International Journal of Computing and Digital Systems

ISSN (2210-142X)
Int. J. Com. Dig. Sys. 16, No.1 (Jul-24)
https://dx.doi.org/10.12785/ijcds/160121

Authentic Signature Verification Using Deep Learning

Embedding With Triplet Loss Optimization And Machine
Learning Classification
Andreas Christianto1 , Jovito Colin1 and I Gede Putra Kusuma Negara1
1
Computer Science Department, BINUS Graduate Program - Master of Computer Science, Bina Nusantara University, Jakarta,
Indonesia, 11480

Received 9 Feb. 2024, Revised 9 Apr. 2024, Accepted 16 Apr. 2024, Published 1 Jul. 2024

Abstract: Various document types (financial, commercial, judicial) necessitate signatures for authentication. With the advancements
of technology and the increasing number of documents, traditional signature verification methods encounter challenges in facing
tasks related to verifying images, such as signature verification. This idea is further reinforced by the growing migration of
transactions to digital platforms. To that end, the fields of Machine learning (ML) and Deep Learning (DL) offer promising
solutions. This study combines Convolutional Neural Network (CNN) algorithms, such as Visual Geometry Group (VGG) and
Residual Network (ResNet) or VGG16 and ResNet-50 specifically, for image embedding alongside ML classifiers such as Support
Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest, and Extreme Gradient Boosting (XGBoost). While
the aforementioned solutions are usually enough, real life scenarios tend to differ in environment and conditions. This problem
leads to difficulty and accidents in the verification process, causing the users to redo the process or even end it prematurely.
To alleviate the issue, this study employs optimization methods such as hyperparameter tuning via Grid Search and triplet loss
optimization to enhance model performance. By leveraging the strengths of CNNs, Machine Learning classifiers, and optimization
techniques, this research aims to improve the accuracy and efficiency of signature verification processes while addressing real-world
challenges and ensuring the trustworthiness of electronic transactions and legal documents. Evaluation is conducted using the
ICDAR-2011 and BHSig-260 datasets. Results indicate that triplet loss optimization significantly improves the performance of
the VGG16 embedding model for SVM classification, notably elevating the Area Under the ROC Curve (AUC) from 0.970 to 0.991.

Keywords: Signature Verification, Signature Authentication, Image Embedding, Triplet Loss, Machine Learning Classifiers

1. Introduction demands attention [2]. For the financial aspect of this

In today’s digital world, the need for strong and issue, financial transactions increasingly migrate to digital
secure handwritten signature verification methods is more platforms, the banking industry, in particular, faces the
important than ever. Since every person has a distinctive challenge of securing electronic signatures against sophis-
signature that is mostly used for personal identification ticated forgery attempts. The implementation and applica-
and the authentication of significant papers or legal trans- tion of robust signature verification systems is imperative
actions [1]. As online electronic transactions and digital to safeguard the integrity of financial transactions, prevent
documents become more prevalent, there’s a growing de- identity theft, and ensure the trustworthiness of electronic
mand for advanced ways to verify signatures. Traditional documents. Furthermore, the real-world applications of
methods face challenges in accurately confirming signa- this research extend beyond the banking sector. Govern-
ture authenticity, leading to a shift towards utilizing the ment agencies, legal institutions, and various industries
latest technologies as a solution. The integration of deep dealing with sensitive information can benefit from an
learning and machine learning techniques into signature advanced signature verification system. The proposed
verification processes presents a promising solution to methodology aims to offer a versatile solution applicable
address the limitations of conventional approaches. Lever- across diverse domains where signature verification is a
aging the capabilities of computer vision in conjunction critical component of security protocols.
with these advanced technologies holds the potential to
drastically improve the accuracy, adaptability, and overall Recent advancements in deep learning have demon-
efficiency of signature verification. strated significant breakthroughs in various computer
vision tasks. Cutting-edge models, such as the VGG-
Every day throughout the world, a great number of 16 architecture, showcase remarkable advancements for
vital financial, commercial, and legal papers are signed, visual recognition tasks. VGG-16 is a robust model that
so validating signatures has become a critical issue that consists of 16 convolution layers and is fully connected,

E-mail address: [email protected], [email protected], [email protected] https:// journal.uob.edu.bh

