Received 31 March 2024, accepted 20 April 2024, date of publication 23 April 2024, date of current version 14 May 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3392640

Novel Transfer Learning Approach for Driver

Drowsiness Detection Using Eye
Movement Behavior
HAMZA AHMAD MADNI1 , ALI RAZA 2, RUKHSHANDA SEHAR 3, NISREAN THALJI 4,

AND LAITH ABUALIGAH 5,6,7,8,9,10

1 College of Electronic and Information Engineering, Beibu Gulf University, Qinzhou 535011, China
2 Department of Software Engineering, The University of Lahore, Lahore 54000, Pakistan
3 Institute of Computer Science, Khwaja Fareed University of Engineering and Information Technology, Rahim Yar Khan 64200, Pakistan
4 Department of Robotics and Artificial Intelligence, Jadara University, Irbid 21110, Jordan
5 Hourani Center for Applied Scientific Research, Al-Ahliyya Amman University, Amman 19328, Jordan
6 Computer Science Department, Al-Albayt University, Mafraq 25113, Jordan
7 Artificial Intelligence and Sensing Technologies (AIST) Research Center, University of Tabuk, Tabuk 71491, Saudi Arabia
8 MEU Research Unit, Middle East University, Amman 11831, Jordan
9 Department of Electrical and Computer Engineering, Lebanese American University, Byblos 13-5053, Lebanon
10 Applied Science Research Center, Applied Science Private University, Amman 11931, Jordan

Corresponding author: Ali Raza ([email protected])

This work was supported in part by the Counterpart Aid Project from Guangxi University for the Discipline Construction Project of Beibu
Gulf University, China, under Grant FG2024005.

ABSTRACT Driver drowsiness detection is a critical field of research within automotive safety, aimed at
identifying signs of fatigue in drivers to prevent accidents. Drowsiness impairs a driver’s reaction time,
decision-making ability, and overall alertness, significantly increasing the risk of collisions. Nowadays, the
challenge is to detect drowsiness using physiological signals, which often require direct contact with the
driver’s body. This can be uncomfortable and distracting. This study aimed at detecting driver drowsiness
through eye movement behavior imagery of the driver. We utilized a standard image dataset based on
the eye movement behavior of drivers to conduct this research experiment. We proposed a novel transfer
learning-based features generation which combined the strengths of the Visual Geometry Group (VGG-16)
and Light Gradient-Boosting Machine (LGBM) methods. The proposed VGLG (VGG16-LGBM) approach
first extracts spatial features from input eye image data and then generates salient transfer features using
LGBM. Experimental evaluations reveal that the k-neighbors classifier outperformed the state-of-the-art
approach with a high-performance accuracy of 99%. The computational complexity analysis shows that
the proposed approach detects driver drowsiness in 0.00829 seconds. We have enhanced the performance
through hyperparameter tuning and validations using k-fold validation. This research has the potential to
revolutionize driver drowsiness detection, aiming to prevent road accidents and save precious lives.

INDEX TERMS Driver drowsiness, eye images, eye behavior, machine learning, deep learning, transfer
leaning.

I. INTRODUCTION and overall alertness. The consequences of driving while

Driver drowsiness refers to a state of fatigue or sleepiness drowsy can be as severe as those associated with drunk
that impairs a driver’s ability to operate a vehicle safely [1]. driving, including a diminished capacity to maintain lane
It significantly contributes to road accidents worldwide, as it position, judge distances, and process the information from
affects a driver’s reaction time, decision-making abilities, the driving environment effectively [2]. The losses due to
driver drowsiness are extensive and multifaceted, impacting
The associate editor coordinating the review of this manuscript and individual and public safety. On a personal level, drowsy
approving it for publication was Joewono Widjaja . driving increases the risk of accidents, leading to potential
2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by/4.0/ 64765
 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

