Signature Verification System Using CNN

Dissertation submitted to
Shri Ramdeobaba College of Engineering & Management, Nagpur
in partial fulfillment of requirement for the award of degree of

Bachelor of Technology (B.Tech)

In

COMPUTER SCIENCE AND ENGINEERING

By

Anup Dhoble (Roll no. 27)

Ankit Bhardwaj (Roll no.25 )
Ajay Mahato (Roll no.74)
Suhani Taran (Roll no. 16)
Of
VI Semester
Guide
Prof. Bhagyashree S. Madan

Department of Computer Science and Engineering

Shri Ramdeobaba College of Engineering & Management, Nagpur 440 013
(An Autonomous Institute affiliated to Rashtrasant Tukdoji Maharaj Nagpur University Nagpur)
April 2024

i
Signature Verification System Using CNN
Dissertation submitted to
Shri Ramdeobaba College of Engineering & Management, Nagpur
in partial fulfillment of requirement for the award of degree of

Bachelor of Technology (B.Tech)

In

COMPUTER SCIENCE AND ENGINEERING

By

Anup Dhoble (Roll no. 27)

Ankit Bhardwaj (Roll no.25 )
Ajay Mahato (Roll no.74)
Suhani Taran (Roll no. 16)
Of
VI Semester
Guide
Prof. Bhagyashree S. Madan

Department of Computer Science and Engineering

Shri Ramdeobaba College of Engineering & Management, Nagpur 440 013
(An Autonomous Institute affiliated to Rashtrasant Tukdoji Maharaj Nagpur University Nagpur)
April 2024

ii
SHRI RAMDEOBABA COLLEGE OF ENGINEERING MANAGEMENT, NAGPUR
(An Autonomous Institute affiliated to Rashtrasant Tukdoji Maharaj Nagpur University Nagpur)

Department of Computer Science and Engineering

CERTIFICATE

This is to certify that the Thesis on “ Signature Verification System Using CNN” is a
Bonafide work of Anup Dhoble, Ankit Bhardwaj, Ajay Mahato, Suhani Taran submitted to
the Rashtrasant Tukdoji Maharaj Nagpur University, Nagpur in partial fulfillment of the
award of a Degree of Bachelor of Technology (B.Tech), in Computer Science and
Engineering. It has been carried out at the Department of Computer Science and Engineering,
Shri Ramdeobaba College of Engineering and Management, Nagpur during the academic
year 2023-2024.

Date:
Place: Nagpur

______________________ ________________
Prof.Bhagyashree Madan Dr. R. Hablani
Project Guide H.O.D
Department of Computer Science Department of Computer Science
and Engineering and Engineering

_______
Dr. R. S. Pande
Principal

iii
 DECLARATION

We hereby declare that the thesis titled “Signature Verification System Using CNN”
submitted herein, has been carried out in the Department of Computer Science and
Engineering of Shri Ramdeobaba College of Engineering and Management, Nagpur. The
work is original and has not been submitted earlier as a whole or part for the award of any
degree/diploma at this or any other institution / University.

Date:
Place: Nagpur

Sign
Anup Dhoble

Ankit Bhardwaj

Ajay Mahato

Suhani Taran

iv
 APPROVAL SHEET

This report entitled “Signature Verification System Using CNN” by Anup Dhoble, Ankit
Bhardwaj, Ajay Mahato, Suhani Taran is approved for the degree of Bachelor of
Technology (B.Tech).

______________________ ______________________
Prof.Bhagyashree Madan
Project Guide External Examiner

_______________
Dr. R. Hablani
H.O.D, CSE

Date:
Place: Nagpur

v
 ACKNOWLEDGEMENTS

We would like to express our deepest appreciation and gratitude to all those who contributed
to the successful completion of this project, both directly and indirectly.

First and foremost, we extend our heartfelt thanks to our esteemed project guide, Prof.
Bhagyashree S. Madan, for her invaluable guidance, unwavering support, and profound
insights throughout the duration of this endeavor. Prof. Bhagyashree S. Madan's mentorship,
expertise, and dedication played a pivotal role in shaping the direction of our research and
fostering an environment of excellence.

We are immensely grateful to our diligent and dedicated team members, Anup Dhoble, Ankit
Bhardwaj, Ajay Mahato, and Suhani Taran, whose collaborative efforts, synergy, and
commitment were integral to the successful execution of this project. Each member's unique
skills, tireless efforts, and unwavering enthusiasm significantly enriched our work and
propelled us towards achieving our objectives.

