Sana'a University Journal of Applied Sciences and Technology https://journals.su.edu.ye/index.

php/jast
Vol. 1 | No. 3 | Page 291 – 303 | 2023 | ISSN : 2958-9568
Doi: 10.59628/jast.v1i3.314

Developing a Model for Offline Signature Verification Using CNN

Architectures and Genetic Algorithm
Abdulbaset M. Qaid Musleh 1,*., Abdoulwase M. Obaid Al-Azzani1
1
Department of Computer Science, Faculty of Computer and Information Technology, Sana'a University, Sana'a, Yemen.
*Corresponding author: [email protected]
ARTICLE INFO KEYWORDS
Article history: 1. Offline Signature Verification
Received: March 13, 2023 2. Convolutional Neural Network (CNN)
Accepted: July 30, 2023 3. Deep Learning (DL)
Published:, August, 2023 4. Genetic Algorithm (GA)

ABSTRACT
The signature verification process has many applications, such as its use in financial operations, providing
the electronic signature of documents, and providing an additional confidentiality standard to verify the
identity of users in computer systems. This process has the advantage of being accepted by the community
and is less intrusive compared to other biological methods. Deep learning (DL) and CNN (Convolutional
Neural Network) are widely used by bioinformaticians. Due to the difficulty in extracting features in other
systems or models, DL and CNN-based signature verification systems have been significantly improved.
Yet Hyperparameter optimization for CNN models remains a challenging problem in designing highly
efficient models with the most accurate results. It is often convolutional neural network (CNN) models
that are manually designed. The proposed method is focused on a genetic algorithm that develops a
population of CNN models in order to find the best fit architecture for designing an offline signature
verification model. Our model is tested on more than one dataset, BHSig260-Bengali, BHSig260-Hindiin,
GPDS, and CEDAR. The result of the approach proposed in this paper has the highest discrimination rate
of FRR of 2.5, FAR 3.2, EER 2.35, and 97.73 %-accuracy rate.
CONTENTS
1. Introduction 5. Training And Testing Stage
2. Problem Definition 6. Experimental Results
3. Related Work 7. Conclusion.
4. Materials And Methods
1. Introduction:
categorized it into two main fields: online and
Biometric systems have been used to identify offline [41]. Online signature verification
individuals based on behavioral or physiological focuses on dynamic information captured
characteristics [39]. Verification, identification, during the writing process and does not consider
and watchlists are primarily and continuously static information, such as pen-up time,
used in biometrics. A handwritten signature is acceleration, or velocity. In contrast, offline
one of the most widely used tools in biometrics signature verification deals with static signature
for verifying secure personal authentication images and is generally considered more
[40]. Signature verification has been extensively challenging, resulting in lower accuracy than
studied over the last two decades, and it is still a online signature verification [1].
promising area of research. Previous studies on Word count: 4,996 words, excluding references.
handwritten signature image verification have

JAST Sana'a University Journal of Applied Sciences and Technology 291

 Abdulbaset M. Qaid Musleh et.al JAST

Funding Statement: The study was supported by depicts the forgery of three types of handwritten
grant NN from the Foundation of Basic Research. signatures.
This work was carried out under research program
NNN of NN University. Author NN was supported
by grant NN from the Ministry of NN.
Ethical Compliance: All procedures performed
in studies involving human participants were in
accordance with the ethical standards of the Figure 1:(a) Signature of Genuine (b) Forgery
institutional and/or national research committee and of Random (c) Forgery of Unskilled (d)
with the 1964 Helsinki Declaration and its later Forgery of Skilled.
amendments or comparable ethical standards.
Data Access Statement: Research data Two classifications of learning are used in
supporting this publication are available from the systems of Handwritten Signature Verification:
NN repository at located at
www.NNN.org/download/.
The writer-independent (WI) and another is
Conflict of Interest declaration: The authors known as writer-dependent (WD) [5], [6]. The
declare that they have NO affiliations with or learning that is performed by all signatures in
involvement in any organization or entity with any the database is known as the Writer-
financial interest in the subject matter or materials Independent (WI) state, but when the learning is
discussed in this manuscript. performed by individual signatures
Author Contributions: AB and MJ contributed to independently, this is known as the Writer-
the design and implementation of the research, JK Dependent (WD) state. In contrast, this made
to the analysis of the results and to the writing of the
manuscript. VK conceived the original and the WI method more popular and was used
supervised the project. when building the WI system; it is easy to add a
Despite its low accuracy, offline signature new person because the classification in the
verification offers several advantages. It does system depends on one category for all persons
not rely on special input devices, making it [7], [8]. Recently, many automated systems
accessible and applicable to a wide range of have been designed to verify the validity of
scenarios. In addition, offline signature handwritten signatures through different
verification encompasses various domains, algorithms and methods, and deep learning has
further expanding its potential applications and the largest share of these systems. This is
relevance. Thousands of financial documents because of the efficiency and ability of deep
and business transactions are continuously learning to classify images using different
authorized through signatures [2]. Hence, the methods and techniques in image processing.
main purpose of handwriting signature The best technique for deep learning is
verification systems is to detect genuine Convolutional Neural Networks [9], [10]. There
signatures (made by the owner writer) or forged is a significant difference in the efficiency of
signatures (made by a fraudulent person) [3]. many deep learning approaches in real-life
Generally, three types of forgeries exist in the applications [11], [12]. The performance and
signature verification field [4]. Unskilled efficiency of CNNs rely heavily on their
forgery can be committed by a person who architectures [13], [14]. Therefore, many
forges another person’s signature without experts in the field of deep learning have
knowing any information about that person. designed several different structures for CNN,
Random forgery can be made by a person who and they are named after their different names
knows the signer’s name without seeing the and versions. VGGNet, GoogleNet [15], ResNet
genuine signature. Skilled forgery can be made [16], CapsNet [17], and DenseNet [18].
by a person who knows the name and shape of However, it is difficult to obtain a CNN model
the genuine signature of the owner. Figure 1 that solves all classification problems.
Therefore, the design of the architecture CNN