266 Andreas Christianto, et al: Authentic Signature Verification. . .

which is usually used to recognize and classify images explores some research papers, each contributing distinct
[3]. Not to mention, it has been pre-trained using the methodologies and insights to the overarching theme
ImageNet dataset, which makes the model good for im- of authentic signature verification through deep learning
age classification tasks. With that in mind, incorporating algorithms.
VGG-16 into authentic signature verification tasks may
potentially enhance the accuracy and efficiency of the Engin et al [6] addresses the challenge of offline sig-
verification process. nature verification in real-world scenarios, particularly fo-
cusing on a banking context where customers’ transaction
Furthermore, notable research efforts by Manish Baj- request documents with occluded signatures are compared
pai [4] and Xamxidin et al [5] have contributed to the ex- to their clean reference signatures. Unlike controlled
ploration of signature verification methodologies. Bajpai’s datasets used in previous research, real-world signatures
research focuses on leveraging the VGGNet model for can include various occlusions such as stamps, seals,
feature extraction in handwritten signature authentication. ruling lines, and signature boxes, leading to high intra-
On the other hand, Xamxidin et al. propose an Improved class variations. The proposed methodology comprises
Inverse Discriminative Network (IDN) to enhance sig- two main components, a stamp cleaning method based
nature verification accuracy. Leveraging the strengths of on CycleGAN and a signature representation method
these methodologies presents an opportunity for further based on Convolutional Neural Networks (CNNs). The
advancement in signature verification tasks. Specifically, experiment results indicate a 76.8% accuracy when em-
utilizing methods like triplet loss optimization reduces ploying VGG-16 for signature representation along with
input requirements, while employing deeper models like the CycleGAN-based cleaning method, which is better
VGG-16 could extract more meaningful information from compared to the 75
input images.
Poddar et al [7] introduces a novel method for signa-
Therefore, to improve checking and verifying sig- ture recognition and forgery detection while considering
natures, this study uses various methods to thoroughly the challenges associated with signature verification due
assess how the combination of image embedding meth- to the variability introduced by individual writing styles
ods, such as VGG-16 and ResNet-50, triplet loss opti- and environmental factors. The proposed approach em-
mization, and different deep learning embedding models ploys Convolutional Neural Networks (CNNs) and the
alongside machine learning classifiers perform on the task Crest-Trough algorithm for signature verification while
of signature verification, such as Support vector machine employing the Harris Algorithm and the Surf Algorithm
(SVM), Random Forest (RF), Artificial Neural Network for forgery detection. In the experimental results, the
(ANN), and XGBoost (XGB). These methods are selected proposed signature recognition system achieves a high
precisely to give an overview of the performance of accuracy of 94% and an accuracy range of 85-89
the combination of the deep learning embedding models
and machine learning classifiers. The main workflow / Lu et al [8] introduces a methodology for handwriting
architecture of this work can be seen in Figure 1. identification that integrates both dynamic and static fea-
tures to enhance the accuracy of signature identification,
In Section 2, we review relevant literature and pre- particularly in the context of forged signatures. The
vious works related to signature verification methodolo- study establishes a Chinese signature forged handwriting
gies, discussing their strengths and limitations. Section database, which contains 44 signatures from different
3 presents the theoretical background and methodologies signers, collected from a dot matrix digital pen tool.
employed in our research, including deep learning em- The data collection involves offline images and online
bedding models, machine learning classifiers, and opti- data, capturing information such as X and Y coordinate
mization techniques such as triplet loss. In Section 4, we points, pressure, timestamp, and pen up-down marks.
propose our novel methodology, which integrates various For classification, the study employs machine learning
image embedding methods, optimization techniques, and algorithms, such as Support Vector Machine (SVM), and
machine learning classifiers to enhance signature verifi- deep learning algorithms, such as Convolutional Neural
cation accuracy. Section 5 details the experimental setup, Network (CNN). The best results for the proposed method
including dataset selection, model training parameters, are 92.2% and 94.4% for SVM and CNN respectively.
and evaluation metrics. In Section 6, we present the
results of our experiments and provide a comprehensive A research paper by Manish Bajpai [4] focused on
discussion of the findings, including insights into the enhancing the accuracy of handwritten signature authen-
performance of different methods and techniques. Finally, tication using the VGGNet model, specifically for feature
Section 7 concludes the paper by summarizing the key extraction, along with hyperparameter tuning. The final
findings, highlighting contributions, and outlining poten- experiment, which produced the best result, reached a
tial avenues for future research in signature verification. testing accuracy of 95% for detecting genuine and forged
handwritten signatures. The conclusion from the experi-
2. Related Works ment’s result emphasizes the critical role of hyperparam-
The process of handwritten signature verification has eter optimization and highlights that a faster learning rate
gone through drastic changes with the integration of does not necessarily enhance efficiency.
deep learning algorithms. In the pursuit of enhancing
accuracy and reliability, researchers have delved into the Borse et al [9] aims to implement a handwritten
world of machine learning, particularly the utilization of signature verification model using machine learning and
deep learning techniques. This literature review section deep learning to distinguish between genuine and forged
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 267

Figure 1. Overall System Architecture

signatures in order to ensure normalizing signature images FC-ResNet, optimizes the ResNet18 structure for size
for accurate comparisons and improving the accuracy of while introducing CBAM in the residual block to better
detecting correct signatures by utilizing deep learning learn correlations between different feature channels and
models and multiple machine learning models such as spatial positions. The study reports an accuracy rate
Multilayer Perceptron (MLP), Support Vector Machine of 96.21% on the CEDAR dataset and 96.41% on the
(SVM), and Random Forest (RF) models. The MLP Uyghur language dataset, demonstrating the method’s
model outperformed SVM and RF models, achieving an effectiveness for signature data with few samples and
accuracy of 95.4%. The MLP model’s effectiveness and its ability to accurately identify signature samples across
robustness suggest its potential in various applications, languages.
particularly in government offices where handwritten sig-
natures play a crucial role in approval and authentication In addition, Lopes et al [11] conducts signature ver-
processes. ification using a modified version of the AlexNet deep
learning model implemented through TensorFlow. This
Melhaoui and Benchaou [10] fixates on the chal- model is designed to recognize and verify individual
lenges and methodologies associated with offline signa- signatures, marking potential forgeries for further manual
ture recognition systems due to its complexity compared verification. In the binary classification model test, mul-
to the online mode. The study utilizes a proprietary signa- tilayer perceptron (MLP) serves as a binary classifier for
ture database, containing 240 signatures from 12 individ- signature/non-signature test data, achieving an accuracy
uals. Feature extraction methods, including Histogram of of 98.4% and an F1 score of 98.3% on test data. However,
Oriented Gradients (HOG), Profile Projection (PP), and this model only confirms the presence or absence of a
Loci, are discussed. Classification methods, specifically signature without verifying its authenticity. The paper
FMMC and K Nearest Neighbors (KNN), are employed concludes by underlining the importance of handwritten
for evaluating the recognition system performance. The signatures in attendance verification. It suggests that the
proposed system, through combining HOG features and proposed methods, including the MLP classifier and CNN
FMMC classification, achieves a recognition rate of 96 model, offer a reliable solution for automating signature
verification.
Xamxidin et al [5], conducted the study on a different
approach and a much broader multilingual dataset to Aljrami et al [12] address the critical task of sig-
test its effectiveness in signature diversity. The paper nature verification and forgery detection, distinguishing
introduces an Improved Inverse Discriminative Network between static (offline) and dynamic (online) methods.
(IDN) to enhance the identification of genuine and false The significance of handwritten signatures as a widely
signatures to address the challenges in said task, such accepted personal attribute for identity verification in
as sparse signature information, language diversity, and various sectors is highlighted. The paper introduces the
the arbitrary nature of signature styles. The experiments use of Deep Convolutional Neural Networks (CNNs)
conducted on this research paper involves testing the for both writer-independent feature learning and writer-
proposed method on various datasets, including Chinese, dependent classification. The proposed methodology con-
Uyghur, Bengali, and Hindi signatures. The conclusion siders handwritten signatures as behavioral biometrics,
that is reached based on the results of the experiments acknowledging the changes in an individual’s signature
emphasizes that the proposed IDN model effectively im- over time. The authors demonstrate the effectiveness of
proves the accuracy of signature verification for single and their approach by showcasing accuracy and loss plots for
mixed languages with an ACC of 92.40% using a Chinese different dataset split ratios. The model achieves its high-
dataset, an ACC of 92.96% using a Uyghur dataset, and an est accuracy rate of 99.7% on the validation dataset with
ACC of 96.33% using a dataset mixed with both Chinese an 8:2 dataset split, suggesting its efficacy in signature
and Uyghur languages. verification.
A paper by Muhtar et al [2] centers on the critical In summary, the exploration of various methodolo-
task of handwritten signature verification, emphasizing gies in handwritten signature verification through deep
its significance in authenticating crucial financial, com- learning algorithms has revealed significant advancements
mercial, and judicial documents globally. The experi- in the field. The discussed papers contribute diverse
mental methodology employs the ResNet18 network and approaches, each addressing specific challenges and in-
introduces the Convolutional Block Attention Module troducing innovative techniques. Engin et al. emphasize
(CBAM) to improve the model. The proposed method, the complexity of real-world scenarios, achieving a com-

https:// journal.uob.edu.bh
268 Andreas Christianto, et al: Authentic Signature Verification. . .

