You are on page 1of 5

2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI)

A Review of Signature Recognition Using Machine


Learning
Elizabeth Ann Soelistio Rafael Edwin Hananto Kusumo Zevira Varies Martan
School of Computer Science School of Computer Science School of Computer Science
Bina Nusantara University Bina Nusantara University Bina Nusantara University
Jakarta, Indonesia 11480 Jakarta, Indonesia 11480 Jakarta, Indonesia 11480
elizabeth.soelistio@binus.ac.id rafael.kusumo@binus.ac.id zevira.martan@binus.ac.id

Edy Irwansyah
School of Computer Science
2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI) | 978-1-6654-4002-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICCSAI53272.2021.9609732

Bina Nusantara University


Jakarta, Indonesia 11480
eirwansyah@binus.edu

Abstract—Signatures have been used for years for requirements. Additionally, results from literature reviews
transactions and consenting to responsibilities. Yet, online or and research questions will be reported in the results section.
offline, signatures can easily be falsified as there are no security This paper concludes by summarizing all the findings of all
measures in place to prevent this. Numerous researches have discussions and providing references to the literature
been carried out to find the most accurate and reliable signature
reviewed during its preparation.
recognition and verification system. This study examines the two
problems previously mentioned. A primary goal of this study is II. THEORITICAL BACKGROUND
to determine the best algorithms for recognizing signatures
based on the signature type. This systematic literature review is A. Offline Signature Recognition Algorithm
conducted using a PRISMA flow diagram. The results indicate
Offline signature faces difficulty in recognition and
that offline signatures mostly use Convolutional Neural
verification because of its non-digitized format. However,
Networks (CNN) for their recognition, while online signatures
use Recurrent Neural Networks (RNN) with other architectures.
there is much research related to offline signature using
models of machine learning and deep learning. Thus, all these
Keywords—Signature Recognition, Offline Signature, Online models will be explained respectively with its development
Signature, Machine Learning, Handwritten Signature and results.

I. INTRODUCTION • Convolutional Neural Network (CNN)


Signature is one of the most common things used in Convolutional Neural Network (CNN) is one of the
different transactions from the government to marketplace. commonly used methods in computer vision for
As one of the widely accepted forms of transactions publicly, image classification [1]. It shows as a research have
it is easy to forge an individual's signature and misuse it. A achieved an accuracy of 85-89% in forgery detection
and 90-94% signature recognition [2] while another
preventative measure taken is to identify the signature and
research achieves an accuracy of 98.8% [3].
verify it with a system that will give an answer if it is a forged
GoogLeNet architecture Inception-v1 and Inception-
or genuine signature. v3 have also been tested as it uses CNN model, and it
Researchers have been working on a preventative achieved an accurate validation of 83% and 75%
measure for a long time. Several approaches such as machine respectively [4]. By adding normalization, a research
learning and deep learning are being utilized to solve this has also achieved an accuracy of 96.41% and 98.30%
problem. In order to recognize and verify offline signatures, [5]. Another research uses CNN model with
they must be scanned and adjusted before being trained. approaches of Support Vector Machine (SVM) and
While online signature only needs to be adjusted in some Radial Basis Function Network (RBFN) achieve an
parts. In both cases, there are many different schemes that use accuracy of 96.6% [6]. When an extra feature
different approaches or combine two or more schemes. extraction method is added, CNN models can achieve
Nevertheless, consistency has its challenges since a 62.5% of success in Writer Independent model and
everyone's signature is affected by external factors such as a 75% success in Writer Dependent model [7].
mood, environment, and many more. Concerns with the DCNN, an improved Convolutional Neural Network
algorithms include the output accuracy, the time required for model, achieved a 100% image recognition rate [8]
the process, and the performance in real-life testing. The and was claimed the most powerful CNN model for
objective of this study is to determine the best model for both image recognition.
online and offline signature identification and verification.
This paper is structured as follows: it begins with a topic • Probabilistic Neural Network (PNN)
introduction, then it reviews literature to analyze other Probabilistic Neural Network (PNN) is often used for
approaches addressing the same problem, followed by a classification problems and follows the rules of
methodology section that explains the manner in which Bayes’ theorem in decision making [9]. PNN can be
literature was reviewed with a scope corresponding to the

