning mechanisms the project took some more time than done usually. Hence,
the project has been developed according to the following time schedule to make
our application schedule feasible.

Table 3. 1 Schedule Plan

S. No. Task Name Duration Start Date End Date

1 Planning 6 Days 22 Oct, 2023 28 Oct, 2023
2 Analysis 8 Days 29 Oct, 2023 11 Nov, 2023
3 Design 12 Days 12 Dec, 2023 26 Dec, 2023
4 Coding 30 Days 26 Dec,2023 14 Feb, 2024
5 Testing 10 Days 15 Feb, 2024 28 Feb, 2024
6 Implementation 6 Days 28 Feb, 2024 11 Feb, 2024
7 Documentation 90 Days 12 Nov, 2023 11 Feb, 2024

15
3.1.3 Analysis

Data Modeling using ER-diagram.

Figure 3. 2 ER-Diagram

In the ER (Entity-Relationship) diagram for a counterfeit money detection system, three

key entities are represented: User, Currency Image, and Detection Result.

User Entity:

The "User" entity is a fundamental component of the counterfeit money detection system,
representing individuals who engage with the system. It is characterized by four main
attributes:

• user_id: This attribute serves as a unique identifier for each user in the system,
ensuring individuality and facilitating relational links.
16
 • username: The username attribute represents the chosen identifier or display name
selected by the user during the registration process.
• email: This attribute stores the user's email address, providing a means of
communication and contact.
• Password: The password attribute stores the secure access key chosen by the user
during registration, ensuring the confidentiality of their account.

Currency Image:

The "Currency Image" entity is a pivotal element in the counterfeit money detection system,
capturing the images submitted by users for analysis. This entity is characterized by two
crucial attributes:

• image_id: This is the attribute where the uploaded image identifier is stored.
• Upload_Date_Time: This attribute represents the timestamp when the currency
image was uploaded to the system. It serves as an essential chronological reference,
aiding in tracking and organizing the images based on their submission time.
• image_data: The image_data attribute stores the binary data or reference to the
actual image file. This includes the content of the currency image uploaded by the
user. It may involve the image file itself or a link/reference pointing to the stored
image data.

Detection Result Entity:

The "Detection Result" entity is a critical component of the counterfeit money detection
system, recording the outcomes of the analysis performed by the system. This entity is
characterized by several key attributes:

• result_id: This attribute serves as a unique identifier for each detection result. It
ensures individuality and allows for easy referencing and tracking of results.
• Prediction Date Time: The prediction_date attribute indicates the timestamp when
the system performed the counterfeit detection analysis. It provides chronological
information, allowing users to understand when the analysis was conducted.
• Watermark Similarity: This attribute represents the degree of similarity between
the uploaded currency image and the expected watermark of genuine currency. It

17
 quantifies the likeness and contributes to determining the authenticity of the
currency note.
• Security Thread Similarity: This attribute measures the resemblance of the
uploaded currency image to the expected characteristics of a genuine currency note,
specifically focusing on features like security features or metallic elements.
• Pred_result: This is the attribute where the result of the uploaded image is stored,
indicating whether it is classified as real or fake.
• Image: This attribute stores the uploaded image in the form of an image blob.

Process Modeling using DFD

DFD-0 Level Diagram

Figure 3. 3 DFD-0 Diagram

In the DFD-0 level, the primary entities include the "user" and the "counterfeit money
detection" process. Users interact with the system through dataflows like "upload image,"
"login," and "register," all directed towards the "counterfeit money detection" process. The
"upload image" dataflow enables users to submit currency images for analysis, while
"login" and "register" processes handle user authentication and account creation,
respectively. The "counterfeit money detection" process processes the uploaded images and
generates predicted results based on the analysis. The outcomes, specifically "predicted
result" and "authentication status," are then communicated back to the user. This DFD-0
level diagram illustrates the essential interactions and dataflows between the user entity
and the counterfeit money detection process, capturing the key components for image
submission, user authentication, and result retrieval.

18
DFD-1 level Diagram

Figure 3. 4 DFD-1 Diagram

In the DFD-1 diagram, a singular user entity interacts with key processes within the
counterfeit money detection system. User authentication is handled through the "user login"
process, while the "receive image" process manages image submissions. Subsequent
processes include "image processing" for quality enhancement, "feature extraction and
analysis" for authenticity evaluation, and the central "counterfeit detection and decision"
process. Three critical data stores— "user table," "currency table," and "prediction"— store
essential information, streamlining interactions and supporting informed decision-making
in the system. This concise breakdown in the DFD-1 diagram provides a clear snapshot of
subprocesses and data management, facilitating efficient system understanding and design
considerations.
19
 Chapter 4: System Design

4.1 Design

4.1.1 Database Design

Figure 4. 1 Database Design

The diagram above shows an entity-Relationship (ER) diagram with three tables: ‘user’,
‘Currency image’ and ‘Detection Result’. It illustrates the transformation of an ER model
into a relational schema:

• Each entity (user, Currency image and Detection Result) is a table.