injuries or fatalities. From an economic perspective, these highlighting the advancements in driver drowsiness detection
accidents can result in substantial financial costs associated methodologies. Section III outlines the design and implemen-
with vehicle repairs, medical bills, insurance premiums, and tation details of our novel approach. Section IV presents the
legal fees. On a broader scale, driver drowsiness contributes findings of our experiments, offering a detailed analysis of
to traffic congestion due to accidents [3], emergency the performance of our proposed methodology compared to
response, and clean-up efforts, further affecting economic existing techniques. Finally, Section V summarizes the key
productivity and the efficiency of transportation networks. contributions of our work, highlighting the significance of our
The loss due to accidents encompasses a broad range of findings in advancing the field of driver drowsiness detection.
financial, emotional, and societal impacts. Between 2000 and
2016, there was a notable increase in both the overall case
fatality rate and human damage resulting from road accidents II. LITERATURE ANALYSIS
in China, with a 19.0% and 63.7% rise respectively [4]. In the literature analysis section dedicated to Driver
A survey conducted in Ontario, Canada in 2002 found Drowsiness Detection Using Eye Movement Behavior,
that over 58% of 750 drivers admitted to driving while a comprehensive review of existing studies reveals a
fatigued or drowsy, and 14.5% reported falling asleep at the multidisciplinary approach that intertwines cognitive science,
wheel within the past year [5]. The United States National computer vision, and automotive engineering. The consensus
Highway Traffic Safety Administration estimated that the among researchers underscores the critical role of eye
economic toll from fatigued or drowsy driving amounted to movement metrics–such as blink rate, blink duration, and
$12.4 billion annually [6]. saccade velocity as reliable indicators of driver fatigue.
This research applied advanced Transfer learning mecha- Technological advancements have facilitated the develop-
nisms [7] for driver drowsiness detection. Transfer learning ment of sophisticated eye-tracking systems capable of real-
is a powerful technique in deep learning that involves taking time monitoring, leveraging algorithms and machine learning
a pre-trained neural network model and fine-tuning it for a models to accurately predict drowsiness onset.
different task. This approach has been increasingly applied The research study [9] aims to detect driver drowsiness
in various domains, including driver drowsiness detection and through eye detection. In this experiment, the NITYMED
leveraging eye movement behavior image data. By utilizing dataset is utilized, which contains various levels of drowsi-
a pre-trained Convolutional Neural Network (CNN) model, ness and videos of drivers. Different deep-learning models
such as VGG-16, we can harness the model’s already are employed in this study. The results demonstrate that using
learned hierarchical feature representations and pass them Convolutional Neural Networks (CNN) for drowsiness detec-
to machine learning models for learning. When applying tion in the eyes yielded the highest accuracy. Specifically,
transfer learning to driver drowsiness detection, the lower ResNet50V2 achieved the best results, with an accuracy of
layers of the pre-trained network, which capture generic 98%. However, the study faced challenges due to low dataset
features like edges and textures, are typically kept frozen. samples and the use of classical approaches.
In contrast, the upper layers are fine-tuned to adapt to the The research study [8] proposed a method for detecting
specific features relevant to drowsiness detection, such as drivers’ eye movements using deep learning models to
eyelid droopiness, blink rate, and eye closure duration. This mitigate road accidents. In this experiment, a drowsiness
fine-tuning process allows the network to learn from the eye dataset, which is publicly available on Kaggle, was utilized.
movement behavior data and accurately classify the driver’s It comprises 2,900 images categorized into four different
state as alert or drowsy. states: open, closed, yawning, and no yawning. The study
The significant research contributions of our study regard- found that the VGG-16 model yielded poor results, whereas
ing driver drowsiness detection are as follows: the CNN model achieved efficient outcomes with an accuracy
• We proposed a novel transfer learning-based features of 97%. The evaluation metrics further indicate a precision of
generation approach VGLG, which combined the 99%, with both recall and F1 scores also at 99%. The study
strengths of the VGG-16 and LGBM methods. The mentioned that classical approaches were used, which were
proposed VGLG approach first extracts spatial features seen as limitations.
from input eye image data and then generates salient The research study [10] introduces driver drowsiness
transfer features using LGBM. detection through eye movement to enhance road safety by
• We have built four machine learning and two deep neural preventing accidents. This proposed study adopts techniques
network techniques for comparison. Each model’s that analyze drivers’ behaviors, specifically focusing on
results are validated using k-fold validation and visual cues to distinguish between opened and closed eyes.
enhanced performance through hyperparameter tuning. For this experiment, images with a resolution of 450 ×
In addition, we determined the computational complex- 320 pixels were utilized. The findings demonstrate that the
ity, entropy feature space, and real-time layers feature deep learning model, specifically a Neural Network (NN),
extractions. attained the highest accuracy rate of 98%. However, the study
The remaining study is organized as follows: Sec- acknowledges limitations associated with the employment of
tion II provides a comprehensive review of existing studies, classical neural network approaches.

64766 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

TABLE 1. The summary of analyzed literature work.