JAST Vol. 1 | No. 3 | 2023 | 292

 JAST Abdulbaset M. Qaid Musleh et.al

depends solely on suitable and specific Convolutional Layer, the Activating function,
parameters to solve the problem by manual pooling layer, and fully connected layer. A filter
design through repeated attempts until the best (also known as a kernel) is applied in the
results are obtained. However, this process can convolution layer that determines the presence
take longer than expected [19]. The method of certain features or patterns in the original
proposed in this paper uses a genetic algorithm image (the input image matrix), and
for hypermetric optimization of the CNN subsequently, several filters can be used to
architecture for the offline signature verification extract different features. The idea behind the
problem. Pooling Layer is very simple: it reduces the size
of large matrices, and the process is performed
2. Problem Definition by applying one of the following two functions:
The signature identifies and proves the signer’s Max and Average. The fully connected layer is
identity. As a result, a person may face some the last layer in the convolutional network, and
difficulties in deciding whether the two is a multilayer perceptron. The neurons were
signatures are identical owing to work pressure fully connected to all the nodes of the previous
and some ambiguity associated with manual layer. The reason it exist, in the end is that the
signatures [40]. During commercial final classification process occurs in it. In other
transactions, there is a growing need in the words, the components of the last array are
public and private sectors, especially in the converted into an array with only one column.
banking system, for an efficient and accurate The method proposed in [22] for offline
automated system capable of verifying the signature verification uses the Siamese network,
genuine signature and detecting skillful forgery which is based on user writer-independent (WI)
of the signature [42]. The main problem is the feature learning. Euclidean distance is used to
automated validation of signatures and signature measure the similarity and dissimilarity between
feature extraction, which enables the system to pairs of Siamese network outputs. In addition, a
differentiate between genuine and forged Siamese Neural Network was used by [23] to
signatures [43]. Therefore, this study aims to build a signature verification system. The
highlight the effective signature features and structures of two similar neural networks were
verify whether the system can automatically trained and evaluated using the same data. The
verify the real signature and detect skilled Siamese network structure helps reduce the
forgery, where the method depends on volume of training data necessary for
optimizing the CNN model hyperparameters implementation. They observed that the system
using a Genetic Algorithm. A genetic algorithm efficiency increased by 13%. Two models for
was used to optimize the CNN architecture for predicting the strength of adhesively bonded
offline signature verification and forgery joints were designed using CNN, the
protection problems. The genetic algorithm architecture CNN: one was manually developed,
helps us select the best hyperparameters to build and the other was designed using a genetic
the best CNN model. algorithm to optimize the architecture. The
results of the two models were then compared.
3. Related Work They noted that the improved model using the
The science of artificial intelligence and its genetic algorithm had a better result, as
modern and contemporary techniques, including proposed in [24]. The following datasets,
deep learning, have been used in many CIFAR10, MNIST, and Caltech256, were used
applications, such as pattern recognition and to test the CNN architecture optimized models
natural language processing, the most important using a genetic algorithm to solve the image
of which is computer vision [21]. Convolutional classification problem. The genetic algorithm
neural networks consist of basic components: a worked by automatically adjusting the

