Signature Verification and Detection

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/371125675
Project Report on Signature Veriﬁcation and Detection by Image Processing
Technical Report · May 2023

DOI: 10.13140/RG.2.2.24971.87846
CITATIONS READS
0 6,117
4 authors:
Alimul Rajee Sumaiya Zafrin

Comilla University Comilla University
38 PUBLICATIONS 17 CITATIONS 1 PUBLICATION 0 CITATIONS
SEE PROFILE SEE PROFILE
Ahad Hossain Shuvo Tasrina Parvin

Comilla University Comilla University
2 PUBLICATIONS 0 CITATIONS 4 PUBLICATIONS 1 CITATION
SEE PROFILE SEE PROFILE
All content following this page was uploaded by Sumaiya Zafrin on 29 May 2023.
The user has requested enhancement of the downloaded file.

Project Report on
Signature Verification and Detection by Image Processing
Group - I
Submitted To Submitted By
AlimulRajee MdFarhanIshrak 11909009
Lecturer Tasrina Parvin 11909014
Dept. of Information and Communication Ahad Hossain 11909022

Technology
SumaiyaZafrin 11909027
Kotbari ,Comilla University. TasmiaraJahanToma 11909048
Submission Date: May 2023

Acknowledgment
First, we would like to sincerely thank AlimulRajee sir, our supervisor, for his unwavering
support of our project report work and associated research, as well as for his patience,
inspiration, and vast expertise. His advice was helpful to us throughout the entire project and
report-writing process. For our project, we could not have asked for a greater mentor and
advisor.
We also want to thank to our course teacher AlimulRajee sir for his encouraging words and
perceptive criticism.
Last but not least, I would want to thank our family, friends, and parents for their spiritual
support while we were writing our report and living our lives in general.
1
Table of Content
Chapter Title Page

Acknowledgment 5
Abstract 6
1 Introduction
1.1 Background Study 7-10
1.2 Aims and Objectives 10
1.3 Outline of This Report 10-11
2 Literature Review 12-19
3 Motivation 20-23
3.1 Introduction 21
3.2 Signature Database 21-22
3.3 Preprocessing 22
3.4 Feature Extraction 23-24
4 Methodology 25-39
4.1 Introduction 25
4.2 Data Collection 26-27
4.2.1 Data Source 26
4.2.2 Signature Types 26
4.2.3 Number of Samples 26
4.2.4 Signature Variations 27
4.2.5 Metadata 27
2
4.2.6 Data Quality 27
4.2.7 Data Security 27
4.3 Pre-Processing 28-29
4.3.1 Image Resizing 28
4.3.2 Image Filtering 28
4.3.3 Binarization 28
4.3.4 Segmentation 28
4.3.5 Normalization 29
4.3.6 Feature Extraction 29
4.3.7 Data Augmentation 29
4.4 Feature Extraction 29-31
4.4.1 Stroke-based Feature 30
4.4.2 Shape-based Feature 30
4.4.3 Texture-based Features 30
4.4.4 Global-Features 30
4.4.5 Local-Features 30
4.4.6 Deep Learning Features 31
4.5 Signature Detection 31-32
4.5.1 Edge Detection 31
4.5.2 Template Matching 31
4.5.3 Hough Transform 32
3
4.5.4 Connected Component Analysis 32
4.5.5 Machine Learning Based Technique 32
4.6 Signature Verification 33-34
4.6.1 Feature Matching 33
4.6.2 Neural Networks 33
4.6.3 Support Vector Machine 33
4.6.4 Dynamic Time Warping 33
4.6.5 Graphology 34
4.6.6 Hybrid Techniques 34
4.7 Types of Forgery 34
4.8 Algorithm Details 35-39
4.8.1 Support Vector Machines 35-36
4.8.2 Convolutional Neural Networks 36-37
4.8.3 Siamese Networks 38-40
5 Performance Measurements 41
6 Result and Discussion 42-46
6.1 Results 42-45
6.2 Discussion 46
7 Conclusion and Future Work 47
References 48-50
4
List of Figure
No Title Page
1.1 Module of Signature Verification System 12
4.1 The Block Diagram of Signature Detection and 27
VerificationMethodology.
5.1 Genuine Signature 54
6.1 Verified Image 44

6.2 Unverified Image 45
6.3 User Interface 45
5
List of Table
No. Title Page
2.1 Different Algorithm and Dataset with Result and 21

Findings.
6.1 Comparison of Various Offline Signature 53

Verification Techniques.
6
List of Abbreviation
Name Abbreviation
CNN Convolutional Neural Networks
DDF Directional Distance Function
DTW Dynamic Time Wrapping
FAR False Acceptance Rate
FFR False Rejection Rate
GLCM Gray Level Co-occurrence Matrix
HMM Hidden Markov Model
HOG Histogram of Oriented Gradients
HSV Handwritten Signature Verification
KNN K-Nearest Neighbour
MLP Multilayer Perceptron
NN Neural Network
PCA Principle Component Analysis
SVM Support Vector Machine

RBF Radial Basis Function
ReLU Rectified Linear Unit
7
Abstract
Signature verification and detection project report focuses on the development of a system for
signature detection and verification using image processing techniques. The goal of the project is
to design a system that can accurately detect and verify signatures on various types of
documents. The system includes several modules such as image preprocessing, feature
extraction, signature detection, and verification. The image preprocessing module is responsible
for enhancing the quality of the input image by removing noise and other artifacts. The feature
extraction module extracts feature from the signature image using various techniques such as
Fourier Transform and Wavelet Transform. The signature detection module uses these features
to detect the presence of a signature in the image. Finally, the signature verification module
verifies the authenticity of the signature by comparing it with the reference signature. The system
which was proposed has been evaluated using a dataset of signature images, and the results show
that it can accurately detect and verify signatures with a high degree of accuracy. The system has
several potential applications in the field of document verification and authentication.Moreover,
we also explored the use of deep learning techniques, specifically convolutional neural networks
(CNNs), for signature detection and verification. We trained a CNN model on a large dataset of
signature images and compared its performance with the traditional approach. Our results show
that the CNN-based approach outperforms the traditional approach, achieving an average
accuracy of 99.5% to 98.6% in signature detection and verification tasks, respectively.In
conclusion, our proposed approach for signature detection and verification in digital images is
effective and can be applied in various practical applications. Additionally, the use of deep
learning techniques can further improve the performance of the approach.
8
Chapter - 1
Introduction
1.1 Background Study

Signature verification is the process of using a digital signature algorithm and a public key to
verify a digital signature on data. It is a form of identity verification.Banks, intelligence services,
and other prestigious institutions employ signature verification to confirm a person's
identification. In bank branches and other branch capture, signature comparison is frequently
employed. The signature verification software compares a direct signature or a picture of a
signature to the recorded signature image. The signature serves as the authority for all legal
transactions. Thus, the necessity for signature verification grows. It is distinct for the handwritten
signatures toindividuals and that cannot be duplicated.In addition to being a well-liked area of
research in the fields of pattern recognition and image processing, signature verification also
plays a significant role in numerous applications, including access control, security, and privacy.
The process of certifying someone based on their handwritten signature is known as signature
verification. Systems for verifying signatures come in two varieties [1].
Online Signature Verification System, which records details like pressure, speed, direction, etc.
using an electronic device like a tablet. Offline Signature Verification System, in which the
signature is written offline and verified using the image of the signature that has previously been
stored. Two distinct methods can be used to verify offline signatures.One involves building
models of real and fake signatures for each writer in a process known as writer dependent
signature verification. Next, a writer's test signature sample is contrasted with its own training
sample. This method's downside is that each new writer must have a model created in order to be
confirmed. Forensic professionals utilize the second method, known as writer-independent
signature verification and accuracy is 84% .[2]
This technique is seen as the most practical cases, since it is not essential to develop a model for
each writer in order to verify its signature. In this instance, a broad model is created using a few
writers selected at random. Yet, due to the significant morphological variation across writers,
writer-independent signature verification is a more challenging issue.[3]
9
A lot of research has been done on signature verification, and it is currently being looked upon,
especially in the offline mode. On-line signature verification has shown substantially higher
verification rates than off-line verification since a lot of dynamic information is not available in
the off-line mode. Online signature verification is therefore frequently more successful. The
majority of the time, signatures denote a writing style rather than a list of alphabets, letters, and
words [4,5,6].
A person's signature frequently alters in response to stress, emotion, time, fatigue, etc. In
signatures are shown to be inconsistent even in accordance with habits,country, physical and
psychological conditions.The resolution of five sub-problems, including data collecting, pre-
processing, feature extraction, comparison process, and performance evaluation, is often
necessary for the design of any signature verification system.Studies on the cutting-edge off-line
signature verification systems up to 1993 are mentioned in [1,2, and 3]. The methods used for
off-line signature verification from 1993 to 2000 were detailed in another survey paper. Here, we
review a few works in this area that weren't discussed in the survey articles. These publications
mostly focus on fuzzy-based off-line signature verification approaches.[7.8]
J.A.Unar et.al propossed for an offline signature system including signature recognition and
verificationS is carried out using fuzzy concepts while the recognition phase is based on the
multi-stage classifier and a combination of global and local features. The HMM-based approach
in determines the author dependent parameters dynamically and automatically to set up an ideal
decision rule for the off-line verification process. Here, not only are the best HMM models
derived using the cross validation method, but also an optimal model.[6]
The pseudo-bacterial genetic algorithm (PBGA), which was used to find fuzzy rules, is also used
to try and verify signatures. The performance of a fuzzy system is acquired synergistically as the
sum of the outputs of numerous rules, even though the rules are units themselves and are made
up of several parameters that need to be tuned. The extraction of personal features for signature
verification was then done using the PBGA.
The signature verification method in is driven by a fuzzy neural network based on a pseudo-outer
product. This method is mostly used to check expert forgeries. The use of the TS model for
signature verification is reported in [PI Key characteristics for this model are taken from [SI's]
10
box method. In this study, we use the same features as in [SI], but we modify the TS model to
improve its forgery detection capabilities.[7]
It wasn't until the introduction of computers in the 1960s that automatic evaluation of disputed
signatures became possible. Designing an automatic forgery detection system became a popular
area of research as computers got more powerful and more accessible. Yet, the majority of the
work in offline forgery detection has experience with simulated forgeries as well as less
competent or random forgeries. We briefly discuss the many sorts of forgeries before examining
the seminal contributions to the field of forgery detection.
Based on their distinguishing characteristics, handwritten signature forgeries have been divided
into different categories.‘The following forms of signature forgeries can be generally
categorized: 1. Random Forgery - The signer creates a forgery known as "the simple forgery" or
"random forgery" by using the victim's name in his own unique manner. 2. Unskilled Forgery -
The signer imitates the signature in his own manner without prior experience or knowledge of
the spelling. 3. Expert Forgery – Without a doubt, the most challenging forgeries are produced
by experienced forgers or professional impostors.[2]
Ammar et.alworked on the detection of competent forgeries in the 1980s. They analyzed the
statistics of dark pixels and used them to spot shifts in the writing's overall flow. Using reference
patterns, specifically the horizontal and vertical placements of the trademark image, Ammar's
following work is based . Euclidean distance is used to compare the projection of the disputed
signature and the reference. An algorithm for the detection of skilled forgeries has been
presented by Guo et al. and is based on a local correlation between a signature under scrutiny and
a model created a priori. At the sub-stroke level, writer-dependent characteristics are assessed,
and a cost function is learned for each writer. Before features are recovered from each of the
original scanned signatures, size normalization, binarization, and thinning are used as pre-
processing steps.[11,12]
Before features are recovered from each of the original scanned signatures, size normalization,
binarization, and thinning are used as pre-processing steps. These qualities make up the
knowledge base that is later used for authenticating signatures and spotting forgeries. We now
provide a quick explanation of the system's several processes for signature verification.
11
Based on their distinguishing characteristics, handwritten signature forgeries have been divided
into different categories [f2]. The following forms of signature forgeries can be generally
categorized:
1. Random Forgery - The signer creates a forgery known as "the simple forgery" or
"random forgery" by using the victim's name in his own unique manner.
2. Unskilled Forgery - The signer imitates the signature in his own manner without prior
experience or knowledge of the spelling.
3. Expert Forgery - Without a doubt, the most challenging forgeries are produced by
experienced forgers or professional impostors. Ammar et.al [I0] worked on the detection of
competent forgeries in the 1980s. They analyzed the statistics of dark pixels and used them to
spot shifts in the writing's overall flow.
1.2 Aims and Objectives
The purpose of this thesis is to suggest a method for online handwritten signature verification.
1. Collecting some offline signatures and take a picture of them.