978-1-6654-4002-8/21/$31.00 ©2021 IEEE

219 28 October 2021, Jakarta - Indonesia


Authorized licensed use limited to: Universidad Tecnica Federico Santa Maria. Downloaded on November 23,2022 at 12:55:53 UTC from IEEE Xplore. Restrictions apply.
2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI)

used with Wavelet Transform Average Framing verification model, it achieves an equal error rate
Entropy (AFE) as done by a research. This resulted in (EER) of 2.03% in a random forgery test [17]. The
an accurate recognition of 92% and verification of lower EER means the model’s accuracy is better.
83% [9].
• Relief Algorithm
• K-Nearest Neighbor (k-NN)
Relief algorithm is a filter-method towards feature
K-Nearest Neighbor is a classification algorithm that selection that is sensitive to interactions [18]. A model
assigns a class to the tested image based on its feature research achieves an average error rate of 5.31% in
values [10]. This research paper adds a feature of local false positive, meaning the model approves of the
binary pattern that allows extraction of texture-based signature as genuine instead of forged [18].
features and is closely related to face recognition [10].
The highest accuracy by using this model is 84.29%. III. METHODOLOGY
This systematic literature review uses PRISMA Flow
• Efficient Fuzzy Kohonen Clustering Network
Diagram. The authors use this because it improves the
(EFKCN)
reporting systematic reviews that are relevant to the topic of
Efficient Fuzzy Kohonen Clustering Network the literature review. The making of this literature review can
(EFKCN) is mostly used for classification by using be seen below:
data clustering [11]. The result of this research shows
70% accuracy in recognition compared to its previous
results of 53% [11].
• Artificial Neural Network (ANN)
ANN is widely used in pattern recognition as it is
powerful and easy to use. Features from the signature
can be selected based on the pattern recognized, then
it will learn the correlation between the classes and
signatures [12]. The research done by using this
algorithm has achieved a success rate of 95% on
average in recognition and verification system [12].
This is done by using MATLAB to design the system.
• Support Vector Machine (SVM)
This research is done by using a multiclass SVM
classifier on radial basis function (RBF) for training
and testing. The result for SIFT with SVM-RBF
achieved an accuracy of 98.75% while SURF with
SVM-RBF achieved 96.25% [13]. [14] uses SVM
with AlexNet as feature extraction and achieved 80%-
100% accuracy in signature recognition for different
datasets.
B. Online Signature Recognition Algorithm
Online signature is widely used in authentication
Figure 1. Prisma Flowchart
technology by using a touchscreen input interface. The
growing popularity in online signatures shows that people are All papers in this review must be written in English and
still familiar with the offline way of signing while putting a must have the following topics: Handwritten Signature
disadvantage at forgery that will attempt to imitate the Recognition, Offline Signature Recognition, Online Signature
signature. The algorithm used for this signature recognition Recognition, or Machine Learning. A second requirement is
and verification vary from Long-term Recurrent that papers have to be published within the year of 2017 up to
Convolutional Network (LRCN) to Weighting Relief 2021. Papers are included or excluded based on their
algorithm. The two models will be explained with its relevance to the topic that is being reviewed, and papers are
development and results. excluded for non-English language, duplicates, and
• Recurrent Neural Network (RNN) unsubstantiated studies. Included and excluded papers were
selected in accordance to these criteria for use in the review
RNN is mostly used in sequential modeling and can paper.
handle a large dataset. A research has improved this
model and shows a 2.37% equal error rate (EER) [15] Records are selected based on 4 steps, which are:
while another research results in below 2.0% EER 1. A database search was conducted using Google
[16]. Scholar, IEEE, Springer Link, and other libraries.
The keywords used to search were (1) Offline
• Long-term Recurrent Convolutional Network Signature Recognition, (2) Online Signature
(LRCN)
Recognition, (3) Handwritten Signature
LRCN is commonly used for visual recognition and Recognition, and (4) Handwritten Signature
description [17]. By using LRCN as a recognition and Biometrics.