We extend our sincere appreciation to ‘Shri Ramdeobaba College Of Engineering And

Management’ for providing us with the necessary resources, infrastructure, and support to
undertake this research endeavor. The conducive academic environment and state-of-the-art
facilities offered by the institution played a crucial role in facilitating our exploration and
experimentation.

Furthermore, we would like to express our gratitude to the participants who generously
contributed their time and signatures, thereby enabling the creation of a comprehensive and
diverse dataset essential for the development and evaluation of our signature verification
system.

vi
 ABSTRACT

In an era marked by rapid technological advancement and evolving security threats, the

demand for robust authentication systems is paramount. This thesis presents the development

of a Signature Verification System employing Convolutional Neural Networks (CNNs) to

authenticate handwritten signatures.

Signatures are pivotal for personal identification in legal, financial, and administrative

contexts, demanding meticulous scrutiny to prevent forgery. Our system, built on the Django

framework, offers an automated solution to enhance security and streamline verification

procedures.

Utilizing CNNs, our system compares two images – an original and a comparison signature –

enabling accurate authentication. Through rigorous testing on a diverse dataset, we ensured

the system's efficacy across various scenarios.

Our results demonstrate promising performance, distinguishing genuine signatures from

imposters, thus enhancing security measures against forgery and falsification.

This research underscores the significance of integrating advanced technologies into

conventional practices to effectively address emerging challenges in authentication and

security.

vii
 TABLE OF CONTENTS

Content Page No.

Cover Page i

Cover Page ii

Certificate iii

Declaration iv

Approval Sheet v

Acknowledgements vi

Abstract vii

List of figures x

Chapter 1. Introduction 1

1.1 Background 1

1.2 Motivation 1

1.3 Objectives 1

Chapter 2. Literature Review 2

2.1 Review of Existing Methods and Technologies for Signature 2

Verification

2.2 Limitations of Traditional Manual Verification Methods 2

2.3 Exploration of Previous Research on CNN-Based Approaches for 3

Signature Verification

Chapter 3. Methodology 4

3.1 Data Collection and Preparation 4

3.2 Model Architecture 4

3.3 Training and Evaluation 5

3.4 Explanation of the Layers Used in the CNN Models 5

Chapter 4. Preprocessing of Signature Images 6

4.1 Image Preprocessing Techniques 6

viii
4.2 Signature Extraction 6

4.3 Resizing and Background Addition 7

Chapter 5. CNN Model Design and Training 8

5.1 Model Version 1 8

5.2 Model Version 2 9

5.3 Model Version 3 11

Chapter 6. System Implementation 14

6.1 Overview of the system architecture and components 14

6.2 Description of the User Interface Design and Functionalities 14

6.3 Integration of the CNN Model into the Signature Verification 15

System

Chapter 7. Results and Discussion 17

7.1 Model Training and Evaluation Results 17

7.2 Performance Analysis of Model Versions 17

7.3 Implications and Areas for Improvement 17

Chapter 8. Discussion and Conclusion 19

8.1 Conclusion 19

8.2 Key Achievements 19

8.3 Future Work 19

List of Publications 20

References 21

Appendices 22

ix
 LIST OF FIGURES

Figure Number Figure Name Page Number

1. Preprocessing of Signatures 7

2. Version 1 Training, Validation Accuracy and Loss 8

3. Version 1 ROC curve 9

4. Version 2 Training, Validation Accuracy and Loss 10

5. Version 2 Test Accuracy and Loss 10

6. Version 2 ROC curve 11

7. Version3 Training, Validation Accuracy 12

8. Version 3 Test Accuracy and Loss 12

9. Version 3 ROC curve 13

x
 CHAPTER 1
INTRODUCTION

______________________________________________________________________

1.1 Background

Handwritten signatures have long served as a fundamental means of personal identification

and document verification in various legal, financial, and commercial contexts. The unique
patterns and strokes of an individual's signature are used to authenticate important
documents, such as checks, contracts, and certificates. Traditionally, manual verification
methods have been employed to compare signatures against known samples to detect forgery
and ensure the authenticity of documents. However, with advancements in technology, the
prevalence of sophisticated forgery techniques has necessitated the development of more
robust and efficient verification systems.

1.2 Motivation

The increasing prevalence of document forgery and falsification poses significant challenges
to the integrity and security of legal and financial transactions. Manual signature verification
processes are often time-consuming, subjective, and prone to human error, making them
inadequate for detecting sophisticated forgery techniques. Moreover, as technology evolves,
the sophistication of forgery methods continues to escalate, further underscoring the need for
automated and reliable signature verification systems. Therefore, there is a pressing need to
develop an efficient and accurate signature verification system that can mitigate the risks
associated with document fraud and ensure the integrity of important transactions.

