2022 IEEE International Conference on Data Science and Information System (ICDSIS)

Handwritten Signature Verification System using

Deep Learning
1st Sudharshan Duth P 2nd Vismaya R N
Department of Computer Science Department of Computer Science
2022 IEEE International Conference on Data Science and Information System (ICDSIS) | 978-1-6654-9801-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICDSIS55133.2022.9915833

Amrita School of Arts and Sciences Amrita School of Arts and Sciences
Mysuru, India Mysuru, India
[email protected] [email protected]

Abstract—Every person has a unique signature, which is handwritten signature. An offline or online signature can be
primarily used for identity documents and the authentication of received by a signature processing system. Processing
significant documents or legal documentation. Static and handwritten signatures offline is harder than processing then
dynamic signature verification are the two types of signature online cases due to a lack of information and is considered a
verification. The basic parameters that must be addressed when difficult problem.
creating a signature verification system are user acceptance, the
level of security necessary, accuracy, cost, and implementation. Consider the following scenario. The cashier verifies the
The objective of Signature verification used to determine signature against saved records of real signatures before
whether a signature is real or a fake. This has proven to be a proceeding with the legal transaction.
difficult endeavor, particularly in the offline scenario, which
employs images of scanned signatures and does not have access Offline signature verification can approach in two ways: a
to dynamic information about the signing process. The offline writer dependent method, in which system is updated
handwritten signature is one of the most essential biometrics whenever new signer is introduced, and a writer independent
used in banking systems, administrative, and financial method, in which a generalized system is constructed to
applications. The study is to evaluate the performance of a few distinguish between authentic and counterfeit signatures.
well-known deep convolutional neural networks as feature
A signature is a sign or mark that a writer writes to
extractors in Handwritten Signature Verification (HSV) using
transfer learning with activation function, as well as to review
establish his or her identity and true intent. It is often stylized
the available convolutional neural network signature and done by hand. A person's handwritten signatures is widely
verification methods. This is achieved by three pretrained accepted as a way of verifying the legal implications of
models, VGG16, VGG19, and ResNet50, which are most often records such as certificates, checks, draught documents,
used generic models in computer vision tasks. Obtained letters, authorizations, visas, passports, and other financial
experimental results, by comparing three models with documents, and is required to avoid falsifying and gerying
SigComp2009 datasets and adding different parameters to each such documents in a variety of financial, legal, administrative,
model. And conclude that VGG19 is best suitable for dataset, academic, and other financial settings.
which has 97.83 percent accuracy than another model. These
findings will help academics construct more effective Deep
Signature verification is vital in numerous transactions and
Learning-based signature verification methods. approval processes, particularly in the banking industry with
adjacent nations. The signature of the applicant of checks will
Keywords—Offline Handwritten Signature, Signature be verified by the Bank using a specimen signature maintained
Verification, Deep Convolutional Neural Network, Transfer in bank custody. The current banking ICDM (Image
Learning. Capturing and Display Module) is a wonderful tool for
recording the applicant's specific signatures and presenting
I. INTRODUCTION them if necessary. However, as other complicated falsification
The practice of automatically and instantly examining strategies arise, such a simple approach may not be sufficient
signs to identify whether they are genuine or not is known as in all circumstances. Using machine learning approaches, this
signature verification and forgery detection. Static and initiative aims to assist and improve the human signature
dynamic signature verification are the two basic signature verification process.
verification kinds Static verification, often known as off-line
II. BACKGROUND STUDY
verification, is a type of verification that takes place offline.,
whereas on a digital tablet or comparable device, a person Kazim Fouladi, A new metrics learning method based on
creates his or her signatures, which is verified dynamically or deep multitasking learning is proposed. SVM and
on-line. The database is then created by comparing the Discriminative Deep Metric Learning are outperformed by the
signature in questions to previous sample of person signature. proposed method [1]. K. C. Santosh, offer a method to
authenticate handwritten signatures in this paper that is based
Handwritten signature verification can be done in a variety on a deep convolution neural network (CNN), a bio-inspired
of ways, and most businesses employ one of them currently. network that acts as if it were a human brain. Deep CNN
Offline signatures make use of the system's static features, retrieves features from the images being analyzed, and then a
which include image processing algorithms for analyzing cubic support vector machine is used to classify them. The
signature correctness. These include a person's initial authors tried their suggested work on three different datasets:
identification via a password. Other multi - modal methods GPDS, BME2, and SVC20, and received positive findings on
strictly verify an individual's identification using the two all three. The findings highlight the importance of deep
biometric features. Biometrics are used by security systems all learning in offline signature verification. The proposed
around the world every day to be able to recognize or test the method boosted the rate of verification and lowered the
validity of people. One of the most essential biometrics in number of incorrect acceptances. Increased verification
tanking, administrative, and financial purposes is the