220 28 October 2021, Jakarta - Indonesia


Authorized licensed use limited to: Universidad Tecnica Federico Santa Maria. Downloaded on November 23,2022 at 12:55:53 UTC from IEEE Xplore. Restrictions apply.
2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI)

2. The screening process eliminates duplicates and Table 2. Publication of algorithm used for online signature.
journal articles that don't correspond to the topic at
Signature Algorithm Number of Paper ID
hand. Type Paper
3. The introduction and abstract are read to determine
eligibility; it eliminates records that are irrelevant to CNN 10 [1], [2], [4],
[5], [7], [8],
the research questions. [20], [21], [22]
4. Included journals are chosen if they provide key or [36]
supporting information to help in the preparation of CNN + SVM 5 [3], [6], [19],
the review paper. [23], [24]
Offline PNN 1 [9]
A. Data extraction and synthesis
KNN 1 [10]
The data are extracted by reading the full text of the journal
then analyzed according to the relevancy for the purpose of EFKCN 1 [11]
writing this paper. The extracted items are: ID, Reference, ANN 2 [12], [36]
Context, Methodologies, and Topic. These are the description
of extracted items. SVM 2 [13], [14]

Table 1. Data Extraction of Each Study


Based on the information extracted, Convolutional Neural
Extracted Data Description
Network (CNN) is the most used algorithm for offline
ID Unique identity of each record. signature recognition and verification. Convolutional Neural
Reference Authors, year of publication, title of
Network is effective in system recognizing because it extracts
the record, and publication location of relevant data for classifications [2]. Many researches use
the reviewed journals. variation of multilayer perceptron (MLP) and introducing
Context The purpose and context of the paper
feature extractions such as LS2Net [5]. An improvement of
written. this model if Deep Convolutional Neural Network (DCNN)
using AlexNet and VGG16 for feature extractions which
Methodologies The methods, models algorithms, and
dataset used in the writing process. deliver a perfect result of 100% accuracy [8]. The dataset used
in [8] is a personal dataset of 600 signatures. However, this
Topic The topic related to the writing of the shows that Convolutional Neural Network (CNN) presents
records.
good results in recognizing signatures and delivering
accuracy.
IV. RESULT
Table 3. Publication of algorithm used for online signature.
The results presented are based on the data extraction
performed previously and the items listed on the previous Signature Algorithm Number of Paper ID
section. Data extraction is done to obtain results of critical Type Paper
literature review based on research questions determined RNN 6 [15], [16], [25],
below. [26], [27], [28]
A. What challenges does signature recognition face? LRCN 1 [17]
Signature recognition faces many challenges in real life Relief 1 [18]
because of the inconsistency of the signatures from time to Algorithm
time [19]. This inconsistency depends on the environment Online
CNN 2 [29], [30]
and human behavior at the time, thus making it a challenge to
create a model to differentiate a genuine and skilled forgery K-ANN 1 [31]
signature. This problem can lead to either a strict model that ANN 1 [32]
only allows the exact same of signature to be recognized as
DTW Cost 2 [33], [34]
genuine, or a model that is too lenient that allows different Matrix
signature to be labeled as genuine.
Furthermore, information loss is another challenge offline
signature recognition must face because of the digitalization Meanwhile, offline signature recognition mostly uses
process of the signature [1]. The signature must be scanned Recurrent Neural Network (RNN). The RNN architecture falls
before it can become a digital image. This create a loss in under machine learning whereas CNN is a deep learning
dynamic information such as position and velocity of the pen, architecture. The difference between CNN and RNN is the
pressure and stroke of the signature, which can determine the learning techniques each uses to learn the pattern. CNN model
genuineness of a signature. usually uses a supervised model while RNN has the ability to
learn in both supervised and semi-supervised methods [35].
B. What are the most used algorithm for signature Thus, if the method and combination are used properly, RNN
recognition? can produce an acceptable result.
The table below contains extractions information from 34
selected papers and journals and 3 review papers. This
information is divided by the type of signatures and extracted
based on the first research question.