1.3 Objectives

The primary goal of this project is to develop a robust signature verification system utilizing
Convolutional Neural Networks (CNNs) within the Django framework. Specifically, the
objectives of the project include:

● Collecting and preprocessing a diverse dataset of handwritten signature images.

● Implementing and training CNN models, leveraging architectures such as ResNet50,
for signature classification.
● Developing a user-friendly web application within the Django framework, allowing
users to authenticate signatures by comparing them with known samples.
● Evaluating the performance of the signature verification system in terms of accuracy,
efficiency, and reliability.
● By achieving these objectives, the project aims to provide a practical solution for
automated signature verification, enhancing the security and efficiency of document
authentication processes.

1
 CHAPTER 2
LITERATURE REVIEW

______________________________________________________________________

2.1 Review of Existing Methods and Technologies for Signature Verification

Handwritten signature verification has been the subject of extensive research, leading to the
development of various methods and technologies aimed at authenticating signatures.
Traditional methods primarily relied on manual inspection, where experts compared
signatures against known samples to detect inconsistencies or signs of forgery. However,
manual verification methods are time-consuming, subjective, and susceptible to human error.
To address these limitations, researchers have explored automated approaches utilizing
computational techniques such as image processing and machine learning.

Image processing techniques involve preprocessing signature images to enhance features and
extract relevant information for analysis. These techniques may include grayscale conversion,
noise reduction, and edge detection algorithms. Additionally, feature extraction methods such
as Histogram of Oriented Gradients (HOG) and Scale-Invariant Feature Transform (SIFT)
have been employed to capture distinctive characteristics of signatures.

Machine learning algorithms, particularly Convolutional Neural Networks (CNNs), have

gained popularity for signature verification tasks due to their ability to learn complex patterns
and features from raw data. CNN-based approaches involve training models on large datasets
of signature images to classify signatures as genuine or forged. These models can
automatically extract features and identify subtle variations in handwriting style, making
them suitable for signature authentication.

2.2 Limitations of Traditional Manual Verification Methods

Traditional manual verification methods suffer from several limitations that hinder their
effectiveness in detecting forged signatures. Firstly, manual inspection is labor-intensive and
time-consuming, especially when dealing with a large volume of documents. Moreover,
human judgment is inherently subjective and prone to bias, leading to inconsistencies in the
verification process. Additionally, experts may lack the necessary expertise or training to

2
identify sophisticated forgery techniques, making it challenging to detect fraudulent
signatures accurately.

Furthermore, manual verification methods are limited in their ability to handle variations in
handwriting styles and signatures across different individuals. As signatures can evolve over
time and exhibit natural variations, manual inspectors may struggle to differentiate between
genuine variations and forged signatures. Consequently, relying solely on manual verification
methods may leave documents vulnerable to undetected forgery, compromising the integrity
of legal and financial transactions.

2.3 Exploration of Previous Research on CNN-Based Approaches for Signature

Verification

Recent advancements in deep learning, particularly CNNs, have paved the way for innovative
approaches to signature verification. Several studies have explored the effectiveness of
CNN-based models in automatically authenticating handwritten signatures. These models are
trained on large datasets of genuine and forged signature images, allowing them to learn
discriminative features and patterns indicative of genuine signatures.

Research has demonstrated the superior performance of CNN-based approaches compared to

traditional methods, achieving higher accuracy rates and robustness against various forms of
forgery. Additionally, CNNs offer scalability and efficiency, enabling rapid processing of
large volumes of documents with minimal human intervention.

3
 CHAPTER 3
METHODOLOGY

______________________________________________________________________

3.1 Data Collection and Preparation

The process of collecting and preparing signature images for training the models involves
several steps to ensure the availability of a diverse and representative dataset. Firstly, a
dataset containing genuine and forged signature images is compiled from various sources,
including public repositories, legal documents, and financial records. Care is taken to include
signatures from different individuals with varying handwriting styles and characteristics.

Once the dataset is curated, preprocessing techniques are applied to standardize and enhance
the quality of the signature images. This typically includes steps such as grayscale
conversion, noise reduction, and normalization to ensure consistency across images.
Additionally, data augmentation techniques may be employed to increase the variability of
the dataset, such as rotation, scaling, and translation, to simulate real-world variations in
signatures.