The research study [11] proposes a method for driver The research study [15] proposes a PBCI (Brain-Computer
drowsiness detection using visual features. This experiment Interface) scheme using EEG (electroencephalogram) brain
utilizes a driver drowsiness detection dataset collected from signals developed to detect human drowsiness during driving.
National Tsing Hua University. The proposed visual features In this experiment, an EEG dataset from 12 right-handed
are extracted from a video obtained from a camera installed male subjects was utilized, with EEG signals recorded
on the dashboard. Various machine learning models are at a frequency of 125Hz. Among all seven classifiers
employed in this study, achieving the highest accuracy tested–decision trees (DT), discriminant analysis (DA),
of 98%. However, the study acknowledges limitations, logistic regression (LR), naïve Bayes (NB), support vector
including the use of classical approaches. machines (SVM), k-nearest neighbor (kNN), and ensemble
The research study [12] proposes a method to detect methods–the ensemble approach achieved the most efficient
drowsiness by integrating both non-intrusive and intrusive results. The overall accuracy reported was 85.6%. However,
approaches. The experiments utilize the NTHU-DDD dataset, the study acknowledges that it achieved lower performance
which includes data from 36 individuals displaying various scores, suggesting limitations with classical methods.
behaviors indicative of drowsiness, such as yawning, slow The research study [16] focuses on the detection of
blinking, dozing off, and the wearing of glasses or sunglasses driver drowsiness by analyzing facial features and landmarks.
under varied lighting conditions both during the day and This experiment utilizes two public datasets, in addition
at night. The dataset categorizes instances into drowsy and to creating a specific video-based VBDDD (Video-Based
non-drowsy states. In this study, a hybrid model combining Driver Drowsiness Detection) dataset. Two distinct types of
MTCNN for facial feature recognition and a GSR sensor for features, temporal and spatial, are extracted for analysis. The
physiological signal analysis achieved an accuracy rate of CSP (Common Spatial Pattern) algorithm is employed to
91%. However, the study reported lower performance scores enhance the performance across different classes of samples.
due to certain limitations. The results demonstrate that this method achieved excellent
The research study [13] proposes a method to address the performance.
significant issue of drowsiness. In this experiment, an eye The research study [17] proposes a real-time drowsiness
image dataset is utilized to detect the driver’s eyes within detection system that utilizes a simple camera to analyze
a specific range. The system alerts the driver by triggering the ratio of eye closure and mouth opening. This system
an alarm when an increased rate of drowsiness is detected. is designed to identify drowsy behavior in drivers and alert
Various machine learning techniques are employed in this them to their condition to prevent accidents. The study is
study, yielding efficient results. The machine learning model segmented into two parts: offline and online procedures.
achieved an accuracy of 80%, which contributes to reducing By employing various deep learning models, the research
accidents. However, the study achieved low performance achieves efficient results. The lowest accuracy observed is
scores due to limitations. with the VGG-16 model at 97%, while the highest accuracy
The research study [14] proposes a wearable device is attained using ResNet50, which achieves 98%. However,
designed to detect drowsiness. To ascertain whether the driver classical approaches were noted as limitations.
is asleep, it utilizes a comfortable wearable device based
on signal processing techniques. This study employs various A. RESEARCH GAP
machine learning models, with Random Forest (RF) achiev- In our investigation, we identified several research gaps
ing an accuracy rate of 89.3%. Furthermore, machine learning pertaining to driver drowsiness detection.
algorithms that utilize bagging and boosting techniques have • Previous research predominantly utilized classical
achieved the highest accuracy, at 89.4%. However, the study machine learning methods and relied heavily on sensor
reports lower performance scores, identifying limitations data, with little to no focus on eye movement analysis
associated with classical methods. specifically.

VOLUME 12, 2024 64767

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

FIGURE 1. The workflow architecture analysis of proposed methodology.