221 28 October 2021, Jakarta - Indonesia


Authorized licensed use limited to: Universidad Tecnica Federico Santa Maria. Downloaded on November 23,2022 at 12:55:53 UTC from IEEE Xplore. Restrictions apply.
2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI)

C. What is the most used feature extractions in CEDAR dataset [20], and [8] using AlexNet and VGG16
Convolutional Neural Network (CNN) for offline feature extractions. The lowest accuracy with 75% are
signature? extracted using GoogLeNet Inception-v3 with GPDS dataset
Feature extractions is the backbone of these algorithm [4].
used for recognizing the signatures. The process increases the
V. CONCLUSION
efficiency of processing without losing any relevant
information while simultaneously reducing the amount of This systematic literature reviews observes and identify
redundancy [36]. Nevertheless, not all algorithms use feature the implemented methods and models that are used for
extractions [37]. This is feature extraction from offline signature recognition and verification. The studies used in the
signatures with the CNN algorithm. writing of this paper are selected based on PRISMA Flow
Diagram with an initial start of 51 studies collected from
Table 4. Feature extraction methods for convolutional neural online databases to 37 relevant studies that are selected to be
network (CNN) algorithm in offline signature recognition. analyzed further. The aim of this systematic literature review
is to determine which algorithms are commonly used for
Algorith Paper Feature Dataset Result offline and online signature recognition. This study focuses on
m ID Extraction
two recognition algorithms, one for each type of signature,
SigNet-F GPDS-160 EER = 1.72% which are Convolutional Neural Network (CNN) for offline
(SVM)
CEDAR EER = 4.63% and Recurrent Neural Network (RNN) for online signature.
[1]
Based on the offline signature recognition, CNN is
MYCT EER = 2.87%
commonly used across different datasets and feature
SigNet Brazilian EER = 2.01% extractions. CNN has been designed to run task for visual
PUC-PR
recognition in computer vision, thus able to perform well in
[2] Crest-Trough 1320 ACC = 94% offline signature since it is feed with images that have either
Algorithm pictures been preprocessed or not.
GoogLeNet ACC = 83% Nevertheless, RNN is mostly used for online signature
Inception-v1 recognition. Since RNN is a machine learning, it will need
[4] GPDS
Convolut GoogLeNet ACC = 75% time to learn and classify the dataset. However, it still gives an
ional Inception-v3 acceptable result depending on the combination used.
Neural
Network MYCT ACC = 96.4% From the analysis of the eligible studies, there are
(CNN) [5] LS2Net
important factors in implementing offline and online signature
CEDAR ACC = 98.30%
recognition in real environments. It must be fulfilled by
LS2Net-v2 GPDS-4000 ACC = 96.91% algorithm equipped to handle a large-scale and noisy dataset.
[8] AlexNet 600 ACC = 100% This algorithm must be tested in real scenarios before
signatures implemented into the real environments. This must be done to
VGG16 ACC = 100%
avoid having miscalculation in the real environment. In
CEDAR ACC = 100% addition, for offline signature type, it must have a standardized
BHSig260 ACC = 97.77% image quality to achieve the same effectiveness for the feature
(in Bengali) extraction method.
[20] InceptionSVG
Net BHSig260 ACC = 95.40%
(GoogLeNet (in Hindi) REFERENCES
Inception-v1) [1] L. G. Hafemann, R. Sabourin, and L. S. Oliveira, “Learning features
Persian ACC = 80.44%
for offline handwritten signature verification using deep convolutional
UTSig
neural networks,” in Pattern Recognit., vol. 70, pp. 163–176, 2017, doi:
[21] ConvNet 300 ACC = 99.7% 10.1016/j.patcog.2017.05.012.
(CNNs) signatures [2] J. Poddar, V. Parikh, and S. K. Bharti, “Offline Signature Recognition
and Forgery Detection using Deep Learning,” in Procedia Comput.
Sci., vol. 170, no. 2019, pp. 610–617, 2020, doi:
Equal error rate (EER) represents a threshold value in 10.1016/j.procs.2020.03.133.
comparison to false acceptance rate (FAR) and false rejection [3] V. L. F. Souza, A. L. I. Oliveira, and R. Sabourin, “A writer-
rate (FRR) [35]. The threshold is commonly used to compare independent approach for offline signature verification using deep
convolutional neural networks features,” in Proc. - 2018 Brazilian
and evaluate biometric authentication systems [35]. The Conf. Intell. Syst. BRACIS 2018, pp. 212–217, 2018, doi:
decreasing of EER values, the model becomes more accurate. 10.1109/BRACIS.2018.00044.
Based on accuracy (ACC), the system can identify whether a [4] Jahandad, S. M. Sam, K. Kamardin, N. N. Amir Sjarif, and N.
signature is fake or authentic. For ACC, the highest number Mohamed, “Offline signature verification using deep learning
is preferred as it displays the level of accuracy of the model. convolutional Neural network (CNN) architectures GoogLeNet
inception-v1 and inception-v3,” in Procedia Comput. Sci., vol. 161, pp.
CNNs use feature extraction and classification as part of 475–483, 2019, doi: 10.1016/j.procs.2019.11.147.
their architectures. Features can be extracted from a dataset [5] N. Çalik, O. C. Kurban, A. R. Yilmaz, T. Yildirim, and L. Durak Ata,
in order to reduce the number of features, while classification “Large-scale offline signature recognition via deep neural networks
is the likelihood that an input will lead to an outcome. and feature embedding,” in Neurocomputing, vol. 359, pp. 1–14, 2019,
doi: 10.1016/j.neucom.2019.03.027.
Furthermore, the difference in feature extraction can lead to
[6] M. Hanmandlu, A. B. Sronothara, and S. Vasikarla, “Deep Learning
accuracy and overfitting improvements depending on the based Offline Signature Verification,” in 2018 9th IEEE Annu.
complexity of the datasets tested. The highest accuracy Ubiquitous Comput. Electron. Mob. Commun. Conf. UEMCON 2018,
achieved using CNN is 100% with InceptionSVGNet with pp. 732–737, 2018, doi: 10.1109/UEMCON.2018.8796678.