mendable accuracy of 76.8% by combining CycleGAN training process and to overcome the degradation problem
and VGG-16. Poddar et al. present a novel method achiev- in learning. The residual learning framework, known as
ing 94% accuracy for signature recognition and 85-89% ResNet, presented a novel framework for residual learning
for forgery detection. Lu et al.’s integration of dynamic [15].
and static features yields promising results of 92.2%
and 94.4% for SVM and CNN, respectively. Notably, ResNet architectures are widely known for their use
Aljrami et al. showcase the highest accuracy of 99.7%, of residual blocks, which include skip connections to
emphasizing the significance of dynamic methods and facilitate the flow of information through the network.
considering signatures as behavioral biometrics. While One of the popular variants of the ResNet architecture
each method demonstrates strengths, the choice of the is ResNet-50. This particular version of the architecture
most effective approach may depend on specific use cases consists of 50 layers. A residual block consists of three
and dataset characteristics. Overall, this review highlights layers, the first layer of the block reduces dimensionality
the evolving trends in deep learning-based signature with a 1x1 kernel, the second employs a 3x3 kernel, while
verification, with a continued emphasis on addressing the third layer restores the original dimensionality. This
real-world challenges and improving accuracy in various design alleviates the vanishing gradient problem which
contexts. enables the training of exceptionally deep networks.

3. Theories and Methods Following the residual blocks, Global Average Pool-
A. VGG16 ing (GAP) is employed. This produces a compact 1x1
In the realm of computer vision, around mid-2010s, feature map. Afterwards, an FC layer with 1000 nodes
some convolutional networks were able to achieve high is employed. The final layer utilizes a softmax activation
accuracy on ILSVRC classification and localisation tasks. function, converting the raw output scores into a proba-
Not only that, they are also tested against other image bility distribution over the classes.
recognition datasets and were able to achieve excellent In summary, ResNet-50 serves as a testament to the
performance [13]. These convolutional networks were benefits of residual learning, facilitating the training of
later known as VGGNet, a family of convolutional net- exceptionally deep neural networks while maintaining
works. VGGNet introduced a novel architecture character- interpretability. In this case study, the signature images
ized by its depth and capability to extract intricate features act as inputs to the network and ResNet-50 obtains the
from images. intricate patterns and features from the images.
One of the members of VGGNet, VGG16, stands out C. Triplet Loss
as an excellent candidate for image recognition tasks. In
the training process, VGG16 receives fixed-size 224x224 To further improve the embedding models, an addi-
RGB images. The architecture consists of a stack of tional method, called triplet loss, was introduced. Origi-
convolutional layers while primarily utilizing 3x3 filters nally, this method was made to improve face recognition
for capturing directional information. Aside from that, tasks by enforcing a margin between each pair of faces
1x1 convolutional filters are employed in one of the from one person to all other faces. This allows the faces of
configurations as linear transformations of input channels. one identity to stay near each other, while still enforcing
The fixed convolutional stride of 1 pixel and spatial the distance to other identities [16]. The steps to the
padding ensure that the spatial resolution is preserved af- triplet loss method are triplet selection, obtaining the
ter convolution. Through five max pooling layers, spatial embeddings through convolutional neural networks, and
pooling is introduced, each operating over a 2x2 pixel utilizing the loss function.
window with a stride of 2. In this case study, triplet loss is utilized in order to
Following the convolutional layers, VGG16 employs enhance the discernment between genuine and forged sig-
three Fully-Connected (FC) layers with the channels of natures. This metric optimizes the arrangement of embed-
4096, 4096, and 1000 respectively. The final layer, which dings in the feature space. In the context of handwritten
exists after the FC layers, employs a soft-max activation signature verification, triplet loss ensures that the distance
to convert the raw output into probability between the between embeddings of genuine signatures is minimized,
values of 0 and 1. It is also important to note that all while the distance between genuine and forged signature
hidden layers incorporate the rectification non-linearity embeddings is maximized. This addition may potentially
(ReLu) which enhances the network’s capability to learn improve the model’s ability to create compact clusters for
intricate features. genuine signatures and increase separation from forged
signatures.
To summarize, the VGGNet algorithm is capable of
D. Signature Verification Methods
extracting features well [14], hence could be beneficial for
image embedding in handwritten signature verification. In In this section, we introduce the machine learning
this case study, the signature images act as inputs to the models proposed as classifiers for the task of handwritten
network, and the deep layers of VGG-16 capture intricate signature verification using advanced deep learning tech-
patterns and features inherent in genuine signatures. niques. The primary objective is to establish robust and
accurate verification methods capable of distinguishing
B. ResNet-50 between genuine and forged signatures. Each classifier
Due to the difficulty of training deep neural networks. utilized deep learning architectures such as VGG-16 and
A residual learning framework was made to ease the
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 269