293 JAST Vol. 1 | No. 3 | 2023 |

 Abdulbaset M. Qaid Musleh et.al JAST

parameters of the model, and through the testing representation scheme, are combined to make a
process of the model, they found that the results final decision regarding the validity of the
were more accurate than those of other models signature. The experimental results indicate that
that were tested with the same dataset [19]. In verification accuracy increases when feature-
the method proposed in [25], a genetic based classifiers are combined. The model was
algorithm was applied to choose the possible built to validate signatures through the transfer
solutions by selecting the following parameters: of learning and activation functions for the three
many filters, filter size, and the number of different CNN models (VGG16, VGG19, and
layers automatically added to the trainable ResNet50) with the addition of some parameters
layers of the CNN transfer model. The results for each model, training, and testing on
show that the proposed method achieved an SigComp2009 dataset, showing that the VGG16
accuracy of 97% in classifying cat and dog model has a high efficiency of 97%- compared
datasets over 15 generations. Reverse inference to other models this approach was proposed by
based on the results of the genetic algorithm [28]. Thus, this study focuses on developing a
showed that the proposed method can detect system for offline signature verification using a
gradient features in network layers, which plays CNN with a genetic algorithm that can search
a role in understanding the transfer of CNN for the best model architecture hyperparameters.
models. They developed a system [26] for the
recognition of finger veins using a CNN with a 4. Materials And Methods
genetic algorithm. The proposed system Genetic Handwritten offline signature verification is
Algorithm with a convolutional neural network a pattern recognition problem. For this purpose,
(GA-CNN) performs better in terms of the proposed model must be able to recognize
accuracy, sensitivity, and precision. The ability and verify genuine handwritten offline
of the Genetic Algorithm (GA) to search is signatures and detect the forgery skills that
utilized to start the training phase of a CNN exist in handwritten offline signatures.
before the model’s training process, rather than Therefore, the offline signature verification
random initializers for the weights of the model includes several stages. The main stages
network created using the genetic algorithm. In are preprocessing of the signature image,
the testing process system, the results were genetic algorithm for selecting the best
97.98% for sensitivity, 98.50% for accuracy, hyperparameters, and CNN for feature
and 99.01%, respectively. The selection of the extraction, training, and testing. Each stage
most suitable set of features for a signer remains consists of several steps. Figure 2 shows the
an open question. Genetic algorithms (GA) have main stages and steps of each stage of the
recently been used a genetic algorithm to select model.
the optimal set of partial curves from the
signature and features of each curve. The A. Signature Images Preprocessing
locations of the partial curves and the features Before starting feature extraction, essential
contained in the partial curves used for processes must be applied to the image
verification are encoded into the chromosome signature. These operations include the
[27]. The focus of the study proposed in [44] is following.
on offline signature verification, in which four 1) Convert the color images into grayscale.
different pattern representation schemes have 2) Resize each image to the same size of 100 x
been implemented. To weigh the individual 100 pixels.
feature components, their pattern 3) Read the image points and store them in an
characterization capabilities were determined image’s matrix.
using Genetic Algorithms. The conclusions of 4) Store the image class name or label name for
the four subsystems, each based on a each image in the label’s matrix.

JAST Vol. 1 | No. 3 | 2023 | 294

 JAST Abdulbaset M. Qaid Musleh et.al

B. The Genetic Algorithm solve the problem. The architecture of a CNN is

often manually optimized by researchers, but a
time-consuming process is difficult to obtain
without previous knowledge of the CNN [19].
Each output chromosome from the genetic
algorithm was tested on the (CNN) model to
calculate the accuracy. The steps of the
proposed algorithm are as follows:
1) Generate 6 initial chromosomes.
2) Compute the accuracy of each chromosome
by CNN model.
3) Sort Chromosomes in descending order based
on accuracy.
4) Select three-chromosome height accuracy.
5) Crossover Function is applied to
chromosomes.
Child1 ← Crossover (Chrom1, Chrom3).
Child2 ← Crossover (Chrom2, Chrom1).
Child3 ← Crossover (Chrom3, Chrom2).
6) Mutation Function is applied to a new child.
Child1 ← Mutation (Child1).
Child2 ← Mutation (Child2).
Child3 ← Mutation (Child3).
7) Accuracy ←Compute the accuracy of each
new child by CNN model.
8) If Accuracy >= 0.99 or Generation = N.
9) Stop: 10) Else 11) Repeat steps 2 - 7 12).
Figure2:Block Diagram for The System of
End.
Proposed. Table (1): Hyperparameter range

Genetic algorithms are population-based search Hyperparameter Range

algorithms that are well suited for solving high- Epoch Random (2, 25)
dimensional and non-convex optimization Filter Size Choice (16,32,64, 96)
problems. In this study, the genetic algorithm Kernel Size Choice [(3x3), (5x5)(7x7)]
Unit Choice (128, 256,512)
helped us to optimize the best hyperparameters Dropout Choice (0,25, 0.50)
to build the best CNN model. However, the
performance depends on the architecture used to
Table (2): Random Generated Initial Population

layer 1 layer 2 layer 3 layer 4 Epoch Dropout Flatten Fitness%

Fitness%
Dropout

Flatten
Epoch
Kernel

Kernel

Kernel

Kernel
Filters

Filters

Filters

Filters

32 3 32 3 64 5 32 5 10 0.50 128 60
64 3 96 7 32 3 32 3 15 0.25 256 58
64 3 64 3 64 5 64 7 12 0.25 512 70
32 3 32 5 96 3 32 5 18 0.50 512 54
64 3 64 3 16 3 16 3 7 0.25 256 49
32 3 32 3 32 5 32 3 9 0.25 128 50