• Attributes of entities become table column.
• ‘Id’ in all tables is a primary key(pk).
• Currency image has foreign key as ‘user_id’ and detection result has foreign key as
‘image_id’

20
Normalizations

1. First Normal Form (1NF):

• Already in 1Nf, as each column contains atomic(indivisible) values.
• No repeating groups or arrays; each column contains atomic values.
2. Second Normal Form (2NF):
• Already in 2NF since there is no composite primary key.
• The ‘user_id’ is a foreign key, and there is no partial dependency on the
primary key.
3. Third Normal Form (3NF):
• Already in 3NF, science there are no transitive dependencies.
• No transitive dependencies, each non-prime attribute depends on the
primary key.

4.1.2 Form Design

Based on our project, forms were designed for customer.

1. User Login & User Sign-Up Form: Allow user to register or log into the platform.
2. Contact-Form: Allows user to reach out the platform.

Figure 4. 2 Form Design

21
4.1.3 Interface Design

Interface design, often referred to as user interface (UI) design, focuses on creating visual
elements and interactions that enable users to interact with a system or application. A well-
designed interface should be intuitive, visually appealing, and user-friendly.

Figure 4. 3 Home Design

Figure 4. 4 Login Design

22
Figure 4. 5 Sign-Up Design

Figure 4. 6 Select Image

23
Figure 4. 7 Result

24
4.2 Algorithm Details

SVM for Classification

SVM stands for Support Vector Machine, which is a popular type of machine learning
algorithm used for classification and regression analysis.

The steps for implementing SVM are as follows:

1. Load Images and Extract Features:

The code loads images from a specified folder using the load_images_from_folder
function. Each image undergoes feature extraction in the extract_features function,
involving resizing, grayscale conversion, and contrast enhancement. The processed
image is further manipulated by denoising, darkening, and region extraction.
Watermark and Security Thread templates are loaded to detect their presence in the
designated regions. This series of steps forms the initial preprocessing stage for
counterfeit money detection.

2. Preprocess and Save Images:

The preprocess_and_save function utilizes the extract_features function to
preprocess and save images. Images are categorized as real or fake based on the
presence of a watermark and security thread.

3. Load and Preprocess Images for SVM:

Load real and fake images from the specified folders. Preprocess and save the
images, maintaining counts for real and fake images.

4. Custom Train-Test Split:

The custom_train_test_split function shuffles and divides the data into separate
training and testing sets. It first shuffles the data randomly and then allocates 20%
of it for testing and 80% for training, according to the specified test_size.
5. Custom Standard Scaler:
The CustomStandardScaler class is a part of the feature scaling process. Feature
scaling is a crucial preprocessing step in machine learning, and it involves
transforming the input features so that they have similar scales. This ensures that no
feature dominates the learning process due to having a larger magnitude.

25
 The formula for standardization is as follows:
𝑂𝑟𝑔𝑖𝑛𝑎𝑙 𝑉𝑎𝑙𝑢𝑒 − 𝑀𝑒𝑎𝑛
𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑 𝑉𝑎𝑙𝑢𝑒 =
𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛

6. Custom SVM Model:

The CustomSVM class is defined for a custom SVM model. The fit method trains
the model using custom gradient descent. The predict method makes predictions
based on the learned weights and bias. The plot_metrics function is a part of the
CustomSVM class, and it is responsible for visualizing the training and testing
accuracy along with the training and testing loss over the epochs.
a. Prediction (Linear SVM):
The prediction is made in the predict method of the CustomSVM class. The
dot product of weights (w) and the input features (X) is computed, and the
sign of the result determines the predicted class (1 or 0).
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 = 𝑠𝑖𝑔𝑛(𝜔. 𝑋 + 𝑏)
b. Update Rule (Gradient Descent):
The weights (𝜔) and bias (b) are updated during training using gradient
descent. The update rule for a binary classification SVM is as follows:
𝜔𝑖 = 𝜔𝑖 + 𝑛. (𝑦 − 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛)
c. Loss (Squared Error Loss):
The squared error loss is used to measure the error during training.
1
𝐿𝑜𝑠𝑠 = (𝑦 − 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛)2
2

7. Load and Standardize Features:

Load all preprocessed images and extract features. Create labels based on whether
the image is real or fake.

8. Evaluate Training and Testing Accuracy:

Make predictions on both training and testing sets. Calculate and print the accuracy
on both sets.

26
 9. Calculate the Confusion Matrix for testing:
Confusion matrix is calculated using the confusion_matrix function. It takes two
arguments: test_labels, which represent the true labels of the test data, and
test_predictions, which represent the predicted labels of the test data.

10. Calculate the Classification Report:

Classification reports generate both the training and testing sets using the
`classification_report` function. For the training set, it calculates a classification
report based on the true labels train_labels and the predicted labels
train_predictions. Similarly, for the testing set, it computes a classification report
using the true labels test_labels and the predicted labels test_predictions, with the
same specified target_names[‘fake’,’real’].