2. Then it needs to be sent to our pc.
3. Then we have derived set of features and from model based classifier.
4. Measuring the similarity of the signature.
5. Then we need to determine the performance of signature verification.
6. Evaluation of performance by comparing it with most widely used technique called
DTW.
1.3 Outline Of This Report
Chapter 1: In this section, we cover the report' primary goal, as well as its motivation, goals and
objectives, and outline of the report.
Chapter 2: This chapter addresses the topic of the literature review.
12
Chapter 3: The background information necessary for a thorough grasp of the report is outlined
in depth, step by step, in this chapter.
Chapter 4: The proposed methodology is thoroughly presented in this chapter.
Chapter 5: The performance evaluation components are thoroughly covered in this chapter.
Chapter 6: In this chapter, the findings and the solutions to the study questions are thoroughly
discussed.
Chapter 7: This chapter discusses the chapter's conclusion and its future direction.
13
Chapter - 2
Literature Review
A literature review is an examination of academic sources on a particular subject. It provides an

overview of current knowledge, helping you to find applicable theories, methodologies, and gaps
in the existing research that you may later apply to your paper, thesis, or dissertation topic. For
many years, handwriting analysis and pattern matching have been the subject of active research.
Several technologies have been applied to Handwriting Signature Verification (HSV),
particularly offline HSV, and the field is continuously being researched. We have examined
several recent papers on offline HSV in this part.
The handwritten signature verification methods and the methodologies used have been the
subject of numerous surveys. Using HMM, SVM, FFT, MLP wavelets, and NN, numerous
methods have recently been presented and extensive study has been done on feature extraction
and classification. Holistic matching, regional matching, and multiple regional matching are
some of the matching techniques used in signature analysis. Euclidean distance, elastic matching,
regional correlation, tree matching, relaxation matching, split and merge, string matching, NN,
HMM, and SVM are a few of the approaches with the most varied applications that have been
described and the problems and difficulties that the signature verification system faces are
explored where accuracy is observed 94%.[2]
Comparative research was done by Hansheng Lei and VenuGovindaraju on aspects that are often
used. A consistency model is created to measure the distances-based measure by generalizing the
already-existing feature-based measure. It was discovered that uniformly re-sampling the
sequences does not always improve verification performance and that the simple features, such
as X- and Y-coordinates, writing speed, and angle with the X-axis, are among the most
consistent and the rate for identifying original signal is 93%.[3]
Dr. Maged and M. M. Fahmy introduced a system for online handwritten signature verification
that is based on discrete wavelet transforms (DWT) feature extraction and classification using
feed-forward back error neural networks [7]. The signature is validated in the DWT domain to
14
increase the distinction between a real signature and a fake.A multi-matcher, which matches for
the same input biometric signal using several representations and six neural networks, is used to
validate signatures. A discussion and comparison of the recognition rates for each of these neural
network recognizers is conducted. Twenty authentic signatures and twenty expertly forged
signatures are tested on five users of the signature database. The success rate for identifying
genuine signatures is 95%.[5]
Christian Gruber et al. put a novel approach to online signature verification using support vector
machines that is based on the LCSS kernel function [8]. Here, the length of an LCSS is
calculated using a kernel function to compare the two time series. It has been demonstrated that
the SVM LCSS can reliably authenticate people with just six real signatures. The similarity
assessment of online signature data based on LCSS turned out to be even better than DTW-based
methods [9].To validate online signatures, Abhishek Sharma and Suresh Sundaram have
introduced a novel model-based technique called GMM within the DTW framework [10].
For signature matching, they first retrieved the writer dependent statistical characteristics. Then,
using a derivation in a warping path-based feature that is useful for verification, the properties of
a warping path are studied.A novel approach to online signature verification using support vector
machines that is based on the LCSS kernel function was put forth by Christian Gruber, Thiemo
Gruber et.al. Here, the length of an LCSS is calculated using a kernel function to compare the
two-time series.The kind of characteristics retrieved, the training process, and the classification
and verification models employed vary amongst research methods.[8]
Hidden Markov Model (HMM) was presented by J. K. Guoetal.Each point in a handwritten