ResNet-50, and triplet loss for enhanced feature extrac- is computed based on weighted sums and activation
tion. The following subsections delve deeper into each functions. The calculated output is then compared to
proposed classifier, illustrating their beneficial attributes the desired output, and the network’s error is computed.
and contributions to the overall signature verification Backward propagation entails adjusting the weights in
framework. the direction opposite to the error gradient, a process
repeated iteratively until the network converges to a state
1) Support Vector Machine (SVM): of accurate predictions [22].
Support Vector Machine (SVM) is a machine learn-
ing algorithm that has shown a good learning ability The characteristics and features of ANN are shown
and generalization ability in classification, regression and to be a fitting algorithm for handwritten signature veri-
forecasting [17]. This algorithm operates by finding an fication. Beginning with how handwritten signatures ex-
optimal hyperplane in a high-dimensional space that hibit intricate and unique patterns, and ANNs, with their
effectively separates data points belonging to different ability to capture complex non-linear relationships, are
classes. In the field of handwritten signature verification, well-suited for recognizing and learning these patterns
SVM could play a pivotal role in distinguishing between [23]. Furthermore, ANN has proven its adaptability to
genuine and forged signatures. The algorithm could be high-dimensional feature spaces. In handwritten signature
integrated with advanced techniques such as VGG-16 verification, features extracted from signature images,
image embedding and triplet loss to enhance the accuracy such as those obtained from deep learning architectures
and reliability of the signature verification process. like VGG-16, often result in high-dimensional spaces.
ANNs can effectively operate in these spaces, providing a
SVM is particularly fitting for handwritten signature suitable framework for processing and learning from the
verification due to a couple noticeable attributes. First of intricate details present in signature images [13]. Lastly,
all, the basic idea of SVM is to translate the input vector the diversity in signature styles demands a model that can
into a high-dimensional space by nonlinear transforma- dynamically adapt to different characteristics. ANNs, by
tion, and then create the best classification surface in said learning and adjusting weights during training, can adapt
space [18], causing it to be effective in high-dimensional to various signature styles, enhancing the model’s ability
spaces and suited for situations such as where signature to accurately verify signatures across a range of writing
data is transformed into complex feature vectors using styles [24].
methods like VGG-16 and ResNet-50. Moreover, SVM
is resilient to overfitting, which is essential for dealing Artificial Neural Networks, as a fitting algorithm for
with limited training data in signature verification tasks. handwritten signature verification, bring essential capa-
Furthermore, its ability to handle non-linear decision bilities such as handling complex patterns, adaptability
boundaries through kernel functions provides flexibility to high-dimensional feature spaces, learning from limited
in capturing intricate relationships within signature data. data, and dynamic adaptation to diverse signature styles.
In closing, Vapnik and Cortes first proposed the support These inherent characteristics make ANNs a powerful
vector machine (SVM) for binary classification in 1995 tool for accurately distinguishing between genuine and
[19]. Since signature verification is inherently a binary forged signatures in various applications, providing a
classification problem which is distinguishing between robust and versatile solution for signature verification
genuine and forged signatures, SVM’s natural binary tasks.
nature aligns well with the task at hand.
3) Random Forest:
In conclusion, the integration of SVM with triplet Random Forest is an ensemble learning algorithm
loss optimized VGG-16 and ResNet-50 image embedding that excels in both classification and regression tasks. In
for handwritten signature verification capitalizes on the fact, they are considered to be one of the most accurate
strengths of each component, resulting in a robust and general-purpose learning techniques available [25]. The
accurate system for authenticating signatures across var- algorithm constructs multiple decision trees during the
ious applications and industries. training phase, with each tree trained on a random sub-
set of the training data and features. This introduction
2) Artificial Neural Network (ANN): of diversity and randomness is crucial in preventing
Artificial Neural Network (ANN) is an algorithm overfitting, contributing to the model’s robustness and
that resembles or mimics the biological human brain adaptability. In classification tasks, the final prediction is
functions to accomplish a given task [20]. Comprising determined by aggregating the predictions of individual
interconnected nodes or artificial neurons organized into trees through a voting mechanism, while in regression
layers, ANNs are designed to learn and recognize com- tasks, the predictions are averaged.
plex patterns within data. These networks consist of input,
hidden, and output layers, with weighted connections The construction of a Random Forest involves the
between neurons. During training, the network adjusts creation of numerous decision trees, each independently
these weights to optimize its ability to capture intricate trained on a subset of the training data and features. The
relationships, enabling ANNs to excel in tasks like pattern process, known as bagging (Bootstrap Aggregating), en-
recognition, classification, and decision-making [21]. sures that each tree is unique, contributing its predictions
to the overall ensemble. . In order to grow these ensem-
The operation of an ANN involves a feedforward and bles, often random vectors are generated that govern the
backward propagation process. During feedforward, input growth of each tree in the ensemble [26]. The ensemble
data is processed through the network, and the output approach enhances the model’s predictive accuracy and
https:// journal.uob.edu.bh
270 Andreas Christianto, et al: Authentic Signature Verification. . .

generalization to new data, making Random Forest a ification, feature extraction methods often result in high-
powerful tool for various machine learning problems. dimensional spaces. XGBoost is well-suited to operate
in these spaces, efficiently handling the intricate and
In the context of handwritten signature verification, complex features extracted from signature images [31].
where signature images are often transformed into high- In addition, XGBoost’s optimized gradient boosting al-
dimensional feature spaces, Random Forest has the po- gorithm enhances accuracy by iteratively improving the
tential and beneficial characteristics to be effective. Tech- model’s predictive performance. This iterative nature is
niques like VGG-16 and ResNet-50 image extraction particularly beneficial when learning and capturing the
generate complex feature representations, and Random subtle nuances present in handwritten signatures [28].
Forest’s ability to handle such intricacies is essential in Since handwritten signature data may contain noise and
distinguishing between genuine and forged signatures. variations. XGBoost’s resilience to noisy data ensures that
the model can discern genuine signatures from forged
The ensemble learning nature of Random Forest is ones, even in the presence of irregularities [32]. Finally,
particularly beneficial in signature verification. By com- XGBoost’s adaptability to diverse data types allows it to
bining predictions from multiple decision trees, the model effectively handle the varying signature styles encoun-
captures the nuanced patterns and variations present in tered in handwritten signature verification tasks, making
different signatures, leading to enhanced accuracy in the it a fitting choice for applications where different writing
verification process. Random Forest is known for its ver- styles must be accommodated [33].
satility and this approach has proved its high accuracy and
superiority with imbalanced datasets [27]. Since hand- XGBoost emerges as a fitting algorithm for handwrit-
written signatures exhibit diverse styles, Random Forest’s ten signature verification due to its ability to handle high-
ability to handle this variability makes it adaptable to dimensional feature spaces, optimized gradient boosting
the inherent complexities of signature verification. The for improved accuracy, robustness to noisy data, and
ensemble of trees allows the model to learn and general- adaptability to diverse signature styles. The algorithm’s
ize across different signature characteristics, providing a efficiency, speed, and customizability contribute to its
versatile solution. suitability for real-world applications, making it a valu-
able tool for accurately distinguishing between genuine
Random Forest emerges as a fitting algorithm for and forged signatures.
handwritten signature verification due to its robustness,
accuracy, and adaptability to high-dimensional feature 4. PROPOSED METHODS
spaces. The ensemble learning approach, combined with A. Dataset
the algorithm’s ability to handle diverse and complex
The dataset used in this case study consists of two sig-
data, positions Random Forest as a reliable solution for
nature datasets taken from kaggle. They are the ICDAR-
distinguishing between genuine and forged signatures in
2011 signature dataset and BHSig260-Bengali signature
various applications.
dataset. The dataset contains 64 and 100 individuals
4) Extreme Gradient Boosting (XGBoost): respectively. Each of these individual’s signatures produce
In closing, this study also includes XGBoost, short multiple images which are divided into two groupings
for Extreme Gradient Boosting, is a powerful machine which are genuine signatures and forged signatures. An
learning algorithm that belongs to the family of gradi- individual in the ICDAR-2011 dataset contains an average
ent boosting methods. Developed to address limitations of 14 genuine signatures and 12 forged signatures while
of traditional gradient boosting techniques, XGBoost is an individual in the BHSig260-Bengali dataset contains
renowned for its efficiency, speed, and high predictive an average of 24 genuine signatures and 30 forged sig-
accuracy. It leverages an ensemble of weak learners, natures. The ICDAR-2011 dataset was chosen because it
typically decision trees, to iteratively optimize a cost func- was used in a competition while the BHSig260-Bengali
tion, enhancing its ability to model complex relationships dataset was chosen to compare the proposed method’s
within data [28]. capability against a larger dataset. The samples of these
datasets can be seen in Figure 2.
XGBoost combines the strengths of boosting algo-
rithms and regularization techniques to improve model B. Preprocessing
performance. During training, weak learners are added In this experiment, the images are resized to a uniform
sequentially, each correcting the errors of its predecessors. size so that they can fit inside the embedding models.
The use of decision trees as base learners, coupled with The size of the images, which were in the range of 200
regularization terms in the objective function, prevents by 200 to 400 by 400, are shrunken down to 64 by 64.
overfitting and enhances the model’s generalization ability In addition, The dataset is split to a ratio of 60/20/20
[29]. Additionally, XGBoost incorporates features like with stratification for training, validation, and testing data
parallel processing, handling missing values, and incor- respectively. This split is chosen to balance between
porating user-defined loss functions, making it a versatile having enough data for training the models effectively
and customizable algorithm [30]. while also ensuring robust evaluation and validation of the
models’ performance. The largest portion of the dataset
XGBoost possesses certain characteristics and fea- is allocated for training the models. A majority of the
tures that make it susceptible to handwritten signature data is needed for training to ensure that the models can
verification. Beginning with how XGBoost handles high- learn meaningful patterns and representations from the
dimensional feature space. In handwritten signature ver- data. With more data for training, the models have a