295 JAST Vol. 1 | No. 3 | 2023 |

 Abdulbaset M. Qaid Musleh et.al JAST

1. Initialization The three most accurate chromosomes were

The initial population is a set of selected from the originals of six chromosomes.
randomly generated chromosomes, as listed in Figure 3 shows the selection process.
Table 1. Each line in Table 2 represents one of
the solutions generated by the genetic
algorithm. Six child chromosomes in the initial
population were generated using a genetic
Figure 3: Selection Process.
algorithm. Each chromosome is represented by
4. Crossover
11 genes, eight pairing genes, epoch, dropout,
The process of mating between the parents
and a fully connected layer. Each gene
selected from the previous process to obtain
represents the structure of the convolutional
new children through the access point, as the
layer in the CNN model. For example,
main objective of this process, is to gather and
chromosome32,3,16,3,32,5,64,7,21,0.25,128is a
identify information from two different sources
representation of the parameters for a CNN
and introduce a new product that is different
structure with four convolution layers. The first
from them. It selects half of the values from
layer contains 32 filters with a kernel size of 3 ×
parent I and the rest from parent II to generate a
3, the second layer contains 16 filters with a
new population. The crossover process is
kernel size of 3 × 3, the third layer consists of
illustrated in Figure 4.
32 filters with a kernel size of 5 × 5, the fourth
layer compares 64 filters with a kernel size of 7
× 7, 21 is referred to as the epoch number, too
0.25 means a dropout ratio, and 128 is fully Figure 4: Crossover Process.
connected in the CNN model. 5. Mutation
When generating new children in the previous
2. Fitness Function process, the mutation process involves a sudden
The CNN model was trained and evaluated change in the chromosome by changing one of
using the training and testing data from the the genes randomly to ensure that a different
dataset, and the CNN model accuracy was chromosome is obtained from its predecessor
calculated for each chromosome. The parents in [29]. In the proposed method, randomly
the next step are selected based on the changing genes (epochs and neurons in the
maximum accuracy value. In this process, each flattened layer). The mutation process is
chromosome is input to the proposed CNN illustrated in Figure 5.
model with the dataset for training and testing
the model to evaluate the performance and
calculate the fitness value.
Figure 5: Mutation Process.
3. Selection
Three chromosomes are selected from the six C. The Feature Extraction
current chromosomes to generate a new The biggest challenge in solving the problem of
generation, and the selection process is based on handwritten signature verification is feature
the optimization function in which the addition extraction, which enables the system to
of each chromosome has priority. The differentiate between real and forged signatures
optimization function is a mathematical [20]. The convolutional neural network can
equation that varies from one problem to learn different features in each stage based on
another, and is appropriate for solving the the layer-by-layer advancement and finally
problem. In this method, selection depends on realize related verification and classification
the chromosomes with the highest accuracy. functions [30]. Convolutional Neural Networks

JAST Vol. 1 | No. 3 | 2023 | 296

 JAST Abdulbaset M. Qaid Musleh et.al

are particularly suitable architectures for connections to all activations in the previous
signature verification [31]. The proposed model layer, as observed in regular neural networks.
consists of four convolution layers: the Their activation can be computed with matrix
convolution layer, in which the size of the input multiplication followed by a bias offset. Figure
image, grayscale dimensions (100 × 100 × 1) of 6 shows the CNN architecture of the proposed
the kernel, and number of filters are specified. model.
The filter of a filter weight array of a certain The model CNN proposed steps are:
size, for example, (3 × 3), wraps the signature 1) Create a Sequential model.
image, starting from the upper left corner, 2) Add a Conv2D layer to the model with
passing through a certain number of rows, until a specified number of filters (32), kernel
it reaches a corner at the bottom right of the size (3 × 3), and activation function
image. This process repeats the number of (activation). The input shape was set to
filters entering the layer with a change in filter the specified input shape (100x100x32).
content each time. The equation below 3) Add another Conv2D layer to the
describes the size of the feature map (output model with a specified number of filters
layer) as the sum of [the receptive field array for (16), kernel size (3 × 3), padding, and
each filter position multiplied by the filter activation function (activation).
weight array]: 4) Add a MaxPooling2D layer was added
to the model with a specified pool size
(2 × 2).
𝐹𝑒𝑎𝑡𝑢𝑟𝑒 𝑀𝑎𝑝 = 𝑅𝑒𝑐𝑖𝑝𝑡𝑖𝑣𝑒𝐹𝑖𝑙𝑒𝑑𝑘 ∗ 𝐹𝑖𝑙𝑡𝑒𝑟 (1)
5) Add another Conv2D layer to the
This implies that if the number of filters used in model with a specified number of filters
the layer is 32, the result of the equation will be (32), kernel size (5 × 5), padding, and
100 × 100 × 32 and equal to 320,000. After activation function (activation).
creating the Convolution Layers and applying 6) Add another MaxPooling2D layer was
the activation functions to them as the ReLU added to the model with a specified pool
function, a method called max-pooling, also size (2 × 2).
called downsampling or subsampling, is used. 7) Add another Conv2D layer to the
Reduce the size of the output matrix and obtain model with a specified number of filters
only the values of the greatest importance. (64), kernel size (7 × 7), padding, and
However, a filter is usually assigned with activation function (activation).
dimensions of 2 × 2 and a pass-through step = 2 8) Add another MaxPooling2D layer was
pixels, and it is applied to the feature map, added to the model with a specified pool
where the filter extracts the maximum value size (2 × 2).
within each effective area. The activation 9) Add a dropout layer was added to the
function is a nonlinear function that is model with a specified dropout rate
considered a bridge to the next layer in the (0.25).
convolutional neural network. Its usefulness lies 10) Add a Flatten layer to the model.
in reducing the number of computations 11) Add a dense layer was added to the
performed by not activating all the feature map model with a specified number of units
points at the same time; in other words, (128) and activation function
activating the points that represent the features (activation).
and excluding the points that do not represent 12) Add another Dense layer to the model
them. It has several types, the most famous of with a specified number of classes
which are Tanh [10], ReLU[33], Softmax [34], (number of classes) and activation
and Sigmoid [38], which will be used in this function (activation).
search. The fully connected layer has full