signature's journey was represented by a series of vectors of values. An effective signature
verification system might be created using the HMM's specified set of feature vectors. These
models were stochastic ones that could take into account both the differences and similarities
across patterns. Stochastic matching of the model and the signature was used in HMM. The
process of matching was carried out through steps of the probability distribution of the features
used in the signatures or the probability used to generate the original signature.[1]
Gabriel raised the issue of training online signature verification systems when the number of
training samples is small and the number of available signatures per user is constrained . Nine
distinct GMM- and UBM-based classification techniques are assessed. These models are created
15
to function in situations with a small sample size that were examined in three independent
studies. Because the training set only contained a portion of the samples, these approaches'
performance fell down more quickly (around 12 signatures per user). By computing the
likelihood ratio and comparing it to the EER decision threshold, the decision was determined.
When the available training subset is at least 50% of the total, the GMM-SVM models' accuracy
is noticeably better than the GMM-UBM models'.
Thomas Hassan-Reza and Beatrice Drott proposed a method for classifying fake and real
signatures that uses binary classification, first with straightforward engineered features, then with
machine learning methods like logistic regression and MLP, and finally with a deep learning
method using a convolutional neural network. The problem of signature verification using deep
learning revealed encouraging results, although they could be better [14].
J. A. Unar et al. proposed Neural networks (NNs) were widely used in pattern recognition due
mostly to their strength and usability. A straightforward method was to first take numerous
samples from various signers and extract a feature set that represents the signature (information
like length, height, duration, etc.). The NN must then understand the connection between a
signature and its class (either "genuine" or "forgery") in order to proceed to the next stage. Once
the network was aware of this link, test signatures that can be identified as coming from a certain
signer can be submitted to it. NNs were thus well adapted to modeling the overall characteristics
of handwritten signatures. The suggested approach makes advantage of new features, such as
surface area, as well as structure features from the signature contour and modified direction
feature.[2]
Template matching was presented by Fang et al. as a means of detecting competent forgeries.
The elastic matching of the strokes in the two-dimensional signature patterns serves as the
foundation for the best matching of the one-dimensional projection profiles of the signature
patterns and the other. The positional variations of a test signature are compared to the statistics
of the training set, and a choice based on a distance measure is determined. We test both binary
and gray-level signature photos. When the local peaks of the vertical projection profiles of the
gray-level signature images were employed for matching and the entire predicted covariance
matrix was taken into account, the average verification error rate of 18.1% was attained.[5]
16
It was simple to determine the relationship, deviation, etc. between two or more data pieces using
statistical expertise. We typically used the idea of correlation coefficients to determine the
relationship between a given set of data points. Refers, generally speaking, to the breaking of two
variables' independence. This method used the correlation principle to determine the degree of
divergence between the entered signature and the average signature, which was derived from the
collection of previously gathered signatures. In this method, different features are retrieved, such
as statistical information derived from the distribution of pixels in a signature, global features
like picture gradient etc. [14]
Griechisch E et al. showed in machine learning, vector machines (SVMs) used a high-
dimensional feature space to estimate disparities across data groups to generalize previously
unknown data. For categorization and verification, the system made advantage of the signatures
global, directional, and grid properties. The 1320 signatures from 70 writers are taken from a
database. A total of 320 signatures were used for training, with 40 writers signing 8 signatures
apiece. With 8 authentic signatures and 8 forgeries, the method obtained FRR 2% and FAR 11%
during early testing.
A signature detection and classification system using wavelet transforms and support vector
machines was proposed. The authors assert that their system had excellent accuracy rates, and
they used trials on a dataset of 1000 signatures to show how successful it was. Using two
publicly accessible datasets, they asserted that their system performs at the cutting edge, and they
contrast their findings with those of alternative methods.This study suggested a fuzzy logic and
evolutionary algorithm-based method for automatic signature verification and forgery detection.
The authors asserted that their approach achieved high rates of accuracy and could precisely
identify frauds. Using a dataset of 1000 signatures, they showed the efficiency of their
approach.[4]
El-Sayed M.El-Horbaty et al. suggested a texture analysis and neural network-based system for
automatic signature verification. They asserted that both online and offline signatures could be
verified using their technology, which they said had high accuracy rates. Using a dataset of 400
signatures, they showed the efficiency of their approach.This study offered a thorough analysis
of signature verification methods. The authors discuss many methods for verifying signatures,
including statistical, structural, and behavioral techniques. They also compared various
17
techniques based on their effectiveness and intricacy.A person could be recognized offline
exclusively through their signature. Almost all documents, including those for financial services,
legal services, letters, checks, security documents, etc., require a signature. Forgery and other
illegal activities were a major worry in our daily lives. Forgery results from negligence in
signature verification. The method described in this study was new for offline signature
verification. The signature was kept manually on paper, obtained using a scanner or camera, and
stored in an image format in this project's signature verification process. The suggested method
relies on geometrical and statistical extraction of qualities, followed by the database where the
attributes were taught, to verify the signature.The suggested method relies on geometrical and
statistical extraction of qualities, followed by the database where the attributes were taught, to
verify the signature. This established whether the signature was real or fabricated by comparing
the reference signature to the investigation signature. This paper employed statistical techniques
while proposing an algorithmic strategy for the verification of handwritten signatures. An
algorithm was used to determine a person's typical signature from a series of signatures. After
examining the correlation between the sample signature and the typical signature, the decision to
accept was made.This paper suggested a brand-new ensemble classifier-based technique for
verifying static handwritten signatures. By merging the output of the three simple classifiers, the
ensemble classifier that was being proposed increases accuracy. This approach computed the
Gabor coefficients in various scales and directions by convolving the signature image with
Gabor wavelets after the pre-processing stage. Three separate feature sets re extracted using
statistical methods employing the obtained Gabor coefficients.[5]
Offline signature verification using long short-term memory and histogram orientation gradient
was presented by Fatima SusilawatiMohamad et al. As a means of detecting signature. The
signing system was a typical step that organizations take to assure the confidently of their data or
datasets and to obvious it against unauthorized entrance or access. Long Short Term
Memory(LSTM) had been suggested, in this paper using input data from USTig and CEDR .The
actual Accuracy efficiency LSTM while using USTig was 92.4% when a run-time of 1.67
seconds and 87.7% for CEDAR in a run-time of 2.98 seconds.[6]
Aliya, et al. presented their work on multilingual handwritten signature recognition using high-
dimensional feature fusion. Due to the widespread use of handwritten signatures for recognition
18
and authentication in various tasks, including the authentication of documents and financial
transactions. Principal component analysis helped to decrease the high-dimensional features that
were controlled by a combination of certain techniques. (PCA). Utilizing the k-nearest neighbors
(k-NN) algorithm, the random forest method was used to match the segmentation. In comparison
to other methods, the presented method had an identification rate of 98.4% using a wide-ranging
signature database. Additionally, it demonstrated how well the technique worked and how it
could be used on sizable datasets of mixed, multilingual signatures. The dearth of works on
offline signature recognition of minority languages and the absence of a publicly accessible
dataset were addressed in this paper. First, a dataset including three languages—Han, Uyghur,
and Kazakh—was successfully constructed. A method for recognizing handwritten signatures
that is mostly based on local maximum occurrence characteristics (LOMO) has been suggested.
On the open dataset CEDAR, this approach achieved a recognition rate of 98.4%.[7]
R and Anagha presented their work on signature recognition and forgery detection.In their paper
titled "Offline Signature Recognition and Verification System utilizing Efficient Fuzzy Kohonen
Clustering Network (EFKCN) Algorithm," DewiSuryani, EdyIrwansyah, and RickiChindra
discuss the collection of images, data collection, image processing, normalization, clustering, and
evaluation. The pre-processing methods included border removal, bounding box extraction,
binary image conversion, binary image inversion, and RGB to grayscale format conversion.
Using this method, an accuracy of around 70% was reached. The pre-processing techniques
employed in TejasJadhav's study, "Handwritten Signature Verification using Local Binary
Pattern Features and KNN," were RGB to Gray Scale conversion, Otsu Thresholding, and
Boundary box cropping. The feature extraction techniques employed were: The feature
extraction techniques are LBP image creation, texture features, and name features. This
methodology combines Euclidian distance with KNN. This tactic's accuracy was 73.34%. The
pre-processing methods employed in the TansinJahan, Md. Shahriar Anwar, and S. M. Abdullah
Al-Mamun study titled "A Study on Preprocessing and Noise Reduction" are thresholding, edge
thinning, noise removal, and noise removal using adaptive filtering. Extraction of Features from
Offline Handwritten Signatures. Finding loops and turning the pixels of a signature into binary
values are two techniques for extracting features from it. In this study, the model was ran using
MATLAB. In their paper titled "An Offline Signature Verification and Forgery Detection
Method Based on a Single Known Sample and an Explainable Deep Learning Approach," Hsin-
19
Hsiung Kao and Che-Yen Wen used Deep Convolutional Neural Networks (DCNN) and
Explainable Deep Learning to implement the model. Rotation, block-based data augmentation,
and RGB to grayscale conversion are some of the pre-processing methods. This model's
accuracy is 94.37%.[8]
Offline Handwritten Signature Recognition was presented by G.A. Khuwaja. Neural Classifiers,.
LVQ Models algorithms were used in this project.
For a random test subset #1 of 10 individuals with 150 handwritten signatures, or 15 signatures
with different covered areas for each subject, the identification rate of the best network design is
98.7%; for subset #2, it is 94.7%. This implies that not all handwritten signatures have the same
average recognition rate, as some signatures resemble those in the second subset in terms of
shape. A training set of 120 signatures covering a range of the same subjects is used to train the
system. The experiment is repeated with 20 additional people, increasing the amount of test
signature images that are hidden from view. (300). the best network classifier's rate of accurate
classification drops to 92.7%. In order to conduct recognition on the input signature image, the
network classifier is trained on random training samples. According to empirical findings, a
network that has been trained with another set of 120 photographs of the same subjects has an
accuracy rate of 98% for a random test set of 150 handwritten signature images of 10 people. [9]
Pascual-Gaspar presented Off-Line Handwritten Signature Recognition by Wavelet Entropy and

Neural Network. We work with two databases in this project. The GPDS960 offline signature
database, or Database A, comprises information from 960 people. For each writer, there were 24
authentic signatures, and for each authentic signature, there were 30 expert forgeries created by
10 forgers from 10 distinct authentic specimens. The included signatures were 300 dpi quality,
greyscale, and in "bmp" format. Database B was a similar offline database containing
information from 20 different authors and 15 signatures per writer. By offering each author a 5-
minute session to complete a form that requires 15 signatures, this database was created..The LG
LSM-100 mouse scanner with a 300 dpi resolution was used as a portable scanner to later
digitize the signatures.In this paper, a signature identification method based on wavelet transform
average framing entropy (AFE) and probabilistic neural network (PNN) is proposed. The system
was tested using wavelet packet (WP) entropy and a discrete wavelet transform (DWT) entropy
and labeled as a WP entropy neural network system (WPENN) and a DWT entropy neural
20
network system, respectively. (DWENN). For comparison, a number of alternative techniques
found in the literature were taken into account. The algorithms were tested using two databases.
WPENN had the highest recognition rate result, with a threshold entropy of 92%.[10]
Table 2.1: Different Algorithm and Dataset with result and findings.
Data set Algorithm Result and findings
Through the following interface, the user

can select signature image of interest
from available database. Then train the
Digital Pixel Images CNN,SVM network with the content of this database
and 'matched ' or 'not matched' status
which indicates if the signature is exact
or forged.
For visual feature extraction, we

designed a Mutual Signature DenseNet
(MSDN), which extracts features in
different levels and combines them, by
learning from pairs of signatures. The
Pixel Image Surf Algorithm,SVM
network inputs either whole signature
image or local regions of signature. The
similarity scores of multiple local
regions are fused to make decision of
verification.
After applying a feature vector of test

signature if the output neuron generates
CNN,Random Sample
Format Images value close to +1 test signature is
Consensus
declared as genuine or if it generates
value close to -1 it is declared as forged.
21
The system is robust and can detect
random, simple and semi-skilled
forgeries but the performance
deteriorates in case of skilled forgeries.
Using a higher dimensional feature
Pixel Image, space and also incorporating dynamic
Deep Learning(DL)
Touch Pen information gathered during the time of
signature can also improve the
performance. The concepts of Neural
Networks as well as Wavelet transforms
hold a lot of promise in building systems
with high accuracy.
The system is robust and can detect
random, simple and semi-skilled
forgeries but the performance
deteriorates in case of skilled forgeries.
Using a higher dimensional feature
space and also incorporating dynamic
Pixel Image CNN , Numpy
information gathered during the time of
signature can also improve the
performance. The concepts of Neural
Networks as well as Wavelet transforms
hold a lot of promise in building systems
with high accuracy.
By applying a feature vector of test

signature if the output neuron generates
Pixel Image CNN value close to +1 test signature is
declared as genuine or if it generates
value close to -1 it is declared as forged.
By applying this technique we can verify