978-1-6654-9801-2/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: Lakehead University. Downloaded on February 28,2024 at 06:28:44 UTC from IEEE Xplore. Restrictions apply.
results can be achieved with a more particular design [2]. the presence of overfitting when doing feature selection in a
Pinzón-Arenas, describes the implementation of a DAG-CNN wrapper mode utilizing Binary Particle Swarm. To prevent
that uses the writer-independent approach to classify and overfitting during the search for the most discriminant
verify the authenticity of three users' offline signatures. Two representation, suggest an approach based on a global
databases (training/validation and testing) were manually verification strategy with an external archive. In addition, an
created in order to complete this project. After the network has evaluation of the usage of the selected features in a transfer
been trained, it is validated and tested, with overall accuracies learning context is conducted. The CEDAR, MCYT, and
of 99.4 percent and 99.3 percent, respectively, demonstrating GPDS datasets are analyzed using a WI technique. The
the features learned by the network and demonstrating the experimental results revealed that when no validation is
ability of this configuration of neural network to be used in utilized during the optimization procedure, overfitting occurs,
applications for offline signatures identification and and that when a global validation method with an external
verification [3]. Suriani Mohd, the GPDS Synthetic Signature archive is applied, overfitting is reduced. In addition, the space
Database, the biggest accessible handwritten signature created following feature selection can be applied to transfer
dataset, was used in this work to classify the signatures of learning [9]. Sharvari, the report compares multiple deep
1000 users, each of whom had 24 genuine signatures and 30 learning models utilizing the Siamese architecture across a
or phoney signatures. In addition, two popular GoogLeNet large collection of signature photos. Based on the analysis of
architecture CNN versions were used, namely Inception-v1 the results, it can be concluded that on the CEDAR and Kaggle
and Inception-v3. To begin, algorithms were trained on datasets, a Siamese network utilizing VGG16 and Gaussian
samples from 20 users, with Inception-v1 achieving an Nave Bayes as the binary classifier detects signature fraud
accuracy of 83 percent and Inception-v3 achieving a with a maximum accuracy of 100 percent. In the ICDAR 2011
validation accuracy of 75 percent. In terms of Equal Error SigComp dataset of handwritten signatures, ResNet50 has the
Rates (EER), Inception-v1 achieved an EER of 17 for a total best accuracy of 98.29% for detecting forgery in Chinese
of 20 users. While the EER for Inception-v3 with 20 users was signatures. MobileNetV2 had 96 percent accuracy using the
24, which is an excellent measure when compared to previous cosine distance function for the CEDAR dataset, and
studies. Although Inception-v3 outperformed Inception-v3 in DenseNet121 had 98.5 percent accuracy using the cosine
the ImageNet picture classification challenge, Inception-v1 distance measure for the Kaggle dataset. Using the Euclidean
outperformed Inception-v3 in the categorization of 2D distance function, Exception had the best accuracy of
signature images. In this study, it is also accepted that 77.6percent on the ICDAR 2011 SigComp Chinese dataset
Inception-v1 spent less time training because it had lesser [10]. Rasheed, Ali Fathel, this paper examines the
operations than Inception-v3 [4]. Navid, the proposed method performance of a variety of characteristic and descriptor
in this research will aid persons in distinguishing signatures in detectors, including SIFT, SURF, ORB, BRIEF, BRISK, and
order to determine whether they are fabricated or genuine. FREAK. In terms of the number of features, the number of
And used Convolutional Neural Network in the system to matching points in areas overlapping between images and
automate the signature verification procedure. The model is subjective stitching accuracy were investigated. The deletion
built on top of the VGG-19, a pre-trained Convolutional of the image's many symmetrical properties increases the
Neural Network. They tested the model on widely verified likelihood that distinct scene views will reliably meet the
signature datasets with a variety of authentic signature requirements. According to one experiment, AGAST, FAST,
samples supplied from ICDAR (3), CEDAR (1), and Kaggle and BRISK detection have the highest number of detected key
(2), and achieved accuracy of 100%, 88.4%, and 94.44 points, whereas STAR, AKAZE, and MSER have the lowest
percent, respectively. According to research findings, the number of extracted key points. In addition, the speed of each
suggested approach can classify signatures that do not nearly method is noted. It achieved an accuracy of 98.8 percent [11].
match real signatures [5]. Foroozandeh, goal is to evaluate the Gornale, in this study, researchers review the most up-to-date
performance of prominent available convolutional neural approaches and applications for assessing neurological
networks in offline handwritten signature verification as disorders utilizing machine learning based on handwriting
features extractor using transfer learning and to review the signatures analysis. In addition, the scientific community
presented signature verification method based on should discuss its accomplishments and difficulties. As a
convolutional neural networks. This is accomplished four pre- result, an integrated discussion of various datasets used,
trained models, including VGG16, VGG19, ResNet50, and feature extraction approaches, and classification systems for
InceptionV3, which are the most often used general models in (PD) Parkinson's and (AD) Alzheimer's diseases is carried out
computer vision applications, as well as two pre-trained and scientifically surveyed. The goal of this research article is
models specifically designed for signature processing tasks, to give a comprehensive overview of scientific literature,
SigNet and SigNetF. And VGG16-99.79 percent accuracy identify vulnerable challenges, and propose new research
was the best [6]. Eman Alajrami, static and dynamic signature avenues in the subject. outlined the state-of-the-art approaches
verification are the two types of signature verification. and invitations for Assessment of Neurological
Dynamic (on-line) verification occurs as a person creates his Disorders Using Machine Learning Methods achievements
or her signature on a digital tablet or similar device, whereas and trials/problems that need be handled to the scientific
static(off-line) verification occurs after an electronic or community [12]. Zhongcheng Wu, the Multi-Path Attention
document signature has been created. In this research, Python Siamese Convolution Network is a novel deep learning model
is used to develop a CNN model for offline signatures, and that consists three modules: the generation module of different
after training and verifying, the testing accuracy was 99.70 size inputs, which ensures that the signature strokes are the
percent [7]. Duth, in this project can recognize cursive same size as the background image; the feature extraction
handwriting in the following styles: San-Serif, Comic Sans, module based on the fusion attention mechanisms, which
Tahoma, and Calibri, both handwritten and printed. The focuses on stroke features; and verification module with
above-mentioned style's upper- and lower-case text will be optimal weight coefficients, which improves the mode. On the
recognized by the approach [8]. Souza, this research explores three public datasets, the experimental findings reveal that