297 JAST Vol. 1 | No. 3 | 2023 |

 Abdulbaset M. Qaid Musleh et.al JAST

13) Compile the model is compiled with the algorithm. This precision was then sent to
specified optimizer (optimizer), loss Genetic Algorithm for its value. The fitness is a
function (loss), and metrics. function that determines the efficiency of the
14) Train the model was trained on the model. If the number of proposed generations
training data (x_train and y_train) with a for the GA is completed, the best accuracy
specified batch size number of epochs obtained is chosen to preserve the weights
(21). produced from these parameters in the CNN.
15) Evaluate the performance for a model Based on the previous steps in our proposed
with data (x_test and y_test). model, we obtained the highest possible
accuracy of 97.7%. The proposed system was
trained and tested with more than one dataset:
(GPDS-160, GPDS-300, CEDAR,
BHSig260Hindi, and BHSig260Bengali). In the
GPDS, the dataset consisted of 24 genuine
signatures and 30 simulated forgeries from 160
and 300 individuals. Fourteen genuine and 20
forgeries signatures from each individual were
used in the training stage, and 10 genuine and
10 forgeries signatures were used in the test
stage [37]. The CEDAR dataset collected 2640
total signatures images for 55 persons. Each
participant had 24 genuine and 24 forged
signatures. Fourteen genuine and 14 forgeries
signatures from each individual were used in the
training stage, and the remaining images were
Figure 6: Block Diagram of CNN Proposed used for testing. The BHSig260-Bengali and
Architecture. BHSig260-Hindi datasets contain 24 genuine
signatures and 30 simulated forgeries, with 100
5. Training and Testing Stage people in BHSig260-Bengali and 160 persons in
At this stage, the genetic algorithm sends the BHSig260-Hindi. Fourteen genuine and 20
generated chromosomes represented by the forgeries signatures from each individual were
parameters to the CNN model. Each used in the training stage, and 10 genuine and
chromosome contains a set of values called 10 forgeries signatures were used in the testing
parameters, where the parameters represent the stage. Table 3 presents the details of the datasets
number of filters for four layers in the CNN used in the proposed system. A genetic
model, the size of the kernel for four layers, the algorithm selects the best hyperparameters for
number of neurons exiting from the layers to the each dataset. The CNN model consists of four
fully connected layer, and the number of cycles layers: the input layer, Conv1, Conv2, Conv3,
for this solution is called EPOCH. These and Conv4. The model includes a set of
parameters are fed to the model with the parameters that are used for signature image
training image array, and then the efficiency of processing and verification of genuine
the model is evaluated using the test dataset to handwriting signatures. These parameters are
obtain the accuracy that represents the specific listed in table 4 according to the different
solution of the parameters sent from the genetic datasets.

JAST Vol. 1 | No. 3 | 2023 | 298

 JAST Abdulbaset M. Qaid Musleh et.al

Table(3):Dataset Used in Proposed System

Type GPDS-160 GPDS-300 CEDAR Bengali Hindi
Signers 160 300 100 100 160
Genuine 24 24 24 24 24
Forged 30 30 24 30 30
Training 5440 10200 1540 3400 5440
Testing 3200 6000 1100 2000 3200

Table(4):The Parameter Setting of The Proposed System.

Parameter GPDS-300 CEDAR BHSig260-B BHSig260- H
Generations 7 2 5 1
Max Epochs 20 21 12 14
Parameters 2,402,731 2,545,911 2,426,120 1,241,250
Layer 1 Conv2D (32, 3x3) (32, 3x3) (64, 3x3) (32, 3x3)
Layer 2 &MaxPool (2, 2) Conv2D (32,3x3)) Conv2D (64,3x3) Conv2D (32, 3x3) Conv2D (32,3x3)
Layer 3 &MaxPool (2, 2) Conv2D (32, 3x3) Conv2D (32,3x3) Conv2D (64, 3x3) Conv2D (32,3x3)
Layer 4 &MaxPool (2, 2) Conv2D (32, 3x3) Conv2D (64,3x3) Conv2D (32, 3x3) Conv2D (32,3x3)
Dropout 0.50 0.25 0.25 0.50
Flatten 512 264 512 256
Accuracy 0.93 0.977 0.958 0.922
FAR. A lower EER indicates better
6. Experimental Results performance for the model). The results
Each dataset was subdivided into two parts: obtained from the proposed method of
training set and testing. The performance of the constructing the CNN model using the genetic
proposed system was measured using three algorithm were compared with those of other
global scales which are as follows: Accuracy is methods using hand-built CNN models. The
the ratio of the number of correctly categorized results were compared with those of other
signatures to the total number of complete studies. Figures (7,8,9, and 10) show the graphs
signatures. These are the False Acceptance Rate of the curves of the results of the proposed
(FAR) and False Rejection Rate (FRR), where study, where the curves are loss, val_loss,
FAR is the presence of the forgeries signature val_accuracy, and accuracy with the used
that are incorrectly classified. An equal Error dataset.
Rate (EER) is applied to evaluate the
equilibrium point where the FRR equals the
Table (5): Comparison Results for CEDAR Dataset.
CEDAR
Method
FAR FRR EER Accuracy%
Surroundedness Features [ [7]] 8.33 8.33 8.33 91.67
Multi-Path Siamese (MA-SCN) [32] 19.21 18.35 18.92 80.75
Siamese CNN [30] 6.78 4.20 – 95.66
Our method 2.5 2.2 2.35 97.73