3.2 Model Architecture

The model architecture plays a crucial role in the performance and effectiveness of the
signature verification system. In this project, Convolutional Neural Networks (CNNs) are
utilized due to their ability to learn hierarchical features from raw image data. The
architecture chosen for the models is ResNet50, a deep CNN architecture known for its
effectiveness in image classification tasks.

ResNet50 consists of multiple layers, including convolutional layers, pooling layers, and fully
connected layers. The convolutional layers extract features from input images through a
series of convolutions, while the pooling layers downsample the feature maps to reduce
computational complexity. Finally, the fully connected layers perform classification based on
the learned features.

4
3.3 Training and Evaluation

The training process involves feeding the preprocessed signature images into the CNN
models and iteratively adjusting the model parameters to minimize the classification error.
During training, the models learn to differentiate between genuine and forged signatures by
optimizing a predefined loss function, typically categorical cross-entropy.

Hyperparameter tuning is performed to optimize the performance of the models, including

parameters such as learning rate, batch size, and number of epochs. Additionally, techniques
such as early stopping and dropout regularization may be employed to prevent overfitting and
improve generalization.

Once trained, the models are evaluated using a separate validation dataset to assess their
performance in terms of accuracy, precision, recall, and other relevant metrics. The models
are then tested on unseen data to measure their effectiveness in real-world scenarios.
Evaluation results are analyzed to identify areas for improvement and refinement of the
signature verification system.

3.4 Explanation of the Layers Used in the CNN Models:

In the CNN models employed for signature verification, several types of layers are utilized to
extract and process features from the input signature images:

● Convolutional Layer: The core building block of a CNN, convolutional layers consist
of learnable filters that scan across the input image, extracting spatial features through
convolution operations.
● Pooling Layer: Pooling layers downsample feature maps, reducing their spatial
dimensions while retaining essential information. Max pooling is commonly used to extract
dominant features from localized regions of the input.
● Fully Connected Layer: Fully connected layers connect every neuron in one layer to
every neuron in the subsequent layer, enabling the model to learn complex patterns and make
predictions based on the extracted features.

5
 CHAPTER 4
PREPROCESSING OF SIGNATURE IMAGES

______________________________________________________________________

4.1 Image Preprocessing Techniques:

The preprocessing of signature images involves several essential steps to enhance their
quality and suitability for further analysis:

● Grayscale Conversion: The input signature image is converted from its original color
space to grayscale using the cv2.cvtColor() function. This step simplifies subsequent
processing by reducing the image to a single channel representing pixel intensities.
● Gaussian Blur: A Gaussian blur is applied to the grayscale image using the
cv2.GaussianBlur() function. This operation helps remove noise and smooth out irregularities
in the image, improving the effectiveness of subsequent processing steps.
● Thresholding: Adaptive thresholding is performed to binarize the blurred image,
separating the foreground (signature) from the background. The cv2.adaptiveThreshold()
function with the cv2.ADAPTIVE_THRESH_GAUSSIAN_C flag adapts the threshold value
for each pixel based on its local neighborhood, enhancing robustness in varying illumination
conditions and background noise.
● Morphological Operations: Morphological closing operations are used to fill small
gaps and smooth out the contours of the signature. The cv2.morphologyEx() function with the
cv2.MORPH_CLOSE flag applies a closing operation using a predefined kernel, ensuring a
continuous and well-defined boundary for the extracted signature.

4.2 Signature Extraction:

Following preprocessing, the signature is extracted from the preprocessed image using
contour detection techniques. The cv2.findContours() function identifies the contours of the
thresholded image, representing the boundaries of distinct objects or regions. The largest
contour, corresponding to the signature, is extracted by iterating over the detected contours
and selecting the contour with the maximum area.

6
4.3 Resizing and Background Addition:

To ensure consistency and compatibility with downstream processing tasks, the extracted
signature is resized to a square shape using the cv2.resize() function. The maximum
dimension of the signature image is determined, and the image is resized to a square of equal
width and height. Additionally, a constant background is added to the resized signature using
numpy operations, ensuring that the output image has a uniform size and aspect ratio.

These preprocessing techniques collectively enhance the quality and suitability of signature
images for subsequent analysis and classification tasks in signature verification systems.

Fig1. Preprocessing of Signature Image

● Image Extraction: The signature is extracted from the document provided as the
original image. This results in the 'Extracted Signature' as mentioned in the context.
● Resizing to Constant Size: The extracted signature is then resized to a constant size,
specifically a square shape. This is necessary for the Convolutional Neural Network (CNN)
to process the image effectively. The resized signature is referred to as the 'Output Image with
Constant Size (Square)' in the context.
● Preparation for Machine Learning Model: The preprocessed signatures are then
prepared for the machine learning model. In your project, a CNN model is used for signature
classification.