Convolution Neural Network (CNN) for matching percentage

Convolutional Neural Networks (CNNs) are a class of deep learning algorithms specifically
designed for processing and analyzing visual data, such as images and videos. [12] Here's
an overview of the key components and operations involved in a matching percentage with
CNN:

1. Define the CNN model:

The custom CNN model is defined using the Keras Sequential API. The model
architecture includes convolutional layers (Conv2D), max-pooling layers
(MaxPooling2D), a global average pooling layer (GlobalAveragePooling2D), and
a dense layer with softmax activation. The model is defined in the custom_model
function.
a. Convolutional Layer (Conv2D):
The convolutional layer applies filters to the input image to extract features.
Each filter is convolved with the input, and the result is passed through an
activation function. Formula:

𝑂𝑢𝑡𝑝𝑢𝑡 = 𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛(∑(𝑖𝑛𝑝𝑢𝑡 × 𝑓𝑖𝑙𝑡𝑒𝑟) + 𝑏𝑖𝑎𝑠)

27
 b. Max-pooling layer (MaxPooling2D):
Max-pooling reduces the spatial dimensions of the feature maps, retaining
the most important information and decreasing computational complexity.
c. Global Average Pooling (GlobalAverageePooling2D):
Global Average Pooling is applied globally across the entire feature map.
Instead of dividing the input into local regions and computing the maximum
or average within each region, Global Average Pooling computes the
average of each feature map across its entire spatial dimensions. The result
is a single value per feature map.
1 𝐻 𝑊
𝐺𝐴𝑃(𝑋𝑖 ) = ∑ ∑ 𝑋𝑖 (ℎ, 𝜔)
𝐻×𝑊 ℎ=1 𝜔=1

d. Fully Connected (Dense) Layers:

Fully connected layers process the flattened features and make predictions.
These layers learn to combine features and capture global patterns.
𝑍 = 𝑋. 𝑊 + 𝑏

Where, X= input matrix, W= Weight matrix, b= bias vector

e. Output Layer:
The output layer produces the final prediction. For binary classification
(counterfeit money detection) multi-class classification, a softmax
activation is used as there is two class Real and Fake. The softmax function
ensures that the sum of all our output probability values will always be equal
to one so that one can easily see which classified choice has the highest
probability of being the right one.
Where, zi represents raw score for class i
N is the total number of classes

2. Preprocess Input Images:

Images from the input folder are preprocessed, including resizing, grayscale
conversion, histogram equalization (CLAHE), denoising, and darkening.
The preprocess_input function is responsible for these preprocessing steps.
𝑔𝑟𝑎𝑦𝑣𝑎𝑙𝑢𝑒 = 0.299 ∗ 𝑅 + 0.587 ∗ 𝐺 + −0.114 ∗ 𝐵

28
 This formula calculates the weighted average of the RGB (Red, Green, Blue) values
to obtain a single grayscale value.

3. Detect Watermark:
The detect_watermark function is used to detect a watermark in the preprocessed
images. It uses template matching (cv2.matchTemplate) with a set of watermark
template images.

4. Detect Security Thread:

The detect_security thread function compares each image with a set of security
thread template images. Template matching (cv2.matchTemplate) is used to find
potential matches. If a match is found based on a threshold, a rectangle is drawn
around the detected security thread in the image.

5. Custom Train-Test Split:

The custom_train_test_split function shuffles and divides the dataset of images into
training and validation sets (with 80% for training and 20% for validation) and
assigns the resulting subsets to variables train_images and val_images.

6. Train the Custom CNN Model:

The train_custom_model function is responsible for training the custom CNN
model. Image data generators are created for the training and validation sets.
The model is compiled with stochastic gradient descent (SGD) optimizer, binary
cross entropy loss, and accuracy as a metric. Model training is performed using
model.fit. A checkpoint is set up to save the best model based on validation loss.
After training, the model, classification report, confusion matrix, and accuracy are
printed and displayed.
a. Optimizer:
The Stochastic Gradient Descent (SGD) optimization algorithm is used for
training machine learning models, including neural networks. For a weight 𝜔 in
the model:
𝜔𝑛𝑒𝑤 = 𝜔𝑜𝑙𝑑 − 𝑛∇𝑗(𝜔(𝑜𝑙𝑑)

𝜔𝑜𝑙𝑑 is the current value of the weight.

29
 ∇𝑗(𝜔(𝑜𝑙𝑑) is the gradient of the loss function ∇𝑗 with respect to the weight
𝜔

𝑤𝑒𝑖𝑔ℎ𝑡 = 𝑤𝑒𝑖𝑔ℎ𝑡 − 𝑙𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑟𝑎𝑡𝑒 ∗ 𝑔𝑟𝑎𝑑𝑖𝑒𝑛𝑡

b. Loss Function:
The loss function quantifies the difference between the predicted values and the
actual labels. It guides the model during training to minimize prediction errors.
Binary loss entropy loss.
Formula:
1 𝑁
− ∑ 𝑦𝑖 log(𝑝𝑖 ) + (1 − 𝑦𝑖 ) log(1 − 𝑝𝑖 )
𝑁 𝑖=1

7. Load and Evaluate the Model:

After training, the saved model (custom_model_class_final.h5) is loaded using
load_model. Predictions are made on the validation set. The model's total accuracy
on the validation set is printed.