Table (6): Comparison Results for GPDS-160 and GPDS-300.

GPDS-160- GPDS-300
Method
FAR FRR EER Accuracy%
CNN_GP [35] 9.08 20.60 12.83 92
GoogLeNet Inception-v1 and Inception-v3 [36] – – 26 72
Our method 9.1 20 11 93

299 JAST Vol. 1 | No. 3 | 2023 |

 Abdulbaset M. Qaid Musleh et.al JAST

Table (7): Comparison Results for BHSig260-B and BHSig260-H

BHSig260- B BHSig260- H
Method
FAR FRR EER Accuracy% FAR FRR EER Accuracy%
Multi-Path Siamese 5.73 4.86 8.18 94.99 9.96 5.85 5.32 92
(MA-SCN) [32]
Siamese CNN [30] 14.25 6.41 – 90.64 12.29 9.6 – 88.98
Multi-scripted with 1.50 3.14 – 95 2.31 6.65 – 90
CNN [6]
Our method 1.3 2.1 1.7 95.82 6.8 4.7 5.2 92.26

Table 5 shows our results for the (CEDAR)

dataset using two methods: the Multi-Path
Attention Siamese Convolution Network (MA-
SCN) model [32] and A Two-Stage Siamese
Network Model for Offline Handwritten
Signature Verification [30]. Surroundedness
features study using local binary pattern (LBP)
and uniform Local Binary Patterns (ULBP) [7].

The results in table 6 compare our system with

other systems using the GPDS-160 and GPDS-
300 datasets. The results in Table 7 are Figure 9 :The Resulting BHSig260-B
displayed as a comparison of some of the
study’s results with our results for BHSig260-B
and BHSig260-H datasets.

Figure 10 :The Resulting BHSig260-B

7. Conclusion
Figure 7 :The Resulting CEDAR Dataset
This study focuses on the development of an
offline signature verification model using a
genetic algorithm and convolutional neural
networks (CNNs). The goal was to optimize the
CNN architecture and hyperparameters to
achieve highly efficient models with accurate
results. The proposed method was tested on
multiple datasets including BHSig260-Bengali,
BHSig260-Hindiin, GPDS, and CEDAR.
The results obtained from the approach
presented in this study demonstrate a promising
Figure 8 :The Resulting GPDS-300