Pixel Image KNN
the original signature
22
We designed a Mutual Signature
DenseNet (MSDN), which extracts
features in different levels and combines
Support Vector
them, by learning from pairs of
Machines (SVM),
Pixel Image signatures. The network inputs either
Random Sample
whole signature image or local regions
Consensus (Ransac).
of signature. The similarity scores of
multiple local regions are fused to make
decision of verification.
Scale-Invariant Feature
Transform (Sift),
RGB Format Images The best accuracy got was 99.7%.
Convolutional Neural
Networks (CNN).
Maximum accuracy of 97% with

Hough Transform ,
personal verification based approach and
Digital Format Images Genetic Algorithm
minimum of 70% with genetic algorithm
,Contour Method etc.
is obtained
The result provided in this research for

Pixel Format Images SVM,CNN training and testing of the system many
signatures are used.
The experimental work in this research

is evaluated based on the number of
Images ,Paper, Pen CNN
correctly identified handwritten
signature images.
The result of the process is fully

clearance and can ensure the purpose
Images , Camera Deep Learning(DL)
which is detect or verifying the signature
with the most accuracy.
Our SIFT-based signature detection

algorithm outperforms existing methods
Touch Pen, Tablet Surf Algorithm, CNN
with a detection rate of 97% on the
Brazilian PUC-PR signature dataset.
23
(SVM),Random Using SVMs for signature recognition
Tablet, Touch Pen,
Sample Consensus achieves a recognition rate of 90% on a
Digital Images
(Ransac). dataset of 1000 signatures
Scale-Invariant Feature
Our CNN-based signature detection
Transform (Sift),
Paper, Pen, Camera algorithm achieves 95% accuracy on the
CEDAR signature dataset.
Networks (CNN).
Uyghur Signature
Dataset In The Self-
Built Database And
The CEDAR
Signature Dataset
This paper handled the problem of the
[12], TheProposed
lack of studies on offline signature
Method Achieved A
recognition of minority languages and
Recognition Rate Of Maximum Relevance
no publicly available dataset. This
98.4% Using A Minimum Redundancy
method achieved a recognition rate of
Diverse Signature Algorithm (MRMR).
96.8% for single-language signatures on
Database Compared
the self-built dataset and 96.41% for the
With Existing
mixed, multilingual signature dataset.
Methods. Effective
And Can Be Applied
To Large Datasets Of
Mixed, Multilingual
Signatures.
In this paper, biometrics involve face or
iris recognition, finger print matching,
The Authentication Of
signature recognition etc. the paper
A Person Can Be
discusses about the signature recognition
Done By Matching Spatial And Frequency
as it is one of the most used recognition
The Signature With Domain Approaches
for biometrics. The authentication of a
The Pre Stored
person can be done by matching the
Signature In Database.
signature with the pre stored signature in
database.
24
This paper proposes a method of off-line
signature recognition using Convolution
Neural Network. The purpose of this
paper is to obtain high accuracy multi-
A Multi-Lingual
class classification with a few training
Signature Database
signature samples. Images are
And CEDAR CNN,KNN
preprocessed to isolate the signature
Database Are Used
pixels from the background/noise pixels
For This Recognition.
using a series of Image processing
techniques. Initially, the system is
trained with 27 genuine signatures of 10
different authors each
Signature’s Database
Is Disseminated As A
RNA Preparation,
Reference Database In Output data with an average success of
Microarray
The Relevant Field 95.62%.
Hybridization
For Evaluation
Purposes.
Support Vector
The proposed system achieves an
The Real Signature Machine (SVM)
accuracy of 95.83% for forgery
Data Sets Algorithm And The
detection
K-Means Algorithm
A biometric cryptosystem using on-line

signature is presented. A function based
signature representation is employed to
generate the users’ templates, while
The Public MCYT
MLP,SVM protection and renewability are obtained
Signature Dataset
by binding the considered templates with
error correcting codes. A method for
selecting the most reliable signature
traits is also introduced.
With the rapid advancement in computer

science and information technology, the
demand for authentication of a person in
The UsingDataset Is A Fine-Tuned Transfer different organizations, institutions,
Used As A Learning-Based banks or online trading, etc. is increasing
Benchmark. Approach day by day. Similarly, signature
verification and recognition systems are
frequently used for forgery and fraud
detection.
25
The purpose of this study was to present
an automated recognition model based
on DL for efficiently classifying genuine
The Handwritten and forged signatures for offline
Signature Dataset, signature verification. The pre-trained
Deep Learning(DL)
TheGPDS Synthetic models VGG 16, VGG 19 ResNet 50,
Signature Dataset ResNet 101, MobileNet, and
EfficientNet achieved the accuracy of
80%, 81%, 77%, 73%, 71%, and 74%,
respectively.
Some Equation
The proposed system achieves an
Related Difference Markov Process Or
accuracy of 95.83% for forgery
Model Are Stated Marker Mode
detection
Here
The Authors Used Using SVMs for signature recognition

Self Made Input For SVM achieves a recognition rate of 90% on a
Conducting That Test. dataset of 1000 signatures
The popularity associated with

1 Literature Survey Surf Algorithm, CNN identification with neural networks
yields an accuracy of 94%.
Several Types Of Network (CNN),
The main steps of the verification and
Signature For Both Hidden Markov Model
identification systems were presented for
Online And Offline (HMM), And Support
both online and offline cases.
Database Vector Machine
(SVM)
26
Identified an 18-gene signature for
detecting severe dengue in patients with
secondary infection
upon hospital admission with a
sensitivity of 0.93 (95% confidence
interval [CI], .82–.98), specificity
Two Cohorts And RNA Preparation,
of 0.67 (95% CI, .53–.80), and area
Training And MicroarrayHybridizati
under the receiver operating
Validation Clinical on
characteristic curve (AUC) of 0.86 (95%
SubcohortsTable
CI, .75–.97). At validation, the signature
had empirical AUCs of 0.85 (95% CI,
.69–1.00) and 0.83 (95% CI, .68–.98) for
the PBMCs and whole blood datasets,
respectively.
The result of the process is fully

Several Types Of clearence and can ensure the purpose
SIFT ,RANSAC
Signatures Database which is detect or verifying the signature
with the most accuracy.

(MSDN), which extracts features in
different levels and combines them, by
Multiple Types Of learning from pairs of signatures. The
SVM, CNN.
Signature As Sample network inputs either whole signature
image or local regions of signature. The
similarity scores of multiple local
verification.
Dataset Which Is (MSDN), which extracts features in
Being Used In This different levels and combines them, by
Research Work learning from pairs of signatures. The
Convolution , CNN
Is A Collection Of network inputs either whole signature
130 Images From 130 image or local regions of signature. The
Individuals. similarity scores of multiple local
verification.
27
It was found to be the best signature
identification method because of its low
Convolutional Neutral
Pixel Image computational requirements and ability
Networks(CNN)
to extract the information necessary for
highly accurate device identification.
Fifteen harmonics having the largest

magnitudes normalized by their
Pixel Image Spectral Analysis corresponding variances were selected
and used in a stepwise discriminant
analysis.
Classification ratio of about 93% was

Pixel Image Recognition Technique
obtained under a threshold of 90%
Through the following interface, the user

can select signature image of interest
from available database. Then train the
A
Illustration Process network with the content of this database
Vortex,DopplerRadar
and 'matched ' or 'not matched' status
which indicates if the signature is exact
or forged.
The results showed that MLPs and SVM

were both able to determine the presence
Protein Sequence MLP,SVM of devices based on their harmonic
signetures with high accuracy around
95%.
The EBI public web interface serves

more than 10000 interactive requests a
month. There are more than 60
Pixel Image Siamese Algorithm installations worldwide of the Perl-based
InterProScan package that has alrady
been used to analyse the complete
genomes on a production scale.
28
Achieves a minimum detection cost
Naïve Bayes function (DCF) value equal to 1.37% for
MCVT
Algorithm random forgeries and 5.42% for skilled
forgeries.
Our results empirically show that SVM,