8. Calculate the confusion matrix for validation data:

The function calculate_confusion_matrix takes a trained model and a validation
dataset directory as input. It initializes an empty confusion matrix based on the
number of classes in the validation set. It then generates predictions for the
validation set using the model and updates the confusion matrix by comparing true
labels with predicted labels.

9. Calculate the classification report:

The classification report is generated by comparing the true labels of a dataset with
the predicted labels generated by a trained model. Predictions are generated for both
the validation and training datasets using the model. The true labels are obtained
from the dataset itself. Then, classification reports are generated for both datasets
using the true labels and predicted labels. These reports summarize the precision,
recall, F1-score, and support for each class in the dataset, providing insights into
the model's performance.

30
 Chapter 5: Implementation and Testing

5.1 Implementation

The process of checking and confirming a software or application is called testing. Initial
testing of the system by the developer themselves comes first. It aids in ensuring that the
system complies with the requirements. This chapter uses pseudocode and diagrams to
explain the theory underlying the implementation of intriguing and difficult features. With
a well-structured plan and the right development approach, the implementation phase of
any project may be the most pleasurable and trouble-free phase.

5.1.1 Tools used.

Different tools and technologies have been used to implement this application. They are
listed in the table below:

1. Programming Languages:
i Python
The system is developed using Python as the primary programming language,
leveraging a set of powerful tools and libraries. File and folder manipulation tasks
are handled using the os module, while image processing functionalities are
achieved through the cv2 (OpenCV) library for reading, processing, and saving
images, alongside the numpy library for numerical operations. The deep learning
and neural network components are implemented with the Keras libraries, allowing
the construction of a custom model architecture designed for watermark and silver
line detection in images. The code also makes use of Python's control flow
statements, mathematical functions from the math module, and text file operations
for conditional execution, rounding, and saving class indices to a file. Additionally,
the matplotlib.pyplot library aids in visualizing training metrics. Overall, Python's
versatility and the robust toolset provided by these libraries enable the effective
implementation of image processing and machine learning functionalities in the
code.

31
ii. HTML
In the system, the structure and styling of a web page for a system named
"DETECTO" are defined. The HTML includes metadata, links to external
stylesheets and scripts, as well as embedded styles and scripts. The page features a
responsive navigation menu, social media icons, and a prominent header with the
DETECTO logo. It also incorporates sections for displaying an introductory
message, an image upload panel, and a button to trigger image processing. The
JavaScript script handles events such as clicking the "See Full Details" button,
redirecting users based on a prediction result. This HTML code is integral to the
user interface and functionality of the system, facilitating image input, processing,
and result presentation. In the documentation, this HTML code should be described
as the front-end structure of the DETECTO system, outlining its key components
and interactions, especially focusing on user input and result presentation features.

iii. CSS
The provided HTML document contains embedded CSS styles that define the visual
presentation of the web page. The styles are written within <style> tags in the
document's <head> section and include rules for various elements such as
containers, images, buttons, and form elements. The CSS utilizes selectors to target
specific HTML elements and applies properties like color, padding, positioning,
and hover effects to achieve the desired appearance. Additionally, the document
incorporates some inline styles directly within HTML elements. The CSS rules also
include media queries for responsiveness, ensuring a consistent layout across
different screen sizes. Overall, the CSS in this document is structured to create a
visually appealing and responsive user interface for the DETECTO website, with
specific styling for buttons, image displays, and other elements to enhance the user
experience.

iv. Bootstraps
Bootstrapping is used to provide more stable training by exposing the models to
different subsets of the data in each iteration. It allows you to observe how the
models' performance changes when certain instances or features are included or
excluded.

32
v. Json
There is a small snippet of JSON code embedded in the HTML within a <script>
tag using the type "application/ld+json". This JSON code defines an organization
with properties such as name, logo, and social media links. JSON is a lightweight
data interchange format that is often used to structure and exchange data between a
server and a web application. In this case, it is used for providing structured data
about the organization for search engines and other applications that may consume
this information.

2. Machine Learning and Computer vision Libraries:

i. OpenCV
In the provided code, OpenCV (Open-Source Computer Vision) plays a pivotal role
in various image processing and computer vision tasks, serving as a fundamental
tool for detecting counterfeit money. The load_images_from_folder function
utilizes OpenCV to read and load grayscale images from a specified folder. The
extract_features function extensively leverages OpenCV for tasks such as resizing,
grayscale conversion, and adaptive histogram equalization for contrast
enhancement. The code also employs OpenCV functions like
fastNlMeansDenoising and convertScaleAbs for image denoising and contrast
adjustment. Notably, template matching methods, such as cv2.matchTemplate, are
applied to detect watermark and security thread thread in preprocessed image
regions. Overall, OpenCV is essential for manipulating, enhancing, and analyzing
images throughout the counterfeit money detection process, showcasing its
versatility in image processing within the context of the developed system.