JAST Vol. 1 | No. 3 | 2023 | 300

 JAST Abdulbaset M. Qaid Musleh et.al

performance. The model achieved a high and Machine Learning in PRML, July 2022, IEEE,
discrimination rate with a False Rejection Rate pp. 391-397.
[6] T. Longjam, D. R. Kisku, and P. Gupta, "Multi-
(FRR) of 2.5, False Acceptance Rate (FAR) of scripted Writer Independent Off-line Signature
3.2, Equal Error Rate (EER) of 2.35, and an Verification using Convolutional Neural Network,"
accuracy rate of 97.73%. Multimedia Tools and Applications, pp. 1-18, Aug.
The use of deep learning and CNN-based 2022.
signature verification systems has proven to be [7] S. Pal, A. Alaei, U. Pal, and M. Blumenstein,
"Performance of an off-line signature verification
effective, providing a less intrusive and widely method based on texture features on a large
accepted method for verifying signatures. The indicscript signature dataset," in 12th IAPR
genetic algorithm plays a crucial role in Workshop on Document Analysis Systems, pp. 72-
developing a population of CNN models and 77, IEEE, April 2016.
finding the best-fit architecture for the offline [8] A. Foroozandeh, A. Hemmat, and H. Rabbani,
"Offline handwritten signature verification and
signature verification task. recognition based on deep transfer learning," in
International Conference on Machine Vision and
Overall, the findings of this study highlight the Image Processing (MVIP), IEEE, Feb. 2020, pp. 1-
potential of utilizing genetic algorithms and 7.
CNNs in signature verification systems. The [9] N. Sharma, S. Gupta, P. Mehta, X. Cheng, A.
Shankar, P. Singh, and S.R. Nayak, "Offline
optimized model presented in this manuscript signature verification using deep neural network
demonstrates promising results and contributes with application to computer vision," Journal of
to the advancement of automated signature Electronic Imaging, vol. 31, no. 4, pp. 041210-1-
verification techniques. Further research and 041210-10, Jul. 2022.
improvements in this field can lead to enhanced [10] Y. LeCun, Y. Bengio, and G. Hinton, "Deep
learning," Nature, vol. 521, no. 7553, pp. 436-444,
security measures, efficient document handling, May 2015.
and increased user confidence in various [11] A. Krizhevsky, I. Sutskever, and G. E. Hinton,
applications such as financial operations and "Imagenet classification with deep convolutional
computer systems. neural networks," Communications of the ACM,
vol. 60, no. 6, pp. 84-90, 2017.
[12] C. Clark and A. Storkey, "Training deep
8. References convolutional neural networks to play go," in
[1] M. A. Ferrer and C. M. Travieso, "Offline signature International Conference on Machine Learning,
verification: An overview and some recent PMLR, Jun. 2015, pp. 1766-1774.
advances," Pattern Recognition Letters, vol. 34, no. [13] N. Purohit, Sapna Purohit, and C. S. Satsangi,
3, pp. 249-256, 2013. "Offline Handwritten Sig- nature Verification Using
[2] B. M. Al-Maqaleh and A. M. Musleh, "An efficient Template Matching and Clustering Technique,"
offline signature verification system using local International Journal of Computer Science and
features," International Journal of Computer Mobile Computing, vol. 2, no. 4, pp. 295-301, 2014.
Applications, vol. 131, no. 10, pp. 39-44, 2015. [14] K. Simonyan and A. Zisserman, "Very deep
[3] L. G. Hafemann, R. Sabourin, and L. S. Oliveira, convolutional networks for large-scale image
"Offline handwritten signature verification— recognition," arXiv preprint, pp.1409-1556, 2014.
literature review," in Seventh International [15] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S.
Conference on Image Processing Theory, Tools and Reed, D. Anguelov, and A. Rabinovich, "Going
Applications (IPTA), Nov. 1-8, 2017, IEEE. deeper with convolutions," in Proceedings of the
[4] R. Verma and D. Rao, "Offline signature IEEE Conference on Computer Vision and Pattern
verification and identification using angle feature Recognition, 2015, pp. 1-9.
and pixel density feature and both methods [16] K. He, X. Zhang, S. Ren, and J. Sun, "Deep
together," International Journal of Soft Computing residual learning for image recognition," in
and Engineering, vol. 2, no. 4, pp. 740-746, 2013. Proceedings of the IEEE Conference on Computer
[5] Y. Muhtar, W. Kang, A. Rexit, and K. Ubul, "A Vision and Pattern Recognition, 2016, pp. 770-778.
Survey of Offline Handwritten Signature [17] E. Parcham, M. Ilbeygi, and M. Amini,
Verification Based on Deep Learning," in 2022 3rd "CBCapsNet: A novel writer- independent offline
International Conference on Pattern Recognition signature verification model using a CNN-based