• Moreover, these studies often reported low performance performance on the test set. The models were further
scores. However, our research introduces an innovative refined through hyperparameter tuning to enhance their
approach by applying transfer learning mechanisms, accuracy and efficiency. The AI model that demonstrated
which have significantly enhanced the detection accu- superior performance through this validation process was
racy. This novel methodology not only addresses the then employed for the detection of driver drowsiness,
limitations of prior studies but also opens new avenues utilizing the eye movement data as its primary input. This
for more effective and reliable drowsiness detection comprehensive methodology aims to leverage the capabilities
systems. of advanced neural networks and transfer learning to improve
the accuracy of driver drowsiness detection systems.
III. PROPOSED METHODOLOGY
This section delved into the materials and methods employed A. EYE MOVEMENTS IMAGE DATABASE
for detecting driver drowsiness by analyzing the eye move- This study utilized a standard dataset [18] comprising 4,103
ments of drivers. We have clearly outlined the image dataset eye images of drivers, meticulously labelled as either open
used to develop the applied neural network approach. The or closed eye movements. This dataset was generated using
performance scores for each method are revealing and UnityEyes, a cutting-edge eye-synthetic simulator renowned
demonstrate their potential for real-time driver drowsiness for its ability to produce high-quality data. UnityEyes facil-
detection. itated the creation of a diverse and representative collection
Figure 1, illustrates the step-wise working flow of our of eye movement recordings, encompassing subjects from a
novel proposed research methodology. In our research wide range of demographics acquired under variable lighting
methodology, we initially accessed a standard dataset com- conditions. These recordings were captured under controlled
prising eye movement records of drivers. Following this, driving scenarios, ensuring a realistic and applicable dataset
we preprocessed the images, ensuring they were properly for the study. The sample images from the dataset are
formatted for further analysis. We then proposed a novel visualized in Figure 2.
approach based on neural networks that employs transfer
learning to extract salient features from the eye movement B. IMAGE PREPROCESSING AND FORMATIONS
dataset of drivers. This extracted data was subsequently In the initial phase of our research, the image dataset is
divided into training and testing sets with an 80:20 split imported and subjected to basic preprocessing to standardize
ratio. We trained several advanced machine learning models the input data for subsequent analysis. A crucial step in
using the training portion of the data and evaluated their this preprocessing is resizing each image to a uniform

64768 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

that are optimally suited for detecting signs of drowsiness

with unprecedented accuracy. Our empirical research con-
clusively demonstrates that the VGLG method substantially
outperforms existing state-of-the-art classical methods in the
domain of drowsiness detection.
Algorithm 1 contains the unique workflow of our novel
proposed transfer learning method.

Algorithm 1 VGLG Algorithm

Input: Eye movement images data.
Output: New Transfer features.
initiate;
1- Svgg16 ←− VGG16prediction (X ) //
X ϵ eye movements images set, here X is input images data
and Svgg16 is generated rich level spatial features.
2- Slgbm ←− LGBMprobabilities prediction (Svgg16 ) // here Slgbm
is the generated novel probabilistic based transfer features.
3- Fprob ←−Slgbm // here Fprob is the final transfer learning
based probabilistic features set used for driver eye movement
FIGURE 2. The sample closed and open eye movement images. detection.
end;

dimension of 256 × 256 pixels, ensuring that all images D. EMPLOYED ARTIFICIAL INTELLIGENCE APPROACHES
are of the same scale and thus compatible with the neural Driver drowsiness is a critical issue that significantly
network architecture designed for image classification tasks. impacts road safety. Fatigue-related accidents can lead
Moreover, to facilitate supervised learning, the target classes to severe consequences, making it essential to develop
were encoded as ’Closed’: 0 and ’Opened’: 1, representing effective drowsiness detection systems. In recent years,
the two categories of interest in our study. A detailed analysis artificial intelligence (AI) techniques have shown promise in
of the dataset’s composition is conducted, with findings addressing this challenge. One approach involves analyzing
presented in Figure 4. This chart elucidated the distribution eye movement behavior to detect signs of drowsiness.
of samples across the two target classes, revealing a nearly
• Convolutional neural network (CNN): CNN Driver
balanced dataset. Such balance is crucial as it prevents
drowsiness detection using CNNs focuses on analyzing
model bias towards the more prevalent class, thus enhancing
eye movement behavior to identify signs of drowsi-
the reliability and generalizability of the predictive model
ness [19], [20]. The process involves data collection
developed from this dataset.
from videos of drivers, feature extraction of eye regions,
and utilization of a custom-designed CNN architecture.
C. NOVEL PROPOSED TRANSFER FEATURES GENERATION This architecture includes convolutional layers, batch
In the realm of drowsiness detection through driver eye normalization, dropout layers, and fully connected
movement behaviors, our research introduces a novel layers. The trained CNN achieves high accuracy in
method named VGLG, combining the strengths of the detecting drowsiness based on eye movement patterns,
Visual Geometry Group (VGG-16) and Light Gradient contributing to safer driving and reduced accidents. The
Boosting Machine (LGBM) approaches. The architecture convolution operation in a convolutional layer can be
of the proposed approach is illustrated in Figure 10. This described mathematically as follows:
novel approach leverages transfer learning for enhanced XX
feature generation, marking a significant advancement over Yij = (F ∗ X )ij = Fmn · Xi+m,j+n (1)
traditional techniques. Initially, the VGLG method processes m n

an image dataset through the VGG-16 architecture training, where:

a deep convolutional network known for its efficacy in – X represents the input image matrix.
extracting spatial features from images. This crucial step – F denotes the filter (or kernel) matrix.
enables the isolation of intricate patterns and characteristics – Y is the output feature map resulting from the
within the eye movement data, which are indicative of convolution.
the driver’s alertness levels. Subsequently, these extracted – i, j are the spatial indices for the output feature map.
features are fed into the LGBM, a powerful gradient-boosting – m, n are the indices iterating over the filter dimen-
approach that excels in generating predictive features from sions.
data. By harnessing LGBM’s capability to refine and enhance This equation represents the core operation within
these spatial features, VGLG produces highly salient features a convolutional layer, where the filter F is applied