Vector Machine SVM,MLP which achieves up to 71% correct
recognition rate, outperforms MLP.
29
Chapter –3
Motivation
In this chapter, the detailed background about signature verification is discussed.
3.1 Introduction
A procedure to determine whether a signature is authentic or a fake is called handwritten

signature verification. A signature is simple to fake. Three categories of signature forgeries exist:
simple, random, and skilled forgeries. [15] [16].
Random Forgery: The forger creates it without knowing the author's name or authentic
signature.
Straightforward Forgery: This type of forgery involves no knowledge of the target signature's
appearance. As it rarely closely resembles a real signature, this kind of counterfeit is the easiest
to spot. Based on the faked signature's handwriting patterns, this form of forgery might
occasionally help an examiner determine who did it.
Expert forgery: When the forger possesses a sample of the target signature.
The level of skill and accuracy required to successfully simulate a signature relies on how much
experience the forger gets before trying the actual forgery. A well-done fake resembles the
original signature more. By moving from simple to professional forgery, the challenge of
signature verification gets progressively more challenging. The design of a signature verification
system becomes a significant problem because there is currently a growing need for the
processing of person identification to be quicker and more precise.
30
3.2 Signature Database
The MCYT project's action plan and development have been supported by the Biometric
Research Laboratory, ATVS, of the Universidad Politécnica de Madrid, where the design and
acquisition of a sizable bio-metric bi-modal database with fingerprint and signature qualities
have been completed [17].
Yet, there are some other business and forensic partners there. 25 client signatures and 25
expertly crafted forgeries (with natural dynamics) are obtained for everyone in the MCYT
Signature sub corpus. The database takes into account both online information (pen trajectory,
pen pressure, and pen azimuth=altitude) and offline information (picture of the written
signature).In the MCYT baseline online corpus, 330 (25 + 25) = 16500 signature samples are
taken into account. A graphics tablet is required since each signature is dynamically acquired;
the acquisition device utilized is a WACOM pen tablet, model. Given that the highest
frequencies of the underlying bio-mechanical movements are always below 2030 Hz, the
sampling frequency of the obtained signals is fixed to 100 Hz using the Nyquist sampling
criterion [17].
3.3 Preprocessing
Preprocessing signatures is a frequent practice to get rid of variants that are deemed unnecessary
for the completion of the verification. Typical preprocessing techniques include resampling, size
normalizing, and rotation normalization. The preprocessing stage involves improving the
signature in order to extract features. Before using any recognition technique, the signature
images need to be altered. This method cleans up and enhances the image's quality in order to get
rid of extraneous details, improve the selection of crucial features for recognition, and increase
the robustness of features that need to be retrieved. Moreover, preprocessing processes are
carried out to lessen input image noise and eliminate the majority of handwriting variability [15].
Some significant preprocessing methods for offline signatures are CNN, Seamens, and SVM.
Before feature extraction, there are further preprocessing stages including pen-up durations, drift
and mean removal, temporal normalization, and stroke concatenation.
31
Time dependencies must be removed from the representation in order to compare the spatial
properties of a signature. The beginning and ending positions of a stroke and the points where a
trajectory changes in a signature are two points that contain crucial information. These points are
extracted and held throughout the procedure because they are the most important ones [16].
3.4 Feature Extraction
Several signature specifications are used by signature verification systems. The accuracy of the
signature verification method is highly dependent on the features that are chosen for extraction.
Due to the many signature shapes and sampling circumstances, it is also the most challenging
stage of the signature verification system. Any signature verification system's feature extraction
procedure is a key challenge. Even yet, there is no assurance that two genuine signatures from
the same person will exactly match (intrapersonal variations).
Its difficulty also results from the fact that expert imitations adhere to the real design
(interpersonal variations). This contrasts with the way that irises and fingerprints from two
separate people might differ greatly from one another.In a perfect world, interpersonal variances
would outweigh intrapersonal variations by a wide margin. Consequently, it is crucial to
recognize and extract the characteristics that promote interpersonal variance while minimizing
intrapersonal variation. There is a lot of versatility in the features that may be used to verify a
signature, and the information that can be extracted from a signature is divided into two
categories:
1. A method based on parameter functions.
2. Function and feature-based methodology.
Using a parameter-basedapproach: Systems for verifying signatures vary in terms of the

features they choose to use and the ways in which they make decisions. There are two categories
of features: global and local.Global features are characteristics of a signature, such as its average
32
speed of signing, its bounding box, and its trajectory as described by Fourier analysis. Local
features relate to a particular sample location along the signature's journey. The distance and
curvature variation between subsequent places on the signature trajectory are two examples of
local characteristics. The most popular tools for capturing online signatures are pressure-
sensitive tablets, which can detect forces at the pen's tip as well as the pen's location.
Another typical local feature is the pressure data at each location along the signature trajectory.
To determine which features are more reliable for use in signature verification, some of these
features are compared.
Function-based, feature-based strategy: The signature is described using a function feature-

based method asposition, velocity, pressure, and other features are all examples of time functions
whose values make up the feature set.
33
Chapter-4
Methodology
In this chapter, we have discussed about the proposed methods of this report.
4.1 Introduction
The suggested approach for signature verification is the subject of this chapter. The
implementation process for an electronic handwritten signature is based on SNN, SVM, and
Siamese. Methodology is an essential component of signature detection and verification
reports. It outlines the process followed to develop the signature detection and verification
system, including data collection, preprocessing, feature extraction, and signature detection
and verification techniques used.
Start
Acquire Image
Binarization
Background
Noise Reduction
Segmentation
Extraction of Selected
Train the Classifier
Verification Methodology.
Perform Matching
End 34
4.2 Data Collection
Signature recognition and verification involve the process of identifying and authenticating an
individual's signature. Data collection is a critical component of signature recognition and
verification, and it is essential to collect accurate and comprehensive data for successful
recognition and verification.
Here are some of the key considerations for data collection for signature recognition and
verification:
4.2.1 Data Source:
The first step is to identify the source of the signature data. The data can be collected from
various sources, such as digital signatures, physical signatures on paper, or signatures captured
using a signature pad or stylus.
4.2.2 Signature Types:
Signatures can be categorized into two types: static and dynamic. Static signatures are those
that are captured as a digital image or scanned from paper, while dynamic signatures are those
that are captured in real-time using a signature pad or stylus. It is essential to collect both static
and dynamic signatures to create a robust signature recognition and verification system.
4.2.3 Number of Samples:
The number of signature samples required for data collection depends on the complexity of the
recognition and verification system. However, it is recommended to collect a minimum of 10
to 20 samples per individual to ensure accuracy and reliability.
35
4.2.4 Signature Variations:
Signature variations are a crucial aspect of data collection. It is essential to collect signatures
that represent the natural variations of an individual's signature. This includes signatures
created at different times, angles, and under different conditions.
4.2.5 Metadata;
Metadata is information that accompanies the signature data, such as the date, time, location,
and user ID. Collecting metadata can help improve the accuracy of the recognition and
verification system.
4.2.6 Data Quality:
It is essential to collect high-quality data for accurate signature recognition and verification.
This includes ensuring that the signature is captured in a clear and consistent manner, and there
are no artifacts or noise in the data.
4.2.7 Data Security:
Finally, data security is a critical consideration in data collection for signature recognition and
verification. It is essential to ensure that the collected data is encrypted and stored securely to
prevent unauthorized access and misuse.
In our project, we used the data those were hand-writing signature and signature images, a
digital camera with scanners and metadata for verification process.
Collecting data for signature recognition and verification requires careful consideration of
various factors, including the data source, signature types, and number of samples, signature
variations, metadata, data quality, and data security. By following these guidelines, you can
collect accurate and comprehensive data to create a robust signature recognition and
verification system.
36
4.3 Pre-Processing
Pre-processing of signature data is an essential step in signature recognition and verification. It

involves several techniques that aim to improve the quality of the signature data, remove any
noise or artifacts, and extract relevant features that can be used for recognition and verification.
Here are some of the key pre-processing techniques used in signature recognition and
verification:
4.3.1 Image Resizing
The first step in pre-processing is to resize the signature image to a standard size. This helps to
reduce the computation time and ensures that all images have the same resolution, making it
easier to compare and analyze them.
4.3.2 Image Filtering
Image filtering is a technique used to remove any noise or artifacts from the signature image.
This involves applying filters such as median filters or Gaussian filters to smooth the image
and remove any unwanted noise.
4.3.3 Binarization
Binarization is the process of converting a grayscale image into a binary image. This technique
is used to separate the foreground (signature) from the background. It involves thresholding the
image based on a specific threshold value.
4.3.4 Segmentation
Segmentation is the process of separating the signature from the surrounding background. This
is done using various techniques such as edge detection or contour detection. Once the
signature is separated, it can be analyzed and processed further.
37
4.3.5 Normalization
Normalization is the process of scaling and translating the signature image to a standard size
and position. This technique helps to reduce the variations in size, orientation, and position
between different signatures, making it easier to compare and analyze them.
4.3.6 Feature Extraction
Feature extraction involves extracting relevant features from the signature image that can be
used for recognition and verification. These features can include stroke thickness, direction,
curvature, and others. Feature extraction can be done using various techniques such as wavelet
transforms, principal component analysis (PCA), or histogram of oriented gradients (HOG).
4.3.7 Data Augmentation
Data augmentation is the process of creating additional data by applying various

transformations to the existing data. This technique can help to increase the size of the dataset,
improve the accuracy of the model, and reduce over fitting. Some common data augmentation
techniques for signature recognition and verification include rotation, scaling, and adding
noise.
Pre-processing of signature data is a crucial step in signature recognition and verification. It

involves several techniques that aim to improve the quality of the data, remove any noise or
artifacts, and extract relevant features for recognition and verification. By following these pre-
processing techniques, you can create a robust signature recognition and verification system
that can accurately identify and authenticate signatures.
4.4 Feature Extraction
Feature extraction is a critical step in signature recognition and verification. It involves