222 28 October 2021, Jakarta - Indonesia


Authorized licensed use limited to: Universidad Tecnica Federico Santa Maria. Downloaded on November 23,2022 at 12:55:53 UTC from IEEE Xplore. Restrictions apply.
2021 1st International Conference on Computer Science and Artificial Intelligence (ICCSAI)

[7] M. M. Yapici, A. Tekerek, and N. Topaloglu, “Convolutional Neural CNNs,” in Expert Syst. Appl., vol. 133, pp. 317–330, 2019, doi:
Network Based Offline Signature Verification Application,” in Int. 10.1016/j.eswa.2019.03.040.
Congr. Big Data, Deep Learn. Fight. Cyber Terror. IBIGDELFT 2018 [23] T. Younesian, S. Masoudnia, R. Hosseini, and B. N. Araabi, “Active
- Proc., no. November 2019, pp. 30–34, 2019, doi: Transfer Learning for Persian Offline Signature Verification,” in 4th
10.1109/IBIGDELFT.2018.8625290. Int. Conf. Pattern Recognit. Image Anal. IPRIA 2019, pp. 234–239,
[8] A. Hirunyawanakul, S. Bunrit, N. Kerdprasop, and K. Kerdprasop, 2019, doi: 10.1109/PRIA.2019.8786013.
“Deep Learning Technique for Improving the Recognition of [24] S. V. Bonde, P. Narwade, and R. Sawant, “Offline Signature
Handwritten Signature,” in Int. J. Inf. Electron. Eng, vol. 9, no. 4, 2019, Verification Using Convolutional Neural Network,” in 2020 6th Int.
doi: 10.18178/ijiee.2019.9.4.709. Conf. Signal Process. Commun. ICSC 2020, pp. 119–127, 2020, doi:
[9] K. Daqrouq, H. Sweidan, A. Balamesh, and M. N. Ajour, “Off-line 10.1109/ICSC48311.2020.9182727.
handwritten signature recognition by wavelet entropy and neural [25] R. Tolosana, R. Vera-Rodriguez, J. Fierrez, and J. Ortega-Garcia,
network,” in Entropy, vol. 19, no. 6, 2017, doi: 10.3390/e19060252. “Exploring Recurrent Neural Networks for On-Line Handwritten
[10] T. Jadhav, “Handwritten Signature Verification using Local Binary Signature Biometrics,” in IEEE Access, vol. 6, pp. 5128–5138, 2018,
Pattern Features and KNN,” in Int. Res. J. Eng. Technol., vol. 6, no. 4, doi: 10.1109/ACCESS.2018.2793966.
pp. 579–586, 2019, [Online]. Available: www.irjet.net. [26] R. Vera-Rodriguez, R. Tolosana, M. Caruana, G. Manzano, C.
[11] D. Suryani, E. Irwansyah, and R. Chindra, “Offline Signature GonzalezGarcia, J. Fierrez and J. Ortega-Garcia, “DeepSignCX:
Recognition and Verification System using Efficient Fuzzy Kohonen Signature Complexity Detection using Recurrent Neural Networks,” in
Clustering Network (EFKCN) Algorithm,” in Procedia Comput. Sci., Proc. International Conference on Document Analysis and
vol. 116, pp. 621–628, 2017, doi: 10.1016/j.procs.2017.10.025. Recognition, 2019
[12] A. U. Rehman, S. ur Rehman, Z. H. Babar, M. K. Qadeer, and F. A. [27] C. Li, X. Zhang, F. Lin, Z. Wang, J. Liu, R. Zhang, and H. Wang, “A