ii. Scikit-learn
the scikit-learn (sklearn) library is used in the provided code. Specifically, it is used
for train-test splitting and scaling features. The custom_train_test_split function is
a custom implementation of a train-test split, which is typically handled by the
train_test_split function in scikit-learn. Additionally, a custom standard scaler
(CustomStandardScaler) is implemented, which is similar to the functionality
provided by scikit-learn's StandardScaler. These scikit-learn functionalities are
utilized to preprocess and split the data for training and testing the custom Support
Vector Machine (SVM) model.
33
3. Deep Learning Frameworks:
Keras
In the provided code, Keras is utilized for constructing, compiling, and training a
convolutional neural network (CNN) model. Custom layers, such as CustomDense
and inverted_residual_block, are implemented using Keras's Layer class. The
MobileNetV2 architecture is defined using Keras layers, and the model is compiled
with the Stochastic Gradient Descent (SGD) optimizer, binary crossentropy loss,
and accuracy as the evaluation metric. Keras's ImageDataGenerator is employed for
data preprocessing and augmentation, facilitating the creation of batches for
training and validation. The training process is orchestrated using the fit method,
and the ModelCheckpoint callback from Keras is employed to save the best model
weights based on validation loss. Finally, the trained model is saved using Keras's
save method, resulting in two files: one for the model checkpoint during training
(watermark_mobilenetmodel_checkpoint.h5) and another for the final trained
model (watermark_mobilenetmodel_final.h5). Keras streamlines the development
of deep learning models by providing a high-level API and abstracting many
complexities associated with neural network implementation.

4. Image Processing Tools:

Adobe Photoshop or GIMP
Adobe Photoshop or GIMP is used to reduce noise or artifacts in the images,
ensuring that the input to the detection models is clean and representative of
genuine currency features. Adjusting the brightness, contrast, and sharpness of the
images can enhance the visibility of subtle features in the currency notes. The role
of these image editing tools is to prepare and enhance the data for effective model
training and evaluation.

5. Development Frameworks:
Flask
In the provided code, Flask is utilized to create a web application for predicting
whether an uploaded image contains counterfeit money. Flask is a Python web
framework that simplifies the process of building web applications. The application
defines routes, such as "/", "/result", and "/predictt", each associated with specific
functionalities. When a user accesses the specified routes, Flask invokes the
34
 corresponding functions, rendering HTML templates, processing image uploads,
and providing predictions. Flask seamlessly handles HTTP requests and responses,
enabling the integration of the image processing logic, SVM prediction, and web
interface. Additionally, the app.run() method is employed to launch the Flask
application, allowing it to run locally with debugging capabilities. The code
demonstrates the use of Flask's simplicity and flexibility in building a web-based
counterfeit money detection system.

6. Diagram tool
Draw.io plays a pivotal role in our Counterfeit Money Detection System's
development by serving as the go-to tool for Entity-Relationship (ER) diagrams and
Data Flow Diagrams (DFD), offering a comprehensive visualization of data
structures and process flows. Meanwhile, Figma takes the lead in crafting the
system's user interface, ensuring an intuitive and visually appealing design.

7. SQLite
In your project, SQLite is employed as a relational database to manage user-related
data, currency images, and detection results. The user_id table likely contains user
information, while the currency_image table stores images along with relevant
details, linked to users via foreign keys. The detection_result table records outcomes
of currency detection processes, associating results with specific users and currency
images. SQL queries enable seamless interaction with these tables, facilitating tasks
such as data insertion, retrieval, and management within the application. Overall,
SQLite provides a lightweight and efficient solution for organizing and accessing
essential data in the context of user accounts, currency images, and detection results.

35
5.1.2 Implementation Details of Algorithm:

Implementation of SVM algorithm for Counterfeit money detection

1. Custom SVM Implementation:

The system uses a custom SVM implementation named CustomSVM. It defines a
custom SVM class (CustomSVM) with an adjustable learning rate and number of
epochs. This class contains methods for model initialization, fitting (training), and
prediction. The SVM model is trained on the scaled training features using the fit
method of the CustomSVM class. The trained SVM model is evaluated on both the
training and testing sets by comparing the predicted labels with the actual labels.
Accuracy metrics for both sets are calculated and printed

2. Model Initialization:
Within the CustomSVM class, the weights and bias of the SVM model are
initialized randomly during instantiation. These parameters will be adjusted during
the training process to create an optimal decision boundary.

3. Model Training (fit)

The SVM model is trained using a simple gradient descent approach in the fit
method of the CustomSVM class. The weights and bias are updated iteratively
based on the error between the predicted and actual labels for each training example.

4. Prediction Function:
The predict method of the CustomSVM class is responsible for making predictions
based on the learned weights and bias. It calculates the dot product of the feature
vector and weights, then applies a threshold to determine the predicted class (1 or
0).

36
 Figure 5. 1 Custom SVM

5. Feature Scaling:
A custom standard scaler (CustomStandardScaler) is defined and applied to scale
the features. The scaler's mean and standard deviation are used to normalize the
data.

37
 Figure 5. 2 CustomStandardScalar

6. Model Saving:
The trained SVM model and the scaler used for feature scaling are saved as joblib
files ('custom_svm_model.joblib' and 'scaler_model.joblib', respectively).

Figure 5. 3 Model Saving

38
Implementation of CNN algorithm:

Implementing a Convolutional Neural Network (CNN) involves several steps, including

defining the architecture, preparing the data, training the model, and evaluating its
performance.