Authorized licensed use limited to: Lakehead University. Downloaded on February 28,2024 at 06:28:44 UTC from IEEE Xplore. Restrictions apply.
adjusting several aspects such as the multi-branch structure, IV. PROPOSED METHODOLOGY
the fusion attention technique, and the weight coefficients The handwritten signature is a behavioural biometric that
improves the accuracy of offline signature verification is based on changing behavior rather than any physical aspects
(CEDAR, Bengali, Hindi) [13]. Ashok Kumar, the hand of the individual signature. Because a person's signature
written signature recognition is a crucial biometric technique changes over time, verification system and authentication may
for determining if a mark is authentic or not. It is required in take a lengthy time, resulting in increased errors in some
order to prevent report falsification and to ensure the legality circumstances. Inconsistent signatures result in a higher rate
of documents such as draughts, identity, visas, and current of false rejection for those who do not sign consistently.
financial activities. The goal of research is to use
Convolutional Neural Networks to automate the process of A. Image acquisition
handwritten signature recognition. The proposed model is
based on the VGG16 plan, and prepared model with move A digital image scanner is used to capture the signature
learning using own dataset. With the use of my restricted image. The desktop (or flatbed) scanner, in which the
dataset, and can virtually always determine if a given mark document is laid on a glass pane for scanning, is widely used
was falsified or genuine [14]. Bipin Nair, as these issues in offices. Hand-held scanners have progressed from text
confront document enhancement, the suggested research scanning "wands" to 3D scanners for industrial design, reverse
attempts to recover damaged text sections that are affected by engineering, test and measurement, orthotics, gaming, and
stain markings, ink seepages, and document fading in other uses. Mechanically driven scanners that move the
document images [15]. Rani, due to the presence of vowels, document are widely utilized because a flatbed design would
consonants, and modifiers, automated reading of handwritten be problematic for large-format papers.
Kannada documents is extremely difficult. The difficulty of B. Data Processing
machine-based reading of handwritten vowels and consonants
is exacerbated by the diverse nature of handwriting styles [16]. Before extracting the features from CNN network, some
Bird, Jordan J, the architecture of convolutional neural data preprocessing is required: background noise reduction,
network and a data augmentation approach are investigated data augmentation, and screen. By splitting a single signature
and tweaked, yielding an 87.12 percent accurate model for image into numerous subframe blocks and filtering the sub
verification of 2,640 human signatures. Two robots are then image as the training dataset with appropriate strike pixels,
given the task of forging 50 signatures, 25 of which will be and may enhance the number of samples. In the preprocessing,
used in the verifications attack and the other 25 will be used the following steps are shown:
to tune the model to defend against them. The Line-us robotic • The raw image is converted to grayscale and saved as
arm can trick the system 24 percent of the time, and iDraw 2.0 a BMP 24-bit file as the first step in Fig. 2
robot can fool the system 32 percent of the time, indicating
that there is an information security risk. With roughly 30 • As the second phase, utilize a block-based data-
percent of fraudulent signatures misclassified as real, a increase method. And create a slider window to get a
conditional GAN obtains equal success [17]. sub-image block from the original image.