Seelro, “Offline Signature Recognition and Verification System Using stroke-based RNN for writer-independent online signature
Artificial Neural Network,” in Univ. Sindh J. Inf. Commun. Technol., verification,” in Proc. Int. Conf. Doc. Anal. Recognition, ICDAR, pp.
vol. 2, no. 1, pp. 73–80, 2018. 526–532, 2019, doi: 10.1109/ICDAR.2019.00090.
[13] A. T. Nasser and N. Dogru, “Signature recognition by using SIFT and [28] C. Nathwani, “Online Signature Verification Using Bidirectional
SURF with SVM basic on RBF for voting online,” in Proc. 2017 Int. Recurrent Neural Network,” in Proc. Int. Conf. Intell. Comput. Control
Conf. Eng. Technol. ICET 2017, vol. 2018-Janua, pp. 1–5, 2018, doi: Syst. ICICCS 2020, no. Iciccs, pp. 1076–1078, 2020, doi:
10.1109/ICEngTechnol.2017.8308208. 10.1109/ICICCS48265.2020.9121023.
[14] K. Kamlesh and R. Sanjeev, “Offline Signature Recognition Using [29] C. S. Vorugunti, R. K. S. Gorthi, and V. Pulabaigari, “Online signature
Deep Features,” in Lect. Notes Networks Syst., vol. 141, pp. 405–421, verification by few-shot separable convolution based deep learning,”
2021, doi: 10.1007/978-981-15-7106-0_18. in Proc. Int. Conf. Doc. Anal. Recognition, ICDAR, pp. 1125–1130,
[15] S. Lai, L. Jin, and W. Yang, “Online Signature Verification Using 2019, doi: 10.1109/ICDAR.2019.00182.
Recurrent Neural Network and Length-Normalized Path Signature [30] C. S. Vorugunti, G. S. Devanur, P. Mukherjee, and V. Pulabaigari,
Descriptor,” in Proc. Int. Conf. Doc. Anal. Recognition, ICDAR, vol. “OSVNet: Convolutional siamese network for writer independent
1, no. 1, pp. 400–405, 2017, doi: 10.1109/ICDAR.2017.73. online signature verification,” in Proc. Int. Conf. Doc. Anal.
[16] R. Tolosana, R. Vera-Rodriguez, J. Fierrez, and J. Ortega-Garcia, Recognition, ICDAR, pp. 1470–1475, 2019, doi:
“DeepSign: Deep On-Line Signature Verification,” in IEEE Trans. 10.1109/ICDAR.2019.00236.
Biometrics, Behav. Identity Sci., vol. 3, no. 2, pp. 229–239, 2021, doi: [31] R. Ravi Chakravarthi and E. Chandra, “Kernel based artificial neural
10.1109/tbiom.2021.3054533. network technique to enhance the performance and accuracy of on-line
[17] C. Y. Park, H. G. Kim, and H. J. Choi, “Robust Online Signature signature recognition,” in J. Internet Technol., vol. 21, no. 2, pp. 447–
Verification Using Long-term Recurrent Convolutional Network,” in 455, 2020, doi: 10.3966/160792642020032102013.
2019 IEEE Int. Conf. Consum. Electron. ICCE 2019, no. c, pp. 1–6, [32] D. I. Dikii and V. D. Artemeva, “Online handwritten signature
2019, doi: 10.1109/ICCE.2019.8662005. verification system based on neural network classification,” in Proc.