7
 CHAPTER 5
CNN MODEL DESIGN AND TRAINING

______________________________________________________________________

5.1 Model Version 1:

● Architecture: Utilizes a ResNet50 model pre-trained on the ImageNet dataset,
followed by custom top layers for classification.
● Training Process: The base ResNet50 layers are frozen, and the model is trained with
a learning rate of 0.001 using the Adam optimizer for 100 epochs.
● Performance Metrics: Achieves a test accuracy of approximately 86%, with
fluctuations observed during training.
● Challenges Faced:
- Limited Model Complexity: The initial ResNet50 model lacked complexity,
consisting of basic layers without much depth, leading to suboptimal performance.
- Training Fluctuations: High fluctuations in accuracy and loss were observed during
training, indicating instability in the learning process.
● Improvements:
- Hyperparameter Tuning: Extensive experimentation was conducted to optimize
hyperparameters such as learning rate, batch size, and number of epochs to stabilize training
and improve convergence.
- Model Complexity: Additional convolutional layers were added to the ResNet50
architecture to increase model complexity and capture more intricate features.

Fig2. Version1 Training, Validation Accuracy and Loss

8
 Fig3. Version 1 ROC curve (Home Dataset)
● Findings:
Performance Metrics:
ROC-AUC (Home Dataset): .>90%
Training Accuracy: 86%
Validation Accuracy: 83%
Training Loss: 0.38
Validation Loss: 0.45

5.2 Model Version 2:

● Architecture: Similar to Version 1, with the addition of minimal data augmentation
using techniques such as rotation, shifting, and flipping.
● Training Process: The model is trained with augmented data generated by
ImageDataGenerator, with the same base architecture as Version 1 but a lower learning rate of
0.0001.
● Performance Metrics: Achieves improved test accuracy of around 90%, demonstrating
better generalization compared to Version 1.
● Challenges Faced:

9
- Generalization Issue: When evaluating the model on a new dataset (ICDAR2011),
poor performance was observed, indicating a lack of generalization beyond the Cedar
Dataset.
● Improvements:
- Fine-tuning Hyperparameters: Further adjustments were made to hyperparameters,
including batch size, learning rate, and epoch count, to optimize model performance and
address the fluctuation issue.

Fig4. Version2 Training, Validation Accuracy and Loss

Fig 5. Version 2 Test Accuracy and Loss

10
 Fig 6. Version 2 ROC curve (ICDAR2011)

● Findings:
Performance Metrics:
ROC-AUC (Home Dataset): 0.85
ROC-AUC (ICDAR2011 Dataset): 0.65
Training Accuracy: 90%
Validation Accuracy: 89.77%
Training Loss: 0.22
Validation Loss: 0.30

Model Version 3:
● Architecture: Retains the base ResNet50 model with custom top layers, along with
further enhancements in data augmentation.
● Training Process: Utilizes extensive data augmentation and hyperparameter tuning to
improve model robustness and convergence.
● Challenges Faced: Increased training time due to augmented data and complexity,
requiring optimization of hyperparameters.
● Performance Metrics: Achieves the highest test accuracy of approximately 92% and
demonstrates improved performance on diverse datasets.

11
● Challenges Faced:
- Training Time: With the introduction of data augmentation and increased model
complexity, training time significantly increased, posing a challenge for scalability.
● Improvements:
- Augmented Data: Leveraging augmented data, the model was trained on a more
diverse dataset, enabling it to learn robust features that could handle variations present in the
Cedar Dataset.
- Optimized Hyperparameters: Hyperparameters were further fine-tuned to strike a
balance between model performance and training efficiency, addressing the challenge of
increased training time.

Fig 7 . Version3 Training, Validation Accuracy

Fig 8. Version 3 Test Accuracy and Loss

12
 Fig 9. Version 3 ROC curve

● Findings:
Performance Metrics:
ROC-AUC (Home Dataset): 0.88
Training Accuracy: 92%
Validation Accuracy: 91.23%
Training Loss: 0.18
Validation Loss: 0.25
For Model Version 3 on a Different Dataset:
ROC-AUC (Unseen Dataset): 0.69
False Positive Rate (FPR): 0.341
True Positive Rate (TPR): 0.730

13
 CHAPTER 6
SYSTEM IMPLEMENTATION

______________________________________________________________________

6.1 Overview of the system architecture and components

The Signature Verification System is architectured as a client-server model, with distinct

front-end and back-end components. This architecture enables efficient communication
between the user interface and the computational components responsible for signature
verification.