301 JAST Vol. 1 | No. 3 | 2023 |

 Abdulbaset M. Qaid Musleh et.al JAST

architecture and capsule neural networks," Expert and Industrial Engineering, vol. 30, no. 4, pp. 1037-
Systems with Applications, p. 115649, Dec. 2021. 1045, 1996.
[18] G. Huang, Z. Liu, K. Q. Weinberger, and L. van [28] D. P. Sudharshan and R. N. Vismaya,
der Maaten, "Densely connected convolutional "Handwritten Signature Verification System using
networks," in Proceedings of 2017 IEEE Deep Learning," in 2022 IEEE International
Conference on Computer Vision and Pattern Conference on Data Science and Information
Recognition, Honolulu, HI, USA, 2017, pp. 2261- System (ICDSIS), IEEE, July 2022, pp. 1-5.
2269. [29] A. S. Mondal, "Evolution of convolution neural
[19] F. Johnson, A. Valderrama, C. Valle, B. network architectures using genetic algorithm," in
Crawford, R. Soto, and R. Ñanculef, "Automating 2020 IEEE Congress on Evolutionary Computation
configuration of convolutional neural network (CEC), IEEE, July 2020, pp. 1-8.
hyperparameters using genetic algorithm," IEEE [30] W. Xiao and Y. Ding, "A Two-Stage Siamese
Access, vol. 8, pp. 156139- 156152, 2020. Network Model for Offline Handwritten Signature
[20] V. Malekian, A. Aghaei, M. Rezaeian, and M. Verification," Symmetry, vol. 14, no. 6, pp. 1216,
Alian, "Rapid off-line signature verification based 2022.
on signature envelope and adaptive density [31] L. G. Hafemann, R. Sabourin, and L. S. Oliveira,
partitioning," in 2013 First Iranian Conference on "Learning features for offline handwritten signature
Pattern Recognition and Image Analysis (PRIA), verification using deep convolutional neural
IEEE, Mar. 2013, pp. 1-6. networks," Pattern Recognition, vol. 70, pp. 163-
[21] J. Donahue, L. Anne, S. Guadarrama, M. 176, 2017.
Venugopalan, S. Rohrbach, K. Saenko, and T. [32] X. Zhang, Z. Wu, L. Xie, Y. Li, F. Li, and J.
Darrell, "Long-term recurrent convolutional Zhang, "Multi-Path Siamese Convolution Network
networks for visual recognition and description," in for Offline Handwritten Signature Verification," in
Proceedings of the IEEE Conference on Computer 2022 The 8th International Conference on
Vision and Pattern Recognition, 2015, pp. 2625- Computing and Data Engineering, Jan. 2022, pp.
2634. 51-58.
[22] S. Dey, A. Dutta, J. I. Toledo, S. K. Ghosh, J. [33] V. Nair and G. E. Hinton, "Rectified linear units
Lladós, and U. Pal, "Signet: Convolutional siamese improve restricted boltzmann machines," in
network for writer independent offline signature Proceedings of the 27th International Conference on
verification," arXiv preprint arXiv:1707.02131, Machine Learning (ICML-10), 2010, pp. 807-814.
2017. [34] I. Goodfellow, Y. Bengio, and A. Courville,
[23] C. Yinka-Banjo and C. Okoli, "Signature Adaptive Computation and Machine Learning
Verification Using Siamese Convolutional Neural Series, 2016.
Networks," Covenant Journal of Informatics and [35] L. G. Hafemann, R. Sabourin, and L. S. Oliveira,
Communication Technology, 2019. "Writer-independent feature learning for offline
[24] E. G. Arhore, M. Yasaee, and I. Dayyani, signature verification using deep convolutional
"Optimisation of convolutional neural network neural networks," in 2016 International Joint
architecture using genetic algorithm for the Conference on Neural Net- works (IJCNN), IEEE,
prediction of adhesively bonded joint strength," July 2016, pp. 2576-2583.
Structural and Multidisciplinary Optimization, vol. [36] S. M. Sam, K. Kamardin, N. N. A. Sjarif, and N.
65, no. 9, pp. 1-16, 2022. Mohamed, "Offline signature verification using
[25] C. Li, J. Jiang, Y. Zhao, R. Li, E. Wang, X. deep learning convolutional neural network (CNN)
Zhang, and K. Zhao, "Genetic Algorithm based architectures GoogLeNet inception-v1 and
hyper-parameters optimization for transfer inception-v3," Proce- dia Computer Science, vol.
Convolutional Neural Network," in International 161, pp. 475-483, 2019.
Conference on Advanced Algorithms and Neural [37] M. A. Ferrer, J. F. Vargas, A. Morales, and A.
Networks (AANN 2022), vol. 12285, SPIE, June Ordonez, "Robustness of offline signature
2022, pp. 232-241. verification based on gray level features," IEEE
[26] O. M. Assim and A. M. Alkababji, "CNN and Transactions on Information Forensics and Security,
Genetic Algorithm for Finger Vein Recognition," in vol. 7, no. 3, pp. 966-977, 2012.
2021 14th International Conference on [38] Venkata Saichand, Nirmala Devi M, Arumugam
Developments in Systems Engineering (DeSE), S, and N. Mohankumar, "FPGA Realization of
IEEE, Dec. 2021, pp. 503- 508. Activation Function for Artificial Neural Net-
[27] Xuhua Yang, Xianyi Zeng, Hongbo Fu, and works," in IEEE Eighth International Conference on
Yaping Zhang, "Selection of features for signature Intelligent Systems Design and Applications, 2008.
verification using the genetic algorithm," Computers [39] A. K. Jain, A. Ross, and K. Nandakumar,
Introduction to Biometrics, Springer, 2016.

JAST Vol. 1 | No. 3 | 2023 | 302

 JAST Abdulbaset M. Qaid Musleh et.al

[40] D. Impedovo and G. Pirlo, "Automatic signature

verification: the state of the art," IEEE Transactions
on Systems, Man, and Cybernetics, Part C
(Applications and Reviews), vol. 38, no. 5, pp. 609-
635, 2008.
[41] C. L. Liu and Y. H. Yin, "Offline handwritten
signature verification - literature review," Pattern
Recognition, vol. 40, no. 8, pp. 2293-2307, 2007.
[42] S. Panigrahi, A. K. Tiwari, and R. Sharma, "A
study of various off- line signature verification
approaches," International Journal of Computer
Science and Network Security (IJCSNS), vol. 12,
no. 10, pp. 78-83, 2012.
[43] L. V. Batista, D. Rivard, R. Sabourin, and P.
Maupin, "State of the art in off-line signature
verification," in Iberoamerican Congress on Pat-
tern Recognition, Springer, Berlin, Heidelberg,
2009, pp. 227-234. DOI: 10.1007/978-3-642-10268-
4_28.
[44] V. E. Ramesh and M. Narasimha, "Off-line
signature verification using genetically optimized
weighted features," Pattern Recognition, vol. 32, no.
2, pp. 217-233, 1999.

303 JAST Vol. 1 | No. 3 | 2023 |