1. Preprocessing Images (preprocess input function):

It primarily utilizes techniques to enhance image quality and identify watermark.
Images are resized, converted to grayscale, and undergo Contrast Limited Adaptive
Histogram Equalization (CLAHE) for improved contrast. Non-Local Means
Denoising is then applied to reduce noise. A watermark detection mechanism uses
template matching, and if a match exceeds a threshold, the watermark is considered
detected. The script also splits the dataset into training and validation sets, aiding
the subsequent training of a custom convolutional neural network (CNN) model for
effective counterfeit money detection.

2. Watermark detection.
The watermark detection involves preprocessing both input images and watermark
templates. Template matching is applied, measuring normalized cross-correlation.
If the match surpasses a threshold, a bounding box is drawn on the image, signifying
watermark detection. The script outputs a matching percentage for confidence
assessment.

Figure 5. 4 Watermark Detection

39
3. Security thread Detection:
The detect_silverline function uses template matching (cv2.matchTemplate) to
identify security thread in the preprocessed images. If a security thread is detected,
a bounding box is drawn around it, and the matching percentage is printed.

Figure 5. 5 Security Thread Detection

4. Custom CNN model

The custom Convolutional Neural Network (CNN) model is designed for detecting
counterfeit money. It utilizes Keras, featuring convolutional and pooling layers,
culminating in a dense layer for binary classification. The model is trained using
SGD with a learning rate of 0.001 and binary cross-entropy loss. ModelCheckpoint
saves the best-performing model during training. The process includes
preprocessing images, training with an ImageDataGenerator, and visualizing
accuracy and loss curves. The trained model is then saved for future use.

40
 Figure 5. 6 Custom CNN model

5. Training custom CNN

The training of the custom Convolutional Neural Network (CNN) model for
counterfeit money detection involves preprocessing images and splitting them into
training and validation sets. Images are resized, enhanced, and subjected to
watermark detection before being organized into respective folders. The CNN
model, designed with convolutional and pooling layers, is trained using an
ImageDataGenerator for augmentation. The training process spans five epochs with
a batch size of 30, employing Stochastic Gradient Descent (SGD) as the optimizer
with a learning rate of 0.001 and binary cross-entropy loss. ModelCheckpoint is
utilized to save the best-performing model based on validation loss. The trained
model, alongside class indices, is saved for future use. The training progress is
visualized through accuracy and loss curves, providing insights into the model's
performance.

41
 Figure 5. 7 Training Custom CNN

5.2 Testing

5.2.1 Test Cases for Unit Testing

Unit testing is performed for testing modules against detailed design. Every step of the
project's design and coding has been tested. While testing, we test the module interface to
make sure that data is properly flowing into and out of the program unit. By inspecting the
local data structure, we ensure that the temporarily stored data keeps its integrity during the
algorithm's execution. Finally, each path that handles errors is tested.

42
 Table 5. 1 Test case For Registration

ID Test Case Test Data Expected Actual Test

Description result Result Result

R_1 User enters Redirects to Redirected Pass

Enter your email:
valid signup Login Page to Login
[email protected]
information page
Create password: 123456
Confirm password: 123456

R_2 User enters Enter Your email: Display “Password Pass

different Error and Confirm
[email protected]
password Message. password
Create password: 123456
and confirm didn’t
Confirm password: 23459
password match.”

Table 5. 2 Test Case for User Signup

ID Test Case Test Data Expected Actual Test

Description result Result Result

L_1 Enter your email: Redirects to Redirected to Pass

User enters
[email protected] Currency Currency
valid login
Enter your password: 123456 detection detection
information
page page

43
L_2 User enters Enter your email: Display “Please Pass
invalid [email protected] Error enter a
login Enter your password: 098765 Message correct
information username
and
password.”

L_3 User leaves Enter your email: Display “Please Pass

the login [email protected] Error enter
fields Enter your password: Message password."
empty.

Table 5. 3 Test for Feature Extraction

S.N. Test Test Description Input Expected Actual Remarks

Case Image Result Result
1 Test-1 Compare feature Real note Real Real Pass
of image with the
template
2 Test-2 Compare feature Fake note Fake Fake Pass
of image with the
template

44
 Table 5. 4 Test Case for Model Accuracy

S.N. Test Test Description Input Expected Actual Remarks

Case Image Result Result
1 Test-1 Test on test data Real note High Pass
of the dataset Accuracy
2 Test-2 Test on test data Fake note Low Pass
of the dataset Accuracy

5.2.2 Test Cases for System Testing

System testing is typically done to look for faults brought on by unexpected interactions
between subsystem and system components. Once the source code is developed, software
must be tested to identify and correct any potential mistakes before being delivered to the
client.

Table 5. 5 System Testing

S.N. Test Case Test Expected Actual Remarks

Description Outcome Outcome

1 Test-1 Upload Real Real Image Real Image Pass

Image
2 Test-2 Upload Fake Fake Image Fake Image Pass
Image

45
Figure 5. 8 Fake image

Figure 5. 9 Real Image

46
5.3 Result Analysis

The analysis of the counterfeit money detection system's results provides crucial insights
into its performance and effectiveness. By meticulously examining various performance
metrics, we aim to ascertain the system's ability to accurately distinguish between genuine
and counterfeit currency. This analysis serves as a critical step in evaluating the system's
reliability, identifying areas for improvement, and ensuring its robustness in real-world
scenarios.