Front-end User Interface: The front-end component encompasses the user interface accessible
through a web browser. Developed using HTML, CSS, and JavaScript, the interface provides
a user-friendly platform for interaction. It includes forms for user authentication, signature
upload functionalities, and result display sections. This component is responsible for handling
user inputs and presenting verification outcomes in an intuitive manner.

Back-end Server: The back-end server serves as the computational engine of the system,
processing user requests and orchestrating the signature verification process. Implemented
using Django, a Python web framework, the server manages user authentication, signature
comparison requests, and communication with the CNN model. It handles data preprocessing
tasks, such as image normalization and feature extraction, before forwarding the processed
data to the CNN model for verification.

CNN Model: The Convolutional Neural Network (CNN) model forms the core component of
the system, responsible for authenticating handwritten signatures. Developed using
TensorFlow or PyTorch, the model is trained on a diverse dataset of genuine and forged
signatures. It learns to extract relevant features from signature images and classify them as
either genuine or forged. The CNN model is deployed on the server and invoked by the
back-end component upon receiving signature comparison requests.

14
6.2 Description of the User Interface Design and Functionalities

The user interface of the Signature Verification System is meticulously designed to ensure
ease of use and seamless navigation. It comprises several key functionalities tailored to
facilitate user interaction and streamline the verification process.

● Signature Upload: Users are directed to the signature verification module, where they
can upload image. The interface provides intuitive file upload functionalities, allowing users
to select images from their local storage.. Clear instructions and visual cues guide users
through the upload process to ensure proper alignment and quality of the signature images.
● Real-time Feedback: During the verification process, the interface provides real-time
feedback on the authenticity of the signatures. Upon initiating the verification, users receive
immediate feedback on whether the signatures match or not. Additionally, the interface
displays a confidence score or probability of the match, providing users with insights into the
reliability of the verification outcome. Detailed results and analysis are also available for
users to review, enhancing transparency and trust in the system.

6.3 Integration of the CNN Model into the Signature Verification System

The integration of the CNN model into the Signature Verification System is meticulously
orchestrated to ensure seamless operation and efficient utilization of computational resources.

● Deployment: The trained CNN model is deployed on the server environment using
Django frameworks. This deployment ensures that the model is readily accessible to the
back-end component for inference tasks. The deployment process involves configuring the
server environment, optimizing model performance, and establishing communication
channels with the back-end server.
● Communication: The back-end server communicates with the deployed CNN model
through well-defined APIs or network protocols. Upon receiving signature comparison
requests from the user interface, the server preprocesses the input images and forwards them
to the CNN model for verification. The model generates predictions on the authenticity of the
signatures, which are then relayed back to the server for further processing.
● Scalability and Modularity The integration is designed to be modular and scalable,
allowing for future enhancements and updates to the CNN model. As new signature samples
are collected and the model undergoes further training, the system can seamlessly incorporate
improvements to enhance its accuracy and effectiveness in combating forgery and

15
falsification attempts. Moreover, the modular architecture facilitates the integration of
additional machine learning models or algorithms to augment the system's capabilities in the
future.

By carefully orchestrating the integration of the CNN model into the Signature Verification
System, the system achieves robust and reliable performance in authenticating handwritten
signatures, thereby enhancing security measures and mitigating the risks associated with
document forgery and falsification.

Hardware Used:
Model Training:
GPU used for training: NVIDIA T4 X2
GPU RAM: 15 +15 GB
CPU Memory: 29GB (MAX)
Platform: Kaggle

16
 CHAPTER 7
RESULTS AND DISCUSSION

______________________________________________________________________

7.1 Model Training and Evaluation Results:

Model Performance Metrics: The performance of each model version was evaluated using
standard metrics such as accuracy, loss, and area under the receiver operating characteristic
curve (ROC-AUC). These metrics provide insights into the model's classification
performance and generalization capabilities.

Training History: The training history, including changes in accuracy and loss over epochs,
was visualized to understand the model's learning process and identify any trends or patterns.

7.2 Performance Analysis of Model Versions:

● Version 1: The initial model achieved moderate performance metrics, with decent
accuracy on the training and validation sets. However, it exhibited limitations in
generalization beyond the training dataset.
● Version 2: By implementing data augmentation techniques and fine-tuning
hyperparameters, the second model showed improvements in both training and validation
accuracy. However, it still struggled with generalization to new datasets.
● Version 3: The final model, leveraging augmented data and optimized
hyperparameters, demonstrated the highest accuracy and improved generalization compared
to previous versions. It achieved the best performance metrics on both the training and
validation sets.