[18] L. Yang, Y. Cheng, X. Wang, and Q. Liu, “Online handwritten 2019 IEEE Conf. Russ. Young Res. Electr. Electron. Eng. ElConRus
signature verification using feature weighting algorithm relief,” in Soft 2019, pp. 225–229, 2019, doi: 10.1109/EIConRus.2019.8657134.
Comput., vol. 22, no. 23, pp. 7811–7823, Dec. 2018, doi: [33] A. Sharma and S. Sundaram, “On the Exploration of Information from
10.1007/s00500-018-3477-2 the DTW Cost Matrix for Online Signature Verification,” in IEEE
[19] A. B. Jagtap, R. S. Hegadi, and K. C. Santosh, “Feature learning for Trans. Cybern., vol. 48, no. 2, pp. 611–624, 2018, doi:
offline handwritten signature verification using convolutional neural 10.1109/TCYB.2017.2647826.
network,” in Int. J. Technol. Hum. Interact., vol. 15, no. 4, pp. 54–62, [34] Y. Jia, L. Huang, and H. Chen, “A two-stage method for online
2019, doi: 10.4018/IJTHI.2019100105. signature verification using shape contexts and function features,” in
[20] R. K. Mohapatra, K. Shaswat, and S. Kedia, “Offline Handwritten Sensors (Switzerland), vol. 19, no. 8, 2019, doi: 10.3390/s19081808.
Signature Verification using CNN inspired by Inception V1 [35] nehal hamdy al-banhawy, H. Mohsen, and N. Ghali Prof., “Signature
Architecture,” in Proc. IEEE Int. Conf. Image Inf. Process., vol. 2019- identification and verification systems: a comparative study on the
November, pp. 263–267, 2019, doi: online and offline techniques,” in Futur. Comput. Informatics J., vol.
10.1109/ICIIP47207.2019.8985925. 5, no. 1, p. 3, 2020.
[21] E. Alajrami, B. A. M. Ashqar, B. S. Abu-Nasser, A. J. Khalil, M. M. [36] S. Utkarsh and B. Vikrant, “Comparison between CNN and ANN in
Musleh, A. M. Barhoom, and S. S. Abu-Naser, “Handwritten Signature offline signature verification,” in Proc. Second Int. Conf. Comput.
Verification using Deep Learning,” in Int. J. Acad. Multidiscip. Res., Commun. Control Technol., vol. 4, pp. 136–140, 2018.
vol. 3, no. 12, pp. 39–44, 2019, [Online]. Available: [37] L. G. Hafemann, R. Sabourin, and L. S. Oliveira, “Offline handwritten
https://philarchive.org/archive/ALAHSV signature verification - Literature review,” in Proc. 7th Int. Conf. Image
[22] S. Masoudnia, O. Mersa, B. N. Araabi, A. H. Vahabie, M. A. Sadeghi, Process. Theory, Tools Appl. IPTA 2017, vol. 2018-Janua, pp. 1–8,
and M. N. Ahmadabadi, “Multi-representational learning for Offline 2018, doi: 10.1109/IPTA.2017.8310112.
Signature Verification using Multi-Loss Snapshot Ensemble of

223 28 October 2021, Jakarta - Indonesia


Authorized licensed use limited to: Universidad Tecnica Federico Santa Maria. Downloaded on November 23,2022 at 12:55:53 UTC from IEEE Xplore. Restrictions apply.

You might also like