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 271

Figure 2. ICDAR-2011 and BHSig260-Bengali Dataset Visualization

TABLE I. Evolution of test-bed clusters

C. Feature Extraction for Image Embedding
Dataset Train Validation Test This study utilizes the strength of Convolutional Neu-
ral Network, specifically the VGG-16 and ResNet-50
ICDAR 857 images 391 images 41 images models, to extract features from signature images for veri-
BENGALI 3200 images 1100 images 1100 images fication purposes. These models take the signature images
dataset as inputs and the pre-trained VGG-16 and ResNet-
50 models will then obtain the necessary information
better chance of generalizing well to unseen examples such as simple edges and textures or even patterns and
and avoiding overfitting. A 60% allocation provides a structures. This information can be obtained by finding
substantial amount of data for training while leaving the Red, Green, and Blue (RGB) values associated with
enough for validation and testing. A smaller portion of the signature images.
the dataset is set aside for validation. This portion is
used during the training process to monitor the model’s The VGG-16 and ResNet-50 models in this experi-
performance and adjust hyperparameters accordingly. The ment are obtained from the Keras library. These models
validation set helps in tuning the model’s parameters to were pre-trained using the ImageNet dataset and further
improve its performance without introducing bias from trained by the signature training dataset. In addition, the
the testing data. A 20% allocation is sufficient for vali- layers after the last convolutional layer are removed and
dation purposes while still allowing for effective model replaced with a Global Average Pooling (GAP) layer.
tuning. The remaining portion of the dataset is reserved This will result in 512 and 2048 embeddings, respectively.
for testing the trained models. This independent dataset is Afterwards, these embeddings will be used for the triplet
crucial for evaluating the final performance of the models loss optimization method to further enhance the em-
and assessing their generalization capabilities on unseen bedding models. Lastly, after obtaining the embeddings
data. By keeping a separate testing set, it ensures that from the respective signature images, a feature gallery
the evaluation metrics are reliable and not influenced by is formed by averaging all the embeddings. Their final
the training or validation process. Allocating 20% of the modified form can be seen in Figure 3 and 4.
data for testing provides a substantial sample size for D. Triplet Loss Optimization
robust evaluation. Overall, the 60/20/20 split strikes a
To further improve the embedding model’s capabil-
balance between effective training, validation, and testing
ities, the proposed method takes advantage of an opti-
of the models, ensuring that they can learn meaningful
mization method called Triplet Loss. This optimization
patterns from the data, generalize well to new examples,
process requires triplets as inputs which contain three
and provide reliable performance metrics. Additionally,
different inputs, which are the anchor image, the positive
the use of stratification ensures that each class or category
image (genuine signature), and negative image (forged
within the dataset is represented proportionally across the
signature). The anchor image is obtained from the first
training, validation, and testing sets, reducing the risk
instance of the genuine signature images of an identity
of bias in the evaluation process. The following split is
while the postive images will take the remaining genuine
executed to each type of sample for a single instance.
signatures.
Lastly, the final data count of this research can be seen
in Table I. However, due to the nature of triplets, the inequality of
the genuine signatures and forged signatures data must be
equalized. This is done by first counting which class has
the most instances and randomly repeating the signature
https:// journal.uob.edu.bh
272 Andreas Christianto, et al: Authentic Signature Verification. . .

Figure 3. Feature Extractor from the VGG-16 architecture [34]

Figure 4. Feature Extractor from the ResNet-50 architecture [35]

Figure 5. Triplet Loss Optimization process [36]

images of the other class until the numbers are the same F. Evaluation
while the anchor images are simply the same for each For the evaluation of this case, the AUC evaluation
triplet. The triplet loss optimization architecture can be metric is used [37], which is a commonly employed
seen in Figure 5. performance measure in binary classification tasks. This
method measures the machine learning model’s discrim-
E. Classifiers for Signature Verification inative ability across different decision thresholds by
In order to fully utilize the image embeddings, this plotting the Receiver Operating Characteristic curve and
research uses several machine learning classifiers. These calculating the area beneath it, which is referred to as
classifiers are SVM, ANN, Random Forest, and XGBoost. Area Under Curve (AUC). In addition, the True Positive
Externally, the classifiers are not much different as all Rate (TPR) values at certain False Positive Rate (FPR)
of them take the vector difference between the feature values such as 0.1, 0.01, and 0.001 are also evaluated.
gallery and a signature’s image embeddings as inputs.
However, their internal structures are different from one 5. EXPERIMENTAL SETUP
another and may offer different perspectives on how the A. Hardware Specification
data can be used. Thus, offering various results on various This research is conducted on a laptop equipped
occasions such as data size. with an Intel(R) Core(TM) i5-8265U CPU @ 1.60GHz
The models are trained using the training data ob- processor, featuring four cores and eight threads. The
tained from preprocessing. Additionally, to make sure CPU operates at a base clock rate of 1.60GHz, with a
that the models are optimal, the proposed method takes maximum turbo boost frequency of [insert boost clock
advantage of hyperparameter tuning method, specifically speed if available]. The system is configured with 8 GB
the grid search method. This process uses the validation of system memory (RAM). Additionally, the laptop is
data instead of the training data to save computational equipped with a 476 GB INTEL SS solid-state drive
power and time. (SSD) for storage.