extracting relevant features from the signature data that can be used for comparison,
classification, and verification. Here are some of the key feature extraction techniques used in
signature recognition and verification:
38
4.4.1 Stroke-based Features
Stroke-based features involve extracting information about the stroke patterns used to create
the signature. These features can include stroke direction, length, curvature, and velocity. They
can be extracted using techniques such as dynamic time warping (DTW) or the directional
distance function (DDF).
4.4.2 Shape-based Features
Shape-based features involve extracting information about the shape of the signature, such as
the width, height, and aspect ratio. These features can be extracted using techniques such as
principal component analysis (PCA) or Fourier descriptors.
4.4.3 Texture-based Features
Texture-based features involve extracting information about the texture of the signature, such
as the smoothness, roughness, or graininess. These features can be extracted using techniques
such as local binary patterns (LBP) or gray-level co-occurrence matrix (GLCM).
4.4.4 Global Features
Global features involve extracting information about the overall shape and structure of the
signature. These features can include the perimeter, area, and compactness of the signature.
They can be extracted using techniques such as morphological operations or moments.
4.4.5 Local Features
Local features involve extracting information about specific regions of the signature, such as
the endpoints, corners, or loops. These features can be extracted using techniques such as
Harris corner detection or wavelet transform.
Hybrid Features: Hybrid features involve combining multiple types of features to create a more
robust and accurate signature recognition and verification system. For example, combining
stroke-based and shape-based features can provide a more comprehensive representation of the
signature.
39
4.4.6 Deep Learning Features
Deep learning features involve using convolutional neural networks (CNNs) or other deep
learning models to extract features automatically from the signature data. These features can be
more complex and abstract than traditional feature extraction techniques and can lead to higher
accuracy in signature recognition and verification.
Feature extraction is a critical step in signature recognition and verification. It involves

extracting relevant information from the signature data, such as stroke patterns, shape, texture,
global, and local features. By using these feature extraction techniques, you can create a robust
and accurate signature recognition and verification system that can accurately identify and
authenticate signatures.
4.5 Signature Detection
Signature detection is the process of locating the signature region within an image or
document. In signature recognition and verification, signature detection is a crucial step, as it
enables the system to isolate the signature from the background and process it for further
analysis. Here are some of the key techniques used in signature detection:
4.5.1 Edge Detection
Edge detection is a common technique used in signature detection. It involves detecting the
edges of the signature using techniques such as the Sobel operator, Canny edge detection, or
Laplacian of Gaussian (LoG) filter. Once the edges are detected, the signature region can be
identified based on its boundary.
40
4.5.2 Template Matching
Template matching is a technique that involves comparing the signature image with a pre-
defined template. The template can be created manually or automatically using a reference
signature. Once the template is generated, it can be used to detect the signature region within
the image.
4.5.3 Hough Transform
The Hough transform is a technique used to detect lines and curves within an image. In
signature detection, the Hough transform can be used to identify the curve of the signature and
locate its boundary.
4.5.4 Connected Component Analysis
Connected component analysis is a technique used to identify and group pixels that belong to
the same object or region within an image. In signature detection, connected component
analysis can be used to identify the connected pixels that make up the signature and separate
them from the background.
4.5.5 Machine Learning-based Techniques
Machine learning-based techniques, such as object detection or semantic segmentation, can

also be used for signature detection. These techniques involve training a machine learning
model on a dataset of signature images and using it to detect the signature region within a new
image.
Signature detection is a critical step in signature recognition and verification. It involves

identifying the signature region within an image or document, enabling the system to isolate
the signature from the background and process it for further analysis. By using techniques such
as edge detection, template matching, Hough transform, connected component analysis, or
machine learning-based techniques, you can create a robust signature detection system that can
accurately locate signatures within images and documents.
41
4.6 Signature Verification
Signature verification is the process of comparing a signature with a reference signature to

determine if they belong to the same person. In signature recognition and verification,
signature verification is a critical step as it enables the system to authenticate the signature and
determine its validity. Here are some of the key techniques used in signature verification:
4.6.1 Feature Matching
Feature matching involves comparing the extracted features of the signature being verified
with the reference signature's features. This comparison can be done using techniques such as
distance-based measures like Euclidean distance, Dynamic Time Warping (DTW), or
Directional Distance Function (DDF). If the features of the two signatures are similar, it can be
concluded that the signatures belong to the same person.
4.6.2 Neural Networks
Neural networks, such as Convolutional Neural Networks (CNNs) or Recurrent Neural

Networks (RNNs), can be trained on a dataset of reference signatures and used for signature
verification. The neural network model can learn to extract relevant features from the signature
data and classify them as belonging to the same or different person.
4.6.3 Support Vector Machines (SVM)
SVM is a machine learning algorithm that can be used for signature verification. SVM can be
trained on a dataset of reference signatures and used to classify a new signature as belonging to
the same or different person based on the extracted features.
4.6.4 Dynamic Time Warping (DTW)
DTW is a technique used to compare time series data, such as signature strokes. DTW can be
used to compare the signature strokes of the reference signature and the signature being
verified and determine if they belong to the same person.
42
4.6.5 Graphology
Graphology is a technique that involves analyzing the characteristics of a signature, such as its
size, shape, and pressure, to determine its authenticity. Graphology involves comparing the
reference signature's characteristics with the signature being verified and determining if they
match.
4.6.6 Hybrid Techniques
Hybrid techniques involve combining multiple techniques, such as feature matching and neural
networks, to create a more robust and accurate signature verification system.
Signature verification is a critical step in signature recognition and verification. It involves

comparing a signature with a reference signature to determine if they belong to the same
person. By using techniques such as feature matching, neural networks, SVM, DTW,
graphology, or hybrid techniques, you can create a robust signature verification system that can
accurately authenticate signatures and determine their validity.
4.7 Types of Forgery
Any signature verification system's primary task is to determine if a signature is real or a fake.
Three categories of forgery in signature verification systems are recognized. Random, basic,
and accomplished [8]. Random Forgeries occur when signs are made without any information
regarding the style or signer's name. Straightforward forgeries are signatures in which the
forger has knowledge of the signatory's name but no preceding sample, whereas skilled
forgeries are signatures in which the forger is aware of the name of the signatory and the
original signature's style. Expert forgeries are the toughest to spot.
43
4.8 AlgorithmDetails
In Signature Detection and Verification, we used SVM, CNN and Siamese Networks
algorithms. Those algorithms are discussed in details are below:
4.8.1 Support Vector Machines (SVM)
Support Vector Machines (SVMs) are a popular machine learning algorithm used for
classification and regression analysis. In the context of signature detection and verification,
SVMs are often used for signature verification based on extracted features.
The basic idea behind SVMs is to find a hyper plane in a high-dimensional feature space that
maximally separates the data points belonging to different classes. The hyper plane is
determined by the support vectors, which are the data points closest to the decision boundary.
To train an SVM model, the data points are first transformed into a high-dimensional feature
space using a kernel function. The choice of kernel function can have a significant impact on
the performance of the SVM model, and common kernel functions used in signature
verification tasks include linear, polynomial, and radial basis function (RBF) kernels.
Once the data points have been transformed into the high-dimensional feature space, the SVM
algorithm tries to find a hyper plane that separates the data points belonging to different classes
with the maximum margin. The margin is defined as the distance between the hyper plane and
the closest data points from each class.
In some cases, the data points may not be linearly separable, meaning that there is no hyper
plane that can perfectly separate the data points belonging to different classes. In such cases,
the SVM algorithm can use a soft margin, which allows some data points to be misclassified in
order to find a hyper plane with the maximum margin.
44
After training the SVM model, it can be used to predict the class label of new data points based
on their feature representations. In the context of signature verification, the SVM model can be
used to distinguish between genuine and forged signatures based on extracted features such as
stroke width, curvature, and pressure.
SVMs have several advantages for signature verification tasks. They can handle high-
dimensional feature spaces and are known for their ability to handle small and noisy datasets.
SVMs also have good generalization performance and can be applied to different types of
signature verification tasks, including online and offline signature verification.
However, SVMs also have some limitations. They can be computationally expensive to train
on large datasets, and the choice of kernel function can have a significant impact on the
performance of the model. SVMs also require careful tuning of hyperparameters such as the
penalty parameter and kernel coefficient, which can be a challenging task in practice.
Random Forest: Random Forest is an ensemble learning algorithm that can be used for
signature detection and verification. It works by constructing a multitude of decision trees
during training and then combining their outputs to make a final decision. Random Forest is
known for its ability to handle complex data and is often used for signature verification based
on extracted features.
4.8.2 Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNNs) are a type of deep learning algorithm used primarily
for image classification, object recognition, and other computer vision tasks. In the context of
signature detection and verification, CNNs are often used to classify and detect signatures
based on their visual features.
CNNs consist of multiple layers of interconnected nodes, each of which performs a specific
computation on the input data. The first layer in a CNN is typically a convolutional layer,
which applies a set of learnable filters to the input image to extract features such as edges,
corners, and textures. The output of the convolutional layer is then passed through a non-linear
activation function such as the Rectified Linear Unit (ReLU) to introduce non-linearity into the
model.
After the convolutional layers, CNNs typically include one or more pooling layers, which
reduce the spatial dimensionality of the feature maps and extract the most important features.
45
The most common pooling operation is max pooling, which selects the maximum value in each
sub-region of the feature map.
The output of the pooling layer is then passed through one or more fully connected layers,
which perform a linear transformation on the features and produce the final output of the
model. The final output of the model is typically a probability distribution over the different
classes in the classification task.
CNNs are typically trained using a variant of the backpropagation algorithm called stochastic
gradient descent (SGD). During training, the weights of the filters in the convolutional layers
and the weights of the fully connected layers are updated using the gradients of the loss
function with respect to the model parameters.
One of the key advantages of CNNs is their ability to learn complex feature representations
directly from the raw input data. Unlike traditional machine learning algorithms that require
hand-crafted features, CNNs can automatically learn features that are optimized for the specific
task at hand.
Another advantage of CNNs is their ability to handle translation invariance and local spatial
relationships. By using shared weights in the convolutional layers, CNNs can detect the same
features in different parts of the input image, which makes them well-suited for tasks such as
object recognition and facial recognition.
However, CNNs also have some limitations. They require large amounts of training data and
can be computationally expensive to train on large datasets. The interpretability of CNNs is
also a challenge, as it can be difficult to understand how the model is making its predictions.
Despite these challenges, CNNs have proven to be highly effective for a wide range of
computer vision tasks and are widely used in industry and academia.
46
4.8.3 Siamese Networks
Siamese Networks are a type of neural network architecture used for tasks such as image
similarity detection, facial recognition, and signature verification. Siamese Networks consist of
two identical neural networks, each of which takes an input image and produces a feature
representation. The two feature representations are then compared using a similarity metric to
determine if the two images are similar or dissimilar.
The Siamese architecture was originally introduced for signature verification by Bromley et al.
in 1994. The Siamese architecture is particularly well-suited for tasks such as signature
verification, where the inputs are variable length sequences, and the similarity metric must take
into account both global and local features of the input.
Siamese Networks are typically trained using a variant of the contrastive loss function, which
encourages the network to produce similar feature representations for similar inputs and
dissimilar feature representations for dissimilar inputs. During training, pairs of input images
are randomly sampled from the training dataset and passed through the Siamese Network. The
two feature representations are then compared using the contrastive loss function, and the
network weights are updated to minimize the loss.
Siamese Networks have several advantages over traditional neural network architectures. They
can learn feature representations that are optimized for the similarity task, which makes them
well-suited for tasks such as facial recognition and signature verification. They can also handle
variable-length inputs and learn to recognize complex patterns and features in the input.
One of the main challenges of Siamese Networks is that they can be computationally expensive
to train, particularly for large datasets. They also require careful tuning of the similarity metric
and loss function to achieve good performance.
Despite these challenges, Siamese Networks have proven to be highly effective for a wide
range of similarity tasks, including facial recognition, signature verification, and even natural
47
language processing. They are widely used in industry and academia and continue to be an
active area of research in the field of deep learning.
FLOW CHART
Start
Acquire Image
Binarization
Background
Noise Reduction
Segmentation
Extraction of Selected
Train the Classifier
Perform Matching
End
Fig-4.2: Flow-Chart of Signature Detection and Verification
48
Chapter-5
Performance Measures
A signature recognition system's effectiveness depends on how precisely it can distinguish