7.3 Implications and Areas for Improvement:

Generalization: Despite improvements in model performance, challenges remain in achieving

robust generalization to new datasets. Further exploration of transfer learning techniques and
domain adaptation methods could enhance the model's ability to handle variations in
signature images from different sources.

Scalability: As the system scales to handle larger datasets and increased user traffic,
optimizations in model training and inference speed will be crucial. Techniques such as

17
distributed training and model compression can help address scalability challenges without
compromising performance.

User Feedback: Incorporating user feedback and iterative testing cycles can provide valuable
insights into the usability and effectiveness of the Signature Verification System. Continuous
improvement based on user input can lead to a more refined and user-centric application.

18
 CHAPTER 8
DISCUSSION AND CONCLUSION

______________________________________________________________________

8.1 Conclusion:
The development of the Signature Verification System represents a significant advancement
in automated signature authentication, offering enhanced accuracy and efficiency compared
to traditional methods. Through the implementation of deep learning models and web-based
interfaces, the system provides users with a seamless and intuitive platform for signature
verification tasks.

8.2 Key Achievements:

Model Performance: The implemented ResNet50-based models exhibited promising
performance metrics, achieving high accuracy and robustness in signature verification tasks.
User Interface Design: The user interface design of the web application offers a user-friendly
experience, incorporating features for user authentication and signature verification with ease.

8.3 Future Work:

● Enhanced Generalization: Further research and development efforts will focus on
improving the generalization capabilities of the models, enabling them to effectively handle
variations in signature images from diverse sources.
● Scalability and Efficiency: Optimization techniques will be explored to enhance the
scalability and efficiency of the system, allowing for seamless operation with larger datasets
and increased user traffic.
● Integration of Advanced Techniques: Integration of advanced techniques such as
ensemble learning, attention mechanisms, and domain adaptation will be explored to further
enhance the performance and robustness of the system.
● User Feedback and Iterative Improvement: Continuous solicitation of user feedback
and iterative testing cycles will guide the refinement and enhancement of the Signature
Verification System, ensuring its effectiveness and usability in real-world scenarios.
In conclusion, the Signature Verification System represents a promising solution for
automated signature authentication, with the potential for further advancements and
improvements through ongoing research and development efforts.

19
 LIST OF PUBLICATIONS

______________________________________________________________________

● Automatic Signature Verification: The State of the Art

Authors: Donato Impedovo, Giuseppe Pirlo
Published in: IEEE TRANSACTIONS ON SYSTEMS, MAN, AND
CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 38, NO. 5,
SEPTEMBER 2008
Link:
https://d1wqtxts1xzle7.cloudfront.net/66035604/Impedovo08-libre.pdf?1615924555=&respo
nse-content-disposition=inline%3B+filename%3DAutomatic_Signature_Verification_The_St
a.pdf&Expires=1713479440&Signature=HNpzu93c4RHZxb1HhoUaiyMRE9jxuDVQaD0Z
QiedkV2YMcBNegPI62-z0BcF~DMg~-5FqPLk54jEz0tN5VxBSBIuG3BZwvnNvJbimWr2
TUl297hmIk7OaEpOtDuc0arl7u5XV~PD0iFQB20CoGW3ZWIDiCoMGAyuKloLyG235fW
FdmmWygHLcuHAqk7C4UUBswk4sFDMagtpI5YUDtDVH5V-kRU5205X3rXaWN~Eqx4
oYmeBaSFMeWuJ-GQ5LINIvjGfqmOgjgp4vGFA2gsQctpUaBIbuu9CueYnRBsqXCMlMjv
lJCLSo6scwXlrb-iUkXy3DSStSaTG5L0n9ZoX2w__&Key-Pair-Id=APKAJLOHF5GGSLR
BV4ZA
● Signature verification: A study
Authors: Saba Mushtaq, Ajaz H Mir
Published in: 2013 4th International Conference on Computer and Communication
Technology (ICCCT)
Link: https://www.researchgate.net/publication/271551455_Signature_verification_A_study
● Offline Signature Identification and Verification Based on Capsule Representations
Authors: Dilara Gumushas, Tulay Yildirim
Published in: BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND
INFORMATRONOV 28, NeSofia 2020
Link:
https://github.com/anupdhoble/6thSemProject_SignatureVerification/blob/main/Research_Pa
pers/BulgarianInstitueOfSciences_OfflineSignatureRecogniation.pdf
● HUMAN SIGNATURE VERIFICATION USING CNN WITH TENSORFLOW
Author: Dr.It. Pabon Kennedy, Nithish Kumar, Vinodh Kanar. Yogend
Published in: INTERNATIONAL JOURNAL OF CREATIVE RESEARCH THOUGHTS
(IJCRT) An International Open Access. Peer-reviewed, Refereed Journal
Link:
https://github.com/anupdhoble/6thSemProject_SignatureVerification/blob/main/Research_Pa
pers/rp1.pdf