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 273

B. Software Tools
The software used to conduct the research is Visual
Studio Code (Version 1.86.2) as the Primary Integrated
Environment (IDE) for coding and project management.
Python (Version 3.11.1) was the programming language
used for implementing the entire system. Additionally,
the libraries numpy, tensorflow, os, cv2, sklearn, and mat-
plotlib for the various tasks. Furthermore, Visual Studio
Code was used on a Windows 10 operating system for
compatibility and ease of development.
6. RESULTS AND DISCUSSION
A. Testing Results without Triplet Loss Optimization

Figure 8. ROC Curve for BENGALI dataset with VGG16 and

without Triplet Loss Optimization

Figure 6. ROC Curve for ICDAR dataset with VGG16 and without
Triplet Loss Optimization

Figure 9. ROC Curve for BENGALI dataset with ResNet-50 and

without Triplet Loss Optimization

It can also be seen from Figure 5 that two curves,

belonging to Random Forest and XGBoost respectively,
stood out as the top two curves the graph due to them
being entirely above the other two curves. However, in
Figure 6 to Figure 8, all curves nearly overlaps with one
another making harder for visual analysis. Thus, for more
information, the numerical analysis in Table II is needed.
For the combination of ICDAR dataset and VGG16
embedding model, the best machine learning classifier is
Random Forest with an AUC of 0.970 and an average
Figure 7. ROC Curve for ICDAR dataset with ResNet-50 and
without Triplet Loss Optimization TPR of 0.698. On the other hand, the best machine
learning classifier for the ResNet-50 embedding model
From Figure 5 to Figure 8, the results of the different is Support Vector Machine with an AUC of 0.952 and an
type of combinations of datasets, deep learning embed- average TPR of 0.564.
ding models without triplet loss optimization, and ma-
Unlike the ICDAR dataset, the combination of the
chine learning classifiers are shown through ROC curves.
BENGALI dataset and the VGG16 embedding model’s
Additionally, the numerical summary of the combinations
best machine learning classifier is XGBoost with an AUC
are also shown in Table II.
of 0.972 and an average TPR of 0.677. While the best
https:// journal.uob.edu.bh
274 Andreas Christianto, et al: Authentic Signature Verification. . .

TABLE II. RESULTS TABLE WITH TRIPLET LOSS OPTI-

MIZATION

Without Triplet Loss Optimization

ICDAR Dataset
Model AUC TPR at TPR at TPR at
FPR 0.1 FPR FPR
0.01 0.001
VGG16 0.887 0.623 0.142 0.024
+ SVM
VGG16 0.928 0.764 0.34 0.113
+ ANN
VGG16 0.97 0.896 0.651 0.547
+ RF
VGG16 0.967 0.896 0.703 0.542
+ XGB
ResNet- 0.952 0.849 0.608 0.236
50 +
SVM
Figure 10. ROC Curve for ICDAR dataset with VGG16 and with
ResNet- 0.935 0.821 0.222 0.094 Triplet Loss Optimization
50 +
ANN
ResNet- 0.943 0.835 0.486 0.08
50 +
RF
ResNet- 0.952 0.844 0.387 0.066
50 +
XGB
BENGALI Dataset
VGG16 0.959 0.854 0.544 0.216
+ SVM
VGG16 0.945 0.886 0.254 0.023
+ ANN
VGG16 0.97 0.916 0.65 0.558
+ RF
VGG16 0.972 0.94 0.716 0.374
+ XGB
ResNet- 0.952 0.89 0.45 0.136
50 +
SVM
ResNet- 0.958 0.906 0.684 0.548 Figure 11. ROC Curve for ICDAR dataset with ResNet-50 and
with Triplet Loss Optimization
50 +
ANN
ResNet- 0.977 0.936 0.684 0.198
50 +
RF
ResNet- 0.976 0.938 0.756 0.286
50 +
XGB

machine learning classifier for the ResNet-50 embedding

model is Random Forest with an AUC of 0.977 and an
average TPR of 0.606.
B. Testing Results with Triplet Loss Optimization
Just like in sub section A, Figure 9 to Figure 12
represents the results of the combinations of datasets,
deep learning embedding models, and machine learning
classifiers. However, these embedding models in the four
figures are optimized through triplet loss optimization.
Additionally, Table III represents the numerical summary Figure 12. ROC Curve for BENGALI dataset with VGG16 and
of the combinations. with Triplet Loss Optimization

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 275

TABLE III. RESULTS TABLE WITHOUT TRIPLET LOSS

OPTIMIZATION

With Triplet Loss Optimization

ICDAR Dataset
Model AUC TPR at TPR at TPR at
FPR 0.1 FPR FPR
0.01 0.001
VGG16 0.991 0.981 0.703 0.651
+ SVM
VGG16 0.987 0.986 0.613 0.057
+ ANN
VGG16 0.989 0.981 0.811 0.618
+ RF
VGG16 0.984 0.967 0.712 0.354
+ XGB
ResNet- 0.941 0.835 0.458 0.448
50 +
Figure 13. ROC Curve for BENGALI dataset with ResNet-50 and
SVM
with Triplet Loss Optimization ResNet- 0.981 0.939 0.684 0.538
50 +
ANN
In Figure 10, there are three curves, which represent ResNet- 0.974 0.925 0.712 0.656
the ANN, Random Forest, and XGBoost classifiers, which 50 +
stood out among the four curves. However, Figure 9, RF
Figure 11, and Figure 12 showed that the curves nearly ResNet- 0.981 0.943 0.83 0.797
overlap with each other. Thus, further numerical analysis 50 +
using Table III is necessary. XGB
BENGALI Dataset
For the ICDAR dataset and VGG16, SVM is the best VGG16 0.968 0.928 0.518 0.156
classifier, achieving an AUC of 0.991 and an average TPR + SVM
of 0.778. Other than that, the XGBoost classifier achieved VGG16 0.971 0.916 0.656 0.008
an AUC of 0.981 and an average TPR of 0.857 for the + ANN
ResNet-50 embedding model. VGG16 0.97 0.918 0.584 0.398
+ RF
In contrast to the ICDAR dataset, the BENGALI and VGG16 0.971 0.928 0.48 0.07
VGG16 combination’s best machine learning classifier is + XGB
ANN, achieving an AUC of 0.971 and an average TPR ResNet- 0.96 0.874 0.562 0.36
of 0.527. Lastly, the best machine learning classifier for 50 +
the combination of BENGALI and ResNet-50 is Random SVM
Forest with an AUC of 0.972 and an average TPR of ResNet- 0.961 0.876 0.576 0.188
0.648. 50 +
C. Summary of Testing Results ANN
ResNet- 0.972 0.91 0.682 0.614
For the ICDAR dataset, the best method without triplet
50 +
loss optimization is through the combination of VGG16
RF
and Random Forest with an AUC of 0.970 and an average
ResNet- 0.957 0.872 0.698 0.534
TPR at 0.698. On the other hand, the best method with
50 +
triplet loss optimization is through the combination of
XGB
VGG16 and SVM with an AUC of 0.991 and an average
TPR of 0.778.
Meanwhile for the BENGALI dataset, the best method necessarily mean that the overall model got worse. While
without triplet loss optimization is through the combina- the AUC indeed decreased, the average TPR received an
tion of ResNet-50 and Random Forest with an AUC of increase of 0.042 which means that the overall model
0.977 and an average TPR of 0.606. In addition, the same improved its sensitivity or consistency under different
combination still retained its spot for the best method with conditions. Furthermore, from the two datasets, the model
triplet loss optimization, having an AUC of 0.972 and an with triplet loss optimization performs better or showed
average TPR of 0.648. more improvements on the ICDAR dataset, which con-
tains a lesser amount of data. Hence, it can also be
It can be summarized that the triplet loss method can summarized that the triplet loss optimization method can
improve the AUC of the proposed methods. Instances help when data are scarce.
such as the best BENGALI method losing around 0.005
points in AUC after triplet loss optimization does not Although the performance metric and the dataset
https:// journal.uob.edu.bh
276 Andreas Christianto, et al: Authentic Signature Verification. . .