between authentic and fake signatures [9]. Various performance evaluation metrics such as
SVM(Support Vector machine), CNN(Convolutional Neural Network), Siamese networks are

evaluated for different signature recognition system on the basis of their performance is
evaluated and compared.
Here, Support Vector Machines (SVMs) are a popular machine learning algorithm used for
classification and regression analysis. Convolutional Neural Networks (CNNs) are used to
deep learning algorithm used primarily for image classification, object recognition, and other
computer vision tasks. Siamese Networks are used for tasks such as image similarity detection,
facial recognition, and signature verification. Siamese Networks consist of two identical neural
networks, each of which takes an input image and produces a feature representation. By
applying these algorithms, we have gained accuracy of 99.69%.
49
Chapter-6
Result and Discussion
6.1 Result
5 real signatures and 20 fake signatures are used in the experiment, with 5 real signatures
serving as the reference signatures.
Here are some potential results and findings of a signature verification and detection process
using Siamese Networks.
Accuracy: Siamese networks have been shown to achieve high accuracy rates in signature
verification tasks. The specific accuracy achieved will depend on the quality and size of the
dataset used for training and validation. Accuracy: The percentage of accurately classified data
samples over all the data is known as accuracy. Accuracy can be calculated by the following
equation. [43]
Accuracy = (TP+TN)/(TP+FP+TN+FN)
Speed: Siamese networks can be relatively fast for signature verification, as they only need to
compare pairs of signatures rather than analyzing entire signature images. However, the speed
may be impacted by the size and complexity of the Siamese network architecture.
Sensitivity to quality: The performance of a Siamese network for signature verification may
be impacted by the quality of the input signatures. Poorly scanned or low-quality signatures
may be more difficult for the system to compare accurately.
Potential for bias: Siamese networks may be susceptible to bias if the training data is not
diverse enough. For example, if the training data is heavily skewed towards signatures from
one demographic, the system may not perform as well on signatures from other demographics.
50
Interpretability: Siamese networks can be more difficult to interpret than other machine
learning models, as they are based on a complex neural network architecture. This may make
it harder to identify the specific features of a signature that are most important for
verification.
Overall, the effectiveness of a signature verification and detection process using siamese
networks will depend on many factors, including the quality and diversity of the training data,
the specific Siamese network architecture used, and the specific use case for the system.
Here is theoutlook of our project’s output
Fig 6.1:Verified Signature Image
We have gained the verified signature by applying the above algorithms.In this output, our
project shows a message like “Signature is verified” if the signature image of user is matched
with the image that was stored on the database. Besides the verified message we have also
gained the identification information of the user like User’s name, Age and email address.
51
Fig 6.2: Unverified Signature Image
When the user’s signature doesn’t match with the signature stored on the database then the
project shows a message that is “Signature is not verified”.
Here is an illustration of the user’s interface of signature verification and detection project.
Fig 6.3: User interface
52
6.2 Discussion
Throughout the last thirty years, numerous approaches for verifying signatures have been
proposed. The two methods utilized to extract the pertinent data/features from the signatures are
global features based and local features. These characteristics are covered. The system must be
tested using the available 5 authentic and 20 fake signatures during implementation to obtain the
performance curves. Testing with each test signature and storing the score values took hours.
This grew to be a serious issue. Hence, a Deep Learning is built with all of the test signatures to
tackle this problem. Once created, it was simple to compute the performance curves by just
clicking the run button.
Table 6.1: Comparison of various offline signature verification techniques.
SI. No Approach FRR FAR Accuracy

1 SVM 4.8% 5.3% 97.54%
2 CNN 12% 23% 98.12%
3 Siamense Network 11.8% 13.1% 98.96%
4 Hough Transform 14% 2.8% 94.25%
5 KNN 1.5% 2.5% 97.9%
53
Chapter-7
Conclusion and Future Work
7.1 Conclusion
The TS model, which incorporates structural characteristics in its exponential membership