20
 REFERENCES

______________________________________________________________________

[1] CEDAR Signature Introduced by Srinivasan et al. in Machine learning for signature
verification CEDAR Signature is a database of off-line signatures for signature verification.
https://paperswithcode.com/dataset/cedar-signature

[2] Signature Verification System Engineering/Diploma/Bsc-IT/Msc-IT Projects, IT Projects

https://nevonprojects.com/signature-verification-system/

[3] "Ahmetozlu / signature_extractor” super lightweight image processing algorithm for

detection and extraction of overlapped handwritten signatures on scanned documents using
OpenCV and scikit-image.https://github.com/topics/signature-verification

[4] Digital Image Processing Basics | geeksforgeeks

https://www.geeksforgeeks.org/digital-image-processing-basics/

[4] GFG | Introduction to convolution neural networks

https://www.geeksforgeeks.org/introduction-convolution-neural-network/

[5] IBM Explainer https://www.ibm.com/topics/convolutional-neural-networks\

[6] Anamika Jain, Satish Kumar Singh, Krishna Pratap Singh Figure - available from:
Multimedia Tools and Applications
https://www.researchgate.net/figure/Accuracy-vs-Data-Distribution-for-sCNN_fig6_3404905
26

[7] Zihan Zeng1 and Jing Tian “Deep Learning Methods for Signature Verification”
https://arxiv.org/pdf/1912.05435.pdf\

[8] Teressa Longjam a, Dakshina Ranjan Kisku b, Phalguni Gupta Writer independent
handwritten signature verification on multi-scripted signatures using hybrid CNN-BiLSTM:
A novel approach
https://www.sciencedirect.com/science/article/abs/pii/S0957417422021297

21
 APPENDICES

______________________________________________________________________

Appendix A: Three Types of Layers in Convolutional Neural Networks (CNNs)

1. Convolutional Layer:

● Explanation of convolution operation and filters.

● Role of convolutional layers in extracting features from input images.
● Discussion on parameters such as filter size, padding, and stride.

2. Pooling Layer:

● Description of pooling operations such as max pooling and average pooling.

● Purpose of pooling layers in downsampling feature maps and reducing computational
complexity.
● Considerations for selecting pooling size and stride.

3. Fully Connected Layer:

● Role of fully connected layers in connecting every neuron from one layer to another.
● Explanation of activation functions such as ReLU, sigmoid, and tanh.
● Importance of fully connected layers in producing final classification outputs.

Appendix B: Project Procedure

1. Data Collection:

● Detailed methodology for acquiring a diverse dataset of signature images.

● Considerations for ensuring dataset representativeness and diversity.

2. Data Preprocessing:

● Steps involved in data normalization, grayscale conversion, and augmentation.

● Techniques for handling data imbalances and enhancing dataset quality.

3. Model Architecture:

● Design principles for optimizing the architecture of a CNN.

22
● Strategies for selecting appropriate network depth, width, and complexity.

4. Model Training:

● Process for training, validating, and evaluating the CNN model.

● Discussion on hyperparameter tuning, loss functions, and optimization algorithms.

5. Verification System:

● Overview of the user interface (UI) design and functionality.

● Integration of the trained CNN model into the verification system.
● Implementation details of signature matching functionality.

6. Deployment & Testing:

● Steps involved in deploying the verification system.

● Procedures for conducting testing, including unit testing and user acceptance testing.
● Methods for gathering user feedback and incorporating improvements.

Appendix C: Sample Signature Images

● Sample images from the acquired dataset, showcasing both genuine and forged
signatures.
● Image preprocessing examples, illustrating normalization, grayscale conversion, and
augmentation techniques applied to the dataset.

Appendix D: System Documentation

● Detailed documentation of the Signature Verification System, including system

architecture, component descriptions, and installation instructions.
● User manual providing guidance on system usage, authentication procedures, and
troubleshooting.

Appendix E: Experimental Results

● Tabulated results of model performance metrics, including accuracy, precision, recall,

and F1-score.
● Graphical representations of training and validation loss curves, demonstrating model
convergence and performance trends.

23