III. ARCHITECTURE • Next lightened the frame by 7.5 percent to eliminate

the paper texture and background noise from the
optical scanning device, and raised the RGB value to
1,075.
• Rotate each sub-image block in the clockwise direction
and repeat the process until the entire 360-degree
rotation is accomplished. Genuine training signature,
fabricated training signature, genuine verification
signature, fraudulent verification signature, real testing
signature, and forged trial signature are the six sub
datasets in the experiment. Each sub-image block will
be evaluated to see if there is enough information to
extract the characteristic and identify experiences. A
pixel over the grey border is considered real if it
exceeds a threshold value of 250. Then identify a block
Fig. 1. Proposed System as a valid sample if it has more than 7.5 percent correct
pixels. On the other hand, consider it an unsatisfactory
In the above Fig. 1 the system, trying to explain that when sample and discard it.
a person signs a document of system checks if the signature is
authentic by extraction of the image content by pixel-by-pixel
analysis. Where we have used Google collab and the CNN
model data set which is extracted from SigComp2009 and
three major models used in this process are as follows 1.
VGG16 2. VGG19 and 3. ResNet-50 after the pixel-based data
extraction and comparison of the extracted feature system
verifies weather the signature is real or forge. Also extracting
the features and accuracy of each signature. And to see which
model perform the best and gives more accuracy.

Authorized licensed use limited to: Lakehead University. Downloaded on February 28,2024 at 06:28:44 UTC from IEEE Xplore. Restrictions apply.
 cerebral cortex pyramidal cell structures. Between
nonlinearities (ReLU) and batch normalizations, the
bulk of ResNet models use double- or triple-layer
skips. And the accuracy was 96.23 percent.
D. Dataset
A dataset is a collection of information. The dataset was
collected from SigComp2009 dataset and also taken the friend
help to collect the handwritten signatures. Merging both 3000
datasets and dividing equally into Training, Testing, and
Validation.
The design of a system is divided into two stages;
A) Training stage
B) Testing stage
C) Validation stage
A. Training Stage- These images will be used to train the
model to recognize them. Image pre-processing, retrieval of
a signature image from a dataset Models training, feature
extraction.
B. Testing Stage- processing these images to see if the
trained model is giving accurate results based on training set.
Obtaining a signature from a database to be tested, pre-
processing of images, Extraction of features, extracting
features and applying them to a trained model.
C. Validation Stage-Validating the model for the best
result. Obtaining a signature from a database for validation,
Fig. 2. Preprocessing to increase the sample size
pre-processing of images, Extraction of features, Validation
of the model using retrieved characteristics.
C. Deep Learning Models
V. RESULT
There are three models applied to compare best model In this method, an image is passed through a sequence of
for Handwritten Signatures. alternating convolution and max pooling layer. When an
* VGG16- ImageNet is a vast visual database that is image is convolutional, a preset number of feature maps are
utilized in the development of visual object recognition generated, which are fed into a max pooling layer, which
software. VGG16 is a simple and commonly used generates pooled feature maps from the feature maps gained
CNN Architecture. VGG-16 is a deep CNN with 16 from the convolution layer before it. The following
layers. You can use ImageNet database to import pre- convolution layer receives this pooling layer map, and so on
trained version of network that has been trained over until the fourth max pooling layers is reached. The feature map
million photos. The activation functions utilized in the of the last layer is flattened and placed in the fully connected
implementation of VGG16 are SoftMax and relu. And layers. After many rounds of forward and backward
95.1 percent accuracy was achieved. propagation, the model has been trained, and a forecast may
now be made. The equations for all of the CNN steps are listed
* VGG19- VGG-19 CNN is used as a pre-processing in Table 1.
model. In comparison to standard convolutional neural
networks, the network depth has been enhanced. TABLE I. LIST OF FORMULAS FOR THE OPERATIONS IN THE CNN
Because it alternates between numerous convolutional
Operation Formula
layers and nonlinear activation layers, it has a better
Max Pooling hl xy = max i=0...s, j=0.s hl-1 (x+i)
structure than a single convolutional layer. The layer (y+j)
structure enables superior image extraction of features, ReLu ReLU(zi)= max (0, zi)
downsampling with Maxpooling, and change of linear Convolution zl = hl-1 * Wl
unit (ReLU) as activation function, which picks the Fully-connected layer z = ∗ h −1
image region's largest value as the pooled value of the Softmax softmax(zi) = ezi / ∑j ezj
area. The down - sampling layer is used to improve the
image's anti-distortion capabilities while keeping the
Many signatures are used in system’s testing, training, and
sample's key features and reducing the number of
validation. A total 3000 signatures were used in these studies
parameters. And 97.78 percent accuracy was achieved.
findings. Those 3000 signatures are made up of 300 sets (i.e.,
* Resnet50- ResNet-50 is a 50-layer fully convolutional 300 different people), with 5 examples of genuine signatures
neural network. A residual neural net (ResNet) is a and 5 samples of forgeries for each person. A portion of this
form of artificial neural network (ANN) based on database was used to train the algorithm, with 5 authentic