function, models an off-line signature verification and forgery detection system. Angle-based
features are retrieved using the box technique. Because to the variances in handwritten
signatures, each characteristic produces a fuzzy set when its values are collected from all
samples. Each rule in this formulation is made up of just one characteristic. A sizable database of
signatures has been used to assess the effectiveness of this approach. The verification method
can precisely identify all varieties of forgeries, including random, skilled, and unskilled ones.
The initial parameter selection is vital but not absolutely necessary. But, we only need to make
the right decision once, and it applies to all forms of signatures. Due to the lack of simplicity at
the implementation stage, we have not utilized global learning methodologies.
7.2 Future work
We provide a state-of-the-art for the most recent techniques utilized in offline signature
verification systems in this study. Although various methods are employed in this field, accuracy
has to be improved, particularly for sophisticated forgeries. In this project,accuracy of several
on-hand techniques are described and compared. The accuracy achieved so far by the current
technologies is not particularly high, necessitating more study into off-line signature verification.
Future work might possibly combine various classifiers to produce better verification outcomes.
54
Reference
[1] J. K. Guo. Forgery detection by local correspondence. PhD thesis, College Park, MD,
USA, 2000. Director-Rosenfeld, Azriel.
[2] J. A. Unar, W. C. Seng, A. Abbasi, A review of biometric technology along with trends
and prospects. Pattern Recognit. (2014).
[3] S. Ghandali and M. EbrahimiMoghaddam, "Off-Line Persian Signature Identification
and Verification based on Image Registration and Fusion", Journal of Multimedia, vol.
4, pp. 137-144, 2009.
[4] Griechisch E, Malik MI, Liwicki M: Online Signature Verification using
Accelerometer and Gyroscope Proc. 16th Bienn. Conf. Int. Graphonomics Soc. 2013;
no. January: pp. 143–146. Heba Mohsen, El-Sayed A. El-Dahshan,
[5] El-SayedM.El-Horbaty, Abdel-Badeeh M. Salem,Classification using deep learning
neural networks for brain tumors,Future Computing and Informatics Journal,Volume 3,
Issue 1,2018,
[6] Naz, Saeeda, KiranBibi, and Riaz Ahmad. "DeepSignature: fine-tuned transfer
learning-based signature verification system." Multimedia Tools and Applications
81.26 (2022): 38113-38122.
[7] L. G. Hafemann, R. Sabourin and L. S. Oliveira, “Offline handwritten signature
verification—literature review,” in Proc. 7th Int. Conf. Image Process. Theory, Tools
Appl. IPTA 2017, 1 –8 (2018).
[8] Attwood, T.K. and Parry-Smith, D.J. Introduction to Bioinformatics 1999 Addison
Wesley Longman
[9] D. Impedovo, G. Pirlo, and R. Plamondon, “Handwritten signature verification: New
advancements and open issues,” in 2012 International Conference on Frontiers in
Handwriting Recognition, Sept 2012, pp. 367–372.
[10] H. Lei and V. Govindaraju, “A comparative study on the consistency of features in on-
line signature verification,” Pattern Recogn. Lett., vol. 26, no. 15, pp. 2483–2489, Nov.
2005.
55
[11] F. J. Zareen and S. Jabin, “A comparative study of the recent trends in biometric
signature verification,” in 2013 Sixth International Conference on Contemporary
Computing (IC3), Aug 2013, pp. 354–358
[12] G. Padmajadevi and K. S. Aprameya, “A review of handwritten signature verification
systems and methodologies,” in 2016 International Conference on Electrical,
Electronics, and Optimization Techniques (ICEEOT), March 2016, pp. 3896–3901.
[13] D. Morocho, J. Hernandez-Ortega, A. Morales, J. Fierrez, and J. OrtegaGarcia, “On the
evaluation of human ratings for signature recognition,” in 2016 IEEE International
Carnahan Conference on Security Technology (ICCST), Oct 2016, pp. 1–5.
[14] R. Plamondon and S. N. Srihari, “On-line and off-line handwriting recognition: A
comprehensive survey,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp. 63–
84, Jan. 2000.
[15] M. M. Fahmy, “Online handwritten signature verification system based on dwt features
extraction and neural network classification,” Ain Shams Engineering Journal, vol. 1,
no. 1, pp. 59 – 70, 2010.
[16] C. Gruber, T. Gruber, S. Krinninger, and B. Sick, “Online signature verification with
support vector machines based on lcss kernel functions,” IEEE Transactions on
Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 40, no. 4, pp. 1088–1100,
Aug 2010.
[17] C. Gruber, T. Gruber, and B. Sick, Online Signature Verification with New Time
Series Kernels for Support Vector Machines. Berlin, Heidelberg: Springer Berlin
Heidelberg, 2005, pp. 500–508.
[18] A. Sharma and S. Sundaram, “A novel online signature verification system based on
gmm features in a dtw framework,” IEEE Transactions on Information Forensics and
Security, vol. 12, no. 3, pp. 705–718, March 2017.
[19] M. Faundez-Zanuy, “On-line signature recognition based on vq-dtw,” Pattern Recogn.,
vol. 40, no. 3, pp. 981–992, Mar. 2007. [12] B. Kar, P. K. Dutta, T. K. Basu, C.
VielHauer, and J. Dittmann, “Dtw based verification scheme of biometric signatures,”
in 2006 IEEE International Conference on Industrial Technology, Dec 2006, pp. 381–
386.
56
[20] G. Zapata, J. D. Arias-Londoño, J. Vargas-Bonilla, and J. R. Orozco, “Online signature
verification using gaussian mixture models and small-sample learning strategies,”
RevistaFacultad de Ingeniería, vol. 2016, 06 2016.
[21] B. Drott and T. Hassan-Reza, “On-line handwritten signature verification using
machine learning techniques with a deep learning approach,” 2015, student Paper. S. Z.
Li, Encyclopedia of Biometrics, 1st ed. Springer Publishing Company, Incorporated,
2009.
[22] A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric recognition,”
IEEE Trans. Cir. and Sys. for Video Technol., vol. 14, no. 1, pp. 4–20, Jan. 2004.
[23] J. Ortega-Garcia, J. Fierrez-Aguilar, D. Simon, J. Gonzalez, M. FaundezZanuy, V.
Espinosa, A. Satue, I. Hernaez, J. J. Igarza, C. Vivaracho, D. Escudero, and Q. I. Moro,
“Mcyt baseline corpus: a bimodal biometric database,” IEE Proceedings - Vision,
Image and Signal Processing, vol. 150, no. 6, pp. 395–401, Dec 2003.
[24] D. A. Reynolds and R. C. Rose, “Robust text-independent speaker identification using
gaussian mixture speaker models,” IEEE Transactions on Speech and Audio
Processing, vol. 3, no. 1, pp. 72–83, Jan 1995.
[25] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, “Speaker verification using adapted
gaussian mixture models,” Digit. Signal Process., vol. 10, no. 1, pp. 19–41, Jan. 2000.
[26] "On-line signature verification," Pattern Recognition, vol. 35, no. 12, pp. 2963-2972,
2002, pattern Recognition in Information Systems. A. K. Jain, F. D. Griess, and S. D.
Connell.
[27] "Classical and subsequence dynamic time warping for recognition of rigid body motion
trajectories," in O. Cigdem, T. D. Laet, and J. D. Schutter, June 2013, pp. 1-6, 9th
Asian Control Conference (ASCC).
[28] Automatic online signature verification based on dynamic function characteristics was
described by Z. Chen, X. Xia, and F. Luan in the 7th IEEE International Conference on
Software Engineering and Service Science (ICSESS), which was held in August 2016.
[29] In "Automatic signature verification: The state of the art," D. Impedovo and G. Pirlo
published their research in IEEE Transactions on Systems, Man, and Cybernetics, Part
C (Applications and Reviews), vol. "Online signature verification based on signatures
turning angle representation using longest common divisor," K. Barkoula, G.
57
Economou, and S. Fotopoulos International Journal on Document Analysis and
Recognition (IJDAR), vol. 16, no. 3, pp. 261-272, September 2013.
[30] Online Handwritten Signature Verification by S. Garcia-Salicetti, N. Houmani, B. Ly-
Van, B. Dorizzi, F. Alonso-Fernandez, J. Fierrez, J. Ortega-Garcia, C. Vielhauer, and
T. Scheidat. 125–165, London: Springer London, 2009.
[31] Fusion of Local and Regional Approaches for On-Line Signature Verification, by J.
Fierrez-Aguilar, S. Krawczyk, J. Ortega-Garcia, and A. K. Jain, 34. 2005, Springer
Berlin Heidelberg, Berlin, pp. 188–196.
[32] D. S. Guru and H. N. Prakash, "Online signature verification and recognition: An
approach based on symbolic representation," IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 31, no. 6, june 2009, pp. 1059–1073.
[33] "Hmmbased on-line signature verification: Feature extraction and signature modeling,"
vol. 28, no. 12, pp. 2325–2334, 12 2007. J. Fierrez, J. Ortega-Garcia, D. Ramos, and J.
Gonzalez-Rodriguez.
[34] D. Cai, C. Zhang, and X. He, "Unsupervised feature selection for multicluster data,"
Proceedings of the 16th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, series KDD '10, New York, NY, USA: ACM, 2010, pp.
333–342.
[35] Online signature verification using probabilistic modeling and neural networks, in 2012
Spring Congress on Engineering and Technology, May 2012, pp. 1–5, by M.
Alhaddad, D. Mohamad, and A. M. Ahsan.
[36] Online signature verification using graph representation is discussed in K. Wang, Y.
Wang, and Z. Zhang's paper from the Sixth International Conference on Image and
Graphics, which was published in August 2011 and covered pp. 943–948.
[37] D. Z. Lejtman and S. E. George, "On-line handwritten signature verification using
back-propagation neural networks," Proceedings of Sixth International Conference on
Document Analysis and Recognition, 2001, pp. 992-996.
[38] "A signature complexity measure to select reference signatures for online signature
verification," 2015 International Conference of the Biometrics Special Interest Group
(BIOSIG), Sept 2015, pp. 1-8. C. Kahindo, S. Garcia-Salicetti, and N. Houmani..
58
[39] "An efficient lowcost approach for on-line signature recognition based on length
normalization and fractional distances," Pattern detection, vol. 42, no. 1, pp. 183–193,
2009. C. Vivaracho-Pascual, M. Faundez-Zanuy, and J. M. Pascual.
[40] Identity authentication using an enhanced online signature verification approach was
described by A. Kholmatov and B. Yanikoglu in their 2005 paper published in Pattern
Recognition Letters, vol. 26, no. 15, pp. 2400–2408.
[41] A writer identification method for online whiteboard data was published in Pattern
Recognition, vol. 41, no. 7, 2008, pp. 2381–2397. A. Schlapbach, M. Liwicki, and H.
Bunke.
[42] Biometric Authentication Using Online Signatures by A. Kholmatov and B. Yanikoglu.
Pages 373–380 in Berlin and Heidelberg: Springer Berlin Heidelberg, 2004.
[43] Samia, Moythry Manir, et al. "Aspect-based Sentiment Analysis for Bengali Text using
Bidirectional Encoder Representations from Transformers (BERT)." International
Journal of Advanced Computer Science and Applications 13.12 (2022).
59
View publication stats

Signature Verification and Detection

Uploaded by

Copyright:

Available Formats

You might also like

Signature Verification and Detection

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Signature Verification and Detection

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Project Report on Signature Veriﬁcation and Detection by Image Processing

Technical Report · May 2023

Alimul Rajee Sumaiya Zafrin

SEE PROFILE SEE PROFILE

Ahad Hossain Shuvo Tasrina Parvin

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

Signature Verification and Detection by Image Processing

Lecturer Tasrina Parvin 11909014

Dept. of Information and Communication Ahad Hossain 11909022

Submission Date: May 2023

Chapter Title Page

1.2 Aims and Objectives 10

1.3 Outline of This Report 10-11

2 Literature Review 12-19

3.2 Signature Database 21-22

3.4 Feature Extraction 23-24

4.2 Data Collection 26-27

4.2.1 Data Source 26

4.2.2 Signature Types 26

4.2.3 Number of Samples 26

4.2.4 Signature Variations 27

4.2.7 Data Security 27

4.3 Pre-Processing 28-29

4.3.1 Image Resizing 28

4.3.2 Image Filtering 28

4.3.6 Feature Extraction 29

4.3.7 Data Augmentation 29

4.4 Feature Extraction 29-31

4.4.1 Stroke-based Feature 30

4.4.2 Shape-based Feature 30

4.4.3 Texture-based Features 30

4.4.6 Deep Learning Features 31

4.5 Signature Detection 31-32

4.5.1 Edge Detection 31

4.5.2 Template Matching 31

4.5.3 Hough Transform 32

4.5.5 Machine Learning Based Technique 32

4.6 Signature Verification 33-34

4.6.1 Feature Matching 33

4.6.2 Neural Networks 33

4.6.3 Support Vector Machine 33

4.6.4 Dynamic Time Warping 33

4.6.6 Hybrid Techniques 34

4.7 Types of Forgery 34

4.8 Algorithm Details 35-39

4.8.1 Support Vector Machines 35-36

4.8.2 Convolutional Neural Networks 36-37

4.8.3 Siamese Networks 38-40

6 Result and Discussion 42-46

6.1 Results 42-45

7 Conclusion and Future Work 47

1.1 Module of Signature Verification System 12

4.1 The Block Diagram of Signature Detection and 27

5.1 Genuine Signature 54

6.1 Verified Image 44

No. Title Page

2.1 Different Algorithm and Dataset with Result and 21