1. Data Collection
Our counterfeit money detection system benefits from a robust dataset comprising
a total of 472 samples. This dataset was meticulously compiled from the repository
https://amitness.com/ml-datasets/ to ensure comprehensive coverage and relevance
to our detection system's objectives. This dataset encompasses images depicting
authentic banknotes alongside their counterfeit counterparts. Notably, it features a
comprehensive array of Nepalese currency denominations, including 10, 50, 100,
and 1000 rupee notes. Despite the availability of various denominations within the
dataset, we deliberately opted to concentrate solely on the 1000-rupee
denomination. This decision was made to streamline our development efforts,
allowing us to focus exclusively on optimizing the detection system's efficacy in
identifying counterfeit 1000-rupee notes within the Nepalese context. Furthermore,
it's important to note that we captured the images ourselves and incorporated them
into the dataset, ensuring a high degree of relevance and authenticity for our training
data.

47
2. Epoch Progress: Training Metrics
The training and testing accuracy of an SVM model trained over 100 epochs is
represented in the graph below.

Figure 5. 10 SVM Training and Testing Accuracy

The training and testing loss of an SVM model trained over 100 epochs is
represented in the graph below.

Figure 5. 11 SVM Training and Testing Loss

48
The training and validation accuracy of CNN model for Watermark is trained over
100 epochs is represented in the graph.

Figure 5. 12 Watermark Training and Validation Accuracy

The training and validation loss of CNN model for Watermark is trained over 100
epochs is represented in the graph.

Figure 5. 13 Watermark Training and Validation Loss

49
The training and validation accuracy of CNN model for Security Thread is trained
over 100 epochs is represented in the graph.

Figure 5. 14 Security Thread Training and Validation Accuracy

The training and validation loss of CNN model for Security Thread is trained over
100 epochs is represented in the graph.

Figure 5. 15 Security Thread Training and Validation Loss

50
3. Performance Metrics
Evaluating our counterfeit money detection system involves assessing key
performance metrics. Accuracy measures how often it correctly identifies
counterfeit bills. Precision and recall indicate the system's ability to accurately
classify counterfeit bills and identify genuine ones, respectively. The F1 Score
provides a balanced evaluation of overall performance, while false positive and
false negative rates reveal the system's tendencies for misclassification. By
analyzing these metrics, we can determine the system's reliability in distinguishing
between genuine and counterfeit currency.

Performance metrics were computed separately for two components of our system:
the Support Vector Machine (SVM) for classification and the Convolutional Neural
Network (CNN) for determining the similarity percentage of watermark and
security thread.
The SVM classifier was trained on a dataset containing examples of both genuine
and counterfeit currency. This dataset provided the basis for the SVM to learn how
to distinguish between the two classes using features extracted from the currency
images. We then assessed how well the classifier performs on this training data to
understand its ability to correctly classify genuine and counterfeit currency
instances.

Figure 5. 16 SVM: Accuracy for Training Data Table

51
 Figure 5. 17 SVM: Accuracy for Testing Data Table

The SVM classifier trained on genuine and counterfeit currency data exhibited exceptional
performance metrics, including perfect classification accuracy, precision, recall, F1-score,
and support for both genuine and counterfeit currency instances. For the SVM classifier
tested on genuine and counterfeit currency data, it achieved an overall accuracy of 94% on
the testing data, demonstrating strong performance with high precision, recall, and F1-
scores for both genuine and counterfeit currency classes.

The performance metrics for counterfeit money detection, specifically focusing on

determining the similarity percentage of watermark and security thread using CNN, provide
valuable insights into the model's efficacy.

Figure 5. 18 CNN (Watermark): Accuracy for Training Data Table

52
 Figure 5. 19 CNN (Watermark): Accuracy for Validation Data Table

For CNN analysis of watermark similarity, the model achieved high precision, recall, F1-
scores, and accuracy of 96%, indicating its effectiveness in distinguishing between genuine
and counterfeit currency.

Figure 5. 20 CNN (Security Thread): Accuracy for Training Data Table

Figure 5. 21 CNN (Security Thread): Accuracy for Validation Data Table

Similarly, for CNN analysis of security thread similarity, the model showcased
commendable proficiency with an accuracy of 93%.

53
In a counterfeit money detection system using SVM classification with CNN for matching
percentage evaluation, the confusion matrix provides a concise summary of the system's
performance, offering insights into the accuracy and reliability of the system in
distinguishing between genuine and counterfeit currency. By analyzing these metrics, the
system's effectiveness can be evaluated, guiding further improvements to enhance its
performance in identifying counterfeit currency accurately.