Authorized licensed use limited to: Lakehead University. Downloaded on February 28,2024 at 06:28:44 UTC from IEEE Xplore. Restrictions apply.
samples acquired from each of the 300 different persons and Letters, vol. 80, pp. 84–90, Sep. 2016, doi:
5 forgeries created by different people each signature. Also 10.1016/j.patrec.2016.05.023.
feature extracting and accuracy of each signature. And [2] A. B. Jagtap, R. S. Hegadi, and K. C. Santosh, “Feature learning for
offline handwritten signature verification using convolutional neural
comparative of different signature verifications models to see network,” International Journal of Technology and Human Interaction,
which model perform the best and gives more accuracy. Table vol. 15, no. 4, pp. 54–62, Oct. 2019, doi: 10.4018/IJTHI.2019100105.
2Error! Reference source not found. shows VGG19 is best [3] J. O. Pinzón-Arenas, R. Jiménez-Moreno, and C. G. Pachón-Suescún,
suitable for the dataset, which has 97.83% accuracy than other “Offline signature verification using DAG-CNN,” International
models. Journal of Electrical and Computer Engineering, vol. 9, no. 4, pp.
3314–3322, Aug. 2019, doi: 10.11591/ijece. v9i4.pp3314-3322.
In the below Table 3, handwritten signature verification [4] Jahandad, S. M. Sam, K. Kamardin, N. N. Amir Sjarif, and N.
classification of precision, recall, f1- score, and support will Mohamed, “Offline signature verification using deep learning
be given result by considering the contour and edges of the convolutional Neural network (CNN) architectures GoogLeNet
image. inception-v1 and inception-v3,” in Procedia Computer Science, 2019,
vol. 161, pp. 475–483. doi: 10.1016/j.procs.2019.11.147.
TABLE II. ACCURACIES OBTAINED FROM SPLITTING OF THE [5] S. M. A. Navid, S. H. Priya, N. H. Khandakar, Z. Ferdous, and A. B.
DATASET Haque, “Signature Verification Using Convolutional Neural Network,”
in 2019 IEEE International Conference on Robotics, Automation,
S.no Models Dataset Dataset Accuracy Artificial-Intelligence and Internet-of-Things, RAAICON 2019, Nov.
(SigComp2009) (SigComp2009) 2019, pp. 35–39. doi: 10.1109/RAAICON48939.2019.19.
Testing Validation [6] A. Foroozandeh, A. Askari Hemmat, and H. Rabbani, “Offline
1 VGG16 60 40 94.22 Handwritten Signature Verification and Recognition Based on Deep
40 60 93.63 Transfer Learning,” in Iranian Conference on Machine Vision and
2 VGG19 60 40 97.78 Image Processing, MVIP, 2020, vol. 2020-January. doi:
40 60 97.83 10.1109/MVIP49855.2020.9187481.
3 Resnet50 60 40 96.23 [7] E. Alajrami et al., “Handwritten Signature Verification using Deep
40 60 96.63 Learning,” 2019. [Online]. Available: www.ijeais.org/ijamr
[8] Vaigai College of Engineering and Institute of Electrical and
TABLE III. CLASSIFICATION REPORT Electronics Engineers, Proceedings of the International Conference on
Intelligent Computing and Control Systems (ICICCS 2020): 13-15
Precision Recall f1-score Support May, 2020.
1 0.96 0.95 0.96 50
0 0.00 0.00 0.00 50 [9] V. L. F. Souza, A. L. I. Oliveira, R. M. O. Cruz, and R. Sabourin, “An
accuracy 0.97 100 investigation of feature selection and transfer learning for writer-
Macro average 0.97 0.98 0.99 100 independent offline handwritten signature verification,” in Proceedings
Weighted average 0.95 0.96 0.93 100 - International Conference on Pattern Recognition, 2020, pp. 1475–