used in this study are different than the other state of In addition, the proposed method already achieved high
the arts methods, comparisons can roughly be made. results even without triplet loss optimization to improve
The proposed method’s best performance achieved the the embedding model, future work may consider the pos-
AUC of 0.991 and 0.977 for the ICDAR and BENGALI sibility of a noisier dataset and more extreme data sizes,
dataset respectively. On the other hand, the state of the be it lower or higher than the ones used in this experiment.
arts methods produced various results as such, we will In addition, this study’s evaluation only considers the
compare this paper’s results with the papers that inspired results of the proposed method, not including efficiency.
this research. They achieved the accuracy of 95% and Therefore, any future work continuing this research may
96.33% respectively [4, 5]. Thus, it can be concluded that consider improving the efficiency aspect in general.
the proposed method can contend with other state of the
arts methods. References
[1] P. Bharti and C. Natarajan, “Deep learning-based signature verifi-
7. Conclusion and Future Work cation,” IJARIIE-ISSN(O)-2395-4396, vol. 9, no. 4, p. 720, 2023.
A. Conclusion
[2] Y. Muhtar, M. Muhammat, N. Yadikar, A. Aysa, and K. Ubul,
In this research, the integration of advanced tech- “Fc-resnet: A multilingual handwritten signature verification
nologies, particularly deep learning and machine learning model using an improved resnet with cbam,” Applied Sciences,
techniques, was explored to overcome the limitations vol. 13, no. 14, p. 8022, Jul 2023. [Online]. Available:
of traditional signature verification methods. The study http://dx.doi.org/10.3390/app13148022
focused on leveraging the combination between two
[3] M. Haratatik and K. Anam, “Comparison of convolutional neural
powerful deep learning models (VGG-16 and ResNet- network architecture on detection of helmet use by humans,”
50) and machine learning signature verification models vol. 8, p. 44, May 2023.
(SVM, ANN, Random Forest, and XGBoost), alongside
the triplet loss optimization method, to enhance the accu- [4] M. Bajpai, “Handwritten signature verification using deep learn-
racy of signature verification. The experiments conducted ing technique in conjunction with image processing,” 2022,
involved the application of these models on two distinct national College of Ireland, Jan. 2022.
signature datasets, ICDAR-2011 and BHSig260-Bengali.
[5] N. Xamxidin, Mahpirat, Z. Yao, A. Aysa, and K. Ubul,
The evaluation metrics, including AUC and TPR at spe- “Multilingual offline signature verification based on improved
cific FPR values, provided a comprehensive assessment inverse discriminator network,” Information, vol. 13, no. 6, p.
of the proposed methods’ performance. The findings indi- 293, Jun 2022. [Online]. Available: http://dx.doi.org/10.3390/
cated that the triplet loss optimization method contributed info13060293
to better overall model performance, emphasizing its role
[6] D. Engin, A. Kantarci, S. Arslan, and H. K. Ekenel,
in enhancing the discernment between genuine and forged “Offline signature verification on real-world documents,” in
signatures. 2020 IEEE/CVF Conference on Computer Vision and Pattern
Recognition Workshops (CVPRW), Jun 2020.
The results highlighted the effectiveness of combining
VGG-16 and Random Forest for the ICDAR dataset and [7] J. Poddar, V. Parikh, and S. K. Bharti, “Offline signature
ResNet-50 with Random Forest for the Bengali dataset recognition and forgery detection using deep learning,” Procedia
without the triplet loss optimization method. On the other Computer Science, vol. 170, pp. 610–617, Apr 2020. [Online].
Available: http://dx.doi.org/10.1016/j.procs.2020.03.133
hand, the combination of VGG-16 and SVM is better for
the ICDAR dataset while ResNet-50 and Random Forest [8] J. Lu, H. Qi, X. Wu, C. Zhang, and Q. Tang, “Research on
remain as the best combination for the BENGALI dataset authentic signature identification method integrating dynamic and
with the triplet loss optimization method. Notably, the static features,” Applied Sciences, vol. 12, no. 19, p. 9904, Oct
triplet loss method demonstrated improvements in the 2022. [Online]. Available: http://dx.doi.org/10.3390/app12199904
models’ sensitivity, especially in scenarios with limited
data, as observed in the ICDAR dataset. In conclusion, [9] Y. Borse, A. Patil, S. Shah, and A. Gharu, “Signature verification
using deep learning,” International Journal of Creative Research
the combination of deep learning embedding models Thoughts (IJCRT), vol. 11, no. 4, pp. h956–h960, Apr 2023.
(VGG-16 and ResNet-50), triplet loss optimization, and
machine learning classifiers (SVM, Random Forest, ANN, [10] O. E. Melhaoui and S. Benchaou, “An efficient signature
XGBoost) presents a versatile and robust solution for recognition system based on gradient features and neural
handwritten signature verification. The proposed methods network classifier,” Procedia Computer Science, vol. 198, pp.
385–390, 2022. [Online]. Available: http://dx.doi.org/10.1016/j.
exhibit promising results, opening avenues for improved procs.2021.12.258
security protocols in various sectors such as banking,
legal institutions, and government agencies. As technol- [11] J. Lopes, B. Baptista, N. Lavado, and M. Mendes, “Offline
ogy continues to advance, the integration of sophisti- handwritten signature verification using deep neural networks,”
cated signature verification systems becomes crucial for Energies, vol. 15, no. 20, p. 7611, Oct 2022. [Online]. Available:
maintaining the integrity of electronic transactions and http://dx.doi.org/10.3390/en15207611
safeguarding sensitive information.
[12] E. Alajrami, B. A. M. Ashqar, B. S. Abu-Nasser, A. J. Khalil,
M. M. Musleh, A. M. Barhoom, and S. S. Abu-Naser, “Hand-
B. Future Work written signature verification using deep learning,” International
In this work, the images were resized down to 64 Journal of Academic Multidisciplinary Research (IJAMR), vol. 3,
by 64 due to the limitations of the devices used in this no. 12, pp. 39–44, Dec 2019.
research. Future works may consider the possibility of
larger image resizes for the possibility of better results. [13] K. Simonyan and A. Zisserman, “Very deep convolutional net-