4. Confusion Matrix
In a counterfeit money detection system using SVM classification with CNN for
matching percentage evaluation, the confusion matrix provides a concise summary
of the system's performance. It outlines four key scenarios: True Positives (genuine
currency correctly classified), True Negatives (counterfeit currency correctly
classified), False Positives (counterfeit currency classified as genuine), and False
Negatives (genuine currency classified as counterfeit), offering insights into the
accuracy and reliability of the system in distinguishing between genuine and
counterfeit currency. By analyzing these metrics, the system's effectiveness can be
evaluated, guiding further improvements to enhance its performance in identifying
counterfeit currency accurately.

The confusion matrix for an SVM algorithm is as follows:

Figure 5. 22 SVM: Confusion Matrix

54
The confusion matrix for a CNN analysis of watermark similarity is as follows:

Figure 5. 23 CNN (Watermark): Confusion Matrix

The confusion matrix for a CNN analysis of security thread similarity is as follows:

Figure 5. 24 CNN (Security Thread): Confusion Matrix

55
 Chapter 6: Conclusion and Future Recommendations

6.1 Conclusion

The implementation of a counterfeit money detection system utilizing Support Vector

Machine (SVM) and Convolutional Neural Network (CNN) algorithms offers a promising
solution to the prevalent issue of fraudulent currency circulation. Through the integration
of SVM and CNN algorithms, the system effectively determines the authenticity of
uploaded currency image by analyzing key features such as watermark and security thread.

The SVM algorithm serves as the initial line of defense, effectively discerning between
authentic and counterfeit currency based on crucial features such as watermark and security
thread present in the uploaded image. Through intuitive user interaction, where a threshold
value can be adjusted to fine-tune the detection process, users gain real-time insights into
the authenticity of the currency, enabling informed decision-making. The accuracy of
testing data using SVM stands at 94%.

Moreover, the incorporation of CNN enhances the system's functionality by providing users
with detailed insights into the matching percentage of watermark and security thread of the
uploaded image. This algorithm offers more comprehensive details about the uploaded
image. The validation data accuracy for detecting the watermark is 91%, while for detecting
the security thread, validation data accuracy is 93%.

6.2 Future Recommendation

Technology is advancing at a rapid pace these days. In this system, we were able to detect
watermark and security thread. We’ve looked at the whole image of money so far, but in
the future, we'll try to include all of the security features of money by using a fair
fundamental structure and providing sufficient preparation information. When a picture is
loaded into the training folder from the outside, it does not provide 100 percent accuracy.
By optimizing the system, we can solve this problem.

56
 References

[1] A. Vidhate, Y. Shah, H. Keshri and R. Nikhare, "Fakecurrency detection

application," Int. Res. J. Eng. Technol. (IRJET), vol. 08(05), pp. 2395-0056, 2021.

[2] A. Upadhyaya and D. V. Shokeen, "“ANALYSIS OF COUNTERFEIT

CURRENCY DETECTION TECHNIQUES FOR CLASSIFICATION MODEL”,"
in International Conference on Computing Communication and Automation
(ICCCA), 2018.

[3] D. S. Virakthamath and T. Kshama, "“REVIEW ON DETECTION OF FAKE

CURRENCY USING IMAGE PROCESSING TECHNIQUES”," in International
Conference on Intelligence Computing and Control System, (ICICCS), 2021.

[4] A. Zarin and J. Uddin, "" A HYBRID FAKE BANK NOTE DETECTION MODEL
USING OCR, FACE RECOGNITION AND HOUGH FEATURES"," in
Cybersecurity and Cyberforensics (CCC), 2019.

[5] S. R. Darade and P. G. R. Gadveer, "“AUTOMATIC RECOGNITION OF FAKE

INDIAN CURRENCY NOTE”," in International Conference on Electrical Power
and Energy System, (ICEPES), 2016.

[6] A. Yadav and T. Jain, "" EVALUATON OF MACHINE LEARNING

ALGORITHMS FOR DETECTION OF FAKE BANK CURRENCY”," in
International Conference on Cloud Computing, Data Science & Engineering
(ICCCDE), 2021.

[7] s. Gothe, K. Naik and V. Joshi, Fake currency detection using image processing
and machine learning, 2018.

[8] V. L. Nadh and S. Prasad, "Support vector machine in the anticipation of currency
markets," Int. J. Eng. Technol, vol. 7(2), p. 66–68, 2018.

[9] V. V and L. M, "Real time fake currency note detection using deep learning.," Int. J.
Eng.Adv. Technol. (IJEAT), vol. 9(1S5), pp. 2249-8958, 2019.

57
[10] A. T, B. G, W. P and C. P, "Fake currency detection using image processing,"
IOPConf.Ser.Mater.Sci.Eng., p. 263, 2017.

[11] T. Agasti, "Fake currency detection using image processing," IOP Conference
Series: Materials Science and Engineering, p. 243, 2017.

[12] "Currency Management Department," Nepal Rastra Bank, [Online]. Available:

https://www.nrb.org.np/departments/cmd/. [Accessed 4 October 2023].

58
 Appendix

Figure 7. 1 Home Page

59
Figure 7. 2 Login Page

Figure 7. 3 Signup Page

60
Figure 7. 4 Image with 0.7 threshold

Figure 7. 5 Matched Percentage

61
Figure 7. 6 Image with 0.9 Threshold

Figure 7. 7 Fake image input

62