1482. doi: 10.1109/ICPR48806.2021.9413073.
VI. FUTURE WORK [10] A. R. Bhat, P. Nerella, P. Baburaj, and S. K. S, “A Comparative Study
of Transfer Learning Models for Offline Signature Verification and
We can still improve accuracy and evaluate a Forgery Detection,” 2AD.
greater number of models to compare which approach [11] A. F. Rasheed and A. M. Alkababji, “A Novel Method for Signature
performs best. In the Database Corpus, we can also Verification Using Deep Learning,” Webology, vol. 19, no. 1, pp.
incorporate digital signatures. 1561–1572, Jan. 2022, doi: 10.14704/web/v19i1/web19104.
[12] S. Gornale, S. Kumar, R. Siddalingappa, and P. S. Hiremath, “Survey
As compared, Offline signature verification, is on Handwritten Signature Biometric Data Analysis for Assessment of
not efficient and slow for the large number of papers. As a Neurological Disorder using Machine Learning Techniques,”
result, despite substantial study on signature verification, there Transactions on Machine Learning and Artificial Intelligence, vol. 10,
no. 2, pp. 27–60, Apr. 2022, doi: 10.14738/tmlai.102.12210.
is a need for greater attention and revision of the online
[13] X. Zhang, Z. Wu, L. Xie, Y. Li, F. Li, and J. Zhang, “Multi-Path
signature verification approach in order to improve efficiency Siamese Convolution Network for Offline Handwritten Signature
utilizing deep learning. Verification,” in ACM International Conference Proceeding Series,
Jan. 2022, pp. 51–58. doi: 10.1145/3512850.3512854.
VII. CONCLUSION [14] A. Kumar and T. Srinivasulu, “HAND WRITTEN SIGNATURE
As mentioned, this paper is a comparative study of three RECOGNITION USING DEEP LEARNING NETWORKS,” 2022,
doi: 10.17605/OSF.IO/9TSY3.
models, that is VGG16, VGG19, and Resnet50. Also using
several parameters to check which model is best suitable for [15] N. Shobha Rani, B. J. Bipin Nair, M. Chandrajith, G. Hemantha Kumar,
and J. Fortuny, “Restoration of deteriorated text sections in ancient
Handwritten Signatures Verification System. we conclude document images using a tri-level semi-adaptive thresholding
that VGG19 is best suitable for the dataset, which has more technique,” Automatika, vol. 63, no. 2, pp. 378–398, 2022, doi:
accuracy than other model. Considering that people live in an 10.1080/00051144.2022.2042462.
age of automation, it is essential that all of the job to be [16] N. Shobha Rani, N. Manohar, M. Hariprasad, and B. R. Pushpa,
completed quickly, easily, and successfully, with no errors or “Robust recognition technique for handwritten Kannada character
confusion. Introducing signature verification will help a lot of recognition using capsule networks,” International Journal of Electrical
and Computer Engineering, vol. 12, no. 1, pp. 383–391, Feb. 2022, doi:
people and eliminate physical documentation and everything 10.11591/ijece. v12i1.pp383-391.
can be digitalized. [17] J. J. Bird, “Robotic and Generative Adversarial Attacks in Offline
Writer-independent Signature Verification,” Apr. 2022, [Online].
REFERENCES Available: http://arxiv.org/abs/2204.0724
[1] A. Soleimani, B. N. Araabi, and K. Fouladi, “Deep Multitask Metric
Learning for Offline Signature Verification,” Pattern Recognition

Authorized licensed use limited to: Lakehead University. Downloaded on February 28,2024 at 06:28:44 UTC from IEEE Xplore. Restrictions apply.