https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 16, No.1, 265-277 (Jul-24) 277

works for large-scale image recognition,” arXiv, 2014, doi: [31] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet clas-
10.48550/ARXIV.1409.1556. sification with deep convolutional neural networks,” in Advances
in Neural Information Processing Systems, 2012.
[14] O. M. Parkhi, A. Vedaldi, and A. Zisserman, “Deep face recog-
nition,” in British Machine Vision Conference, 2015. [32] D. Sculley et al., “Hidden technical debt in machine learning
systems,” 2015.
[15] K. He, X. Zhang, S. Ren, and J. Sun, “Deep resid-
ual learning for image recognition,” arXiv, 2015, doi: [33] S. Rendle, “Factorization machines,” in Proceedings of the 2012
10.48550/ARXIV.1512.03385. IEEE International Conference on Data Mining, 2012.

[16] F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified [34] F. Liu, Y. Wang, F.-C. Wang, Y.-Z. Zhang, and J. Lin, “Intelligent
embedding for face recognition and clustering,” 2015, arXiv, and secure content-based image retrieval for mobile users,”
2015, doi: 10.48550/ARXIV.1503.03832. IEEE Access, vol. 7, pp. 119 209–119 222, Aug 2019. [Online].
Available: http://dx.doi.org/10.1109/access.2019.2935222
[17] H. Wang, J. Xiong, Z. Yao, M. Lin, and J. Ren, “Research survey
on support vector machine,” Dec 2017. [Online]. Available: [35] V. N. Vinaykumar, J. A. Babu, and J. Frnda, “Optimal
http://dx.doi.org/10.4108/eai.13-7-2017.2270596 guidance whale optimization algorithm and hybrid deep
learning networks for land use land cover classification,”
[18] Y. Guerbai, Y. Chibani, and B. Hadjadji, “The effective use of EURASIP Journal on Advances in Signal Processing, vol. 2023,
the one-class svm classifier for handwritten signature verification no. 1, Jan 2023. [Online]. Available: http://dx.doi.org/10.1186/
based on writer-independent parameters,” Pattern Recognition, s13634-023-00980-w
vol. 48, no. 1, pp. 103–113, Jan 2015. [Online]. Available:
http://dx.doi.org/10.1016/j.patcog.2014.07.016 [36] H. Xuan, A. Stylianou, X. Liu, and R. Pless, “Hard negative
examples are hard, but useful,” pp. 126–142, Nov 2020. [Online].
[19] P.-Y. Hao, J.-H. Chiang, and Y.-D. Chen, “Possibilistic Available: http://dx.doi.org/10.1007/978-3-030-58568-6 8
classification by support vector networks,” Neural Networks, vol.
149, pp. 40–56, 2022. [Online]. Available: http://dx.doi.org/10. [37] A. P. Bradley, “Auc: A statistically consistent and more
1016/j.neunet.2022.02.007 discriminating measure than accuracy,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 20,
[20] M. Madhiarasan and M. Louzazni, “Analysis of artificial neural no. 3, pp. 289–293, Mar 1998. [Online]. Available: http:
network: Architecture, types, and forecasting applications,” //dx.doi.org/10.1109/34.667881
Journal of Electrical and Computer Engineering, vol. 2022,
Apr 2022. [Online]. Available: http://dx.doi.org/10.1155/2022/ Andreas Christianto is a Computer Sci-
5416722
ence master’s student in Bina Nusantara
University. His interest is in the field of
[21] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
Press, 2016. machine learning and deep learning.

[22] S. Haykin, Neural Networks and Learning Machines. Pearson,

2009.

[23] S. Hochreiter and J. Schmidhuber, “Long short-term memory,”

Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.

[24] J. Schmidhuber, “Deep learning in neural networks: An Jovito Colin is a student from Bina Nu-
overview,” Neural Networks, vol. 61, pp. 85–117, 2015. santara University currently pursuing a
Master’s in Computer Science. His re-
[25] G. Biau, “Analysis of a random forests model,” Journal of search interest’s include deep learning,
Machine Learning Research, vol. 13, p. 1063, Oct 2012. computer vision, embedded systems, and
data mining.
[26] L. Breiman, “Random forest,” Machine Learning, vol. 45, no. 1,
pp. 5–32, Oct 2001. [Online]. Available: http://dx.doi.org/10.
1023/a:1010933404324

[27] A. S. More and D. P. Rana, “Review of random forest

classification techniques to resolve data imbalance,” in 2017 1st Gede Putra Kusuma received PhD de-
International Conference on Intelligent Systems and Information
gree in Electrical and Electronic Engi-
Management (ICISIM), 2017.
neering from Nanyang Technological Uni-
[28] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting versity (NTU), Singapore, in 2013. He
system,” in Proceedings of the 22nd ACM SIGKDD International is currently working as a Lecturer and
Conference on Knowledge Discovery and Data Mining, 2016. Head of Department of Master of Com-
puter Science, Bina Nusantara University,
[29] J. H. Friedman, “Greedy function approximation: A gradient Indonesia. Before joining Bina Nusantara
boosting machine,” The Annals of Statistics, vol. 29, no. 5, pp. University, he was working as a Research
1189–1232, 2001.
Scientist in I2R – A*STAR, Singapore.
[30] O. L. Mangasarian and E. W. Wild, “Multiple additive regres- His research interests include computer vision, deep learning,
sion trees with application in artificial intelligence,” Journal of face recognition, appearance-based object recognition, gamifi-
Human–Robot Interaction, vol. 2, no. 1, pp. 9–32, 2013. cation of learning, and indoor positioning system.

https:// journal.uob.edu.bh