VOLUME 12, 2024 64769

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

FIGURE 3. The transfer features generations mechanism of novel proposed VGLG approach.

representation of the VGG-16 model applied to this task

can be summarized as follows: Given an input image X ,
the model applies a series of convolutional operations.
Each convolutional layer Li in the VGG-16 architecture
transforms the input as follows:

Li (X ) = ReLU (Wi ∗ X + bi ) (2)

where Wi and bi are the weights and bias of the i-th

layer, X is the input to the layer, and ∗ denotes the
convolutional operation. ReLU (Rectified Linear Unit)
is applied for non-linear activation.
After certain convolutional layers, a max-pooling oper-
ation is applied:
FIGURE 4. The Bar chart-based target label data distributions analysis.

Pi (X ) = MaxPooling(Li (X )) (3)

across the input image X to produce a feature map Y , • K-Neighbors-Classifier (KNC): KNC algorithm ana-
highlighting specific features in the input image. lyzes images of a driver’s eyes to assess their drowsiness
• Visual Geometry Group (VGG-16): VGG-16 model, status [23], [24]. Specifically, it measures the Eye
proposed by Karen Simonyan and Andrew Zisserman, Aspect Ratio (EAR), which represents the duration
is a deep neural network designed for large-scale image of eye closure (blinking). When the EAR indicates
recognition [21], [22]. It consists of 16 layers, including drowsiness, warning alarms are triggered at different
13 convolutional layers and 3 fully connected layers. levels of drowsiness during driving. The KNC algorithm
VGG-16 employs small 3 × 3 convolutional filters, classifies these levels based on the time eyes remain
replacing larger filters used in previous models. These closed or the blinking rate, enhancing road safety by
smaller filters enhance non-linearity and allow the alerting fatigued drivers. Let X = {x1 , x2 , . . . , xn } be
network to converge faster. The model achieves remark- the feature vector extracted from the image, and let
able accuracy in image classification tasks, making D = {(X1 , y1 ), (X2 , y2 ), . . . , (Xm , ym )} be the training
it a valuable choice for drowsiness detection based dataset where Xi is the feature vector of the ith training
on eye movement behavior. The basic mathematical example, and yi is its label (e.g., 0 for alert, 1 for drowsy).

64770 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

The Euclidean distance between the feature vector X of The log-odds (logit) can be expressed as:
the new image and a feature vector Xi in the dataset can  
P(Y = 1|X )
be calculated as: log = βX (7)
v 1 − P(Y = 1|X )
u n
uX
d(X , Xi ) = t (xj − xij )2 (4) • Random Forest (RF): RF approach combines multiple
j=1 decision trees to create an ensemble model [29], [30].
In the context of drowsiness detection, it leverages
where xj is the jth feature of vector X , and xij is the jth
features extracted from images of the driver’s eyes.
feature of vector Xi .
Initially, a single-shot scale-invariant face detector
• Light Gradient-Boosting Machine (LGBM): LGBM
identifies the face in the image. Subsequently, facial
is a gradient-boosting algorithm that efficiently handles
features related to eye fatigue, such as yawns, head
large datasets [25], [26]. In the context of driver
posture, and eye closure, are extracted. Finally, the RF
drowsiness detection, it leverages features extracted
technique analyzes these features to assess the driver’s
from eye movement behavior images. These features
drowsiness level. The RF model’s robustness and ability
could include eye closure duration, blink rate, and other
to handle complex feature interactions make it effective
relevant metrics. The LGBM constructs an ensemble
for real-time monitoring and accident prevention. Let:
of decision trees, optimizing for accuracy and speed
during prediction. Learning from labeled data iden- – X be the feature vector extracted from eye move-
tifies patterns associated with drowsiness and non- ment behavior images.
drowsiness, enabling real-time assessment of a driver’s – Y be the binary label (1 for drowsiness, 0 for
alertness level. Its robustness and ability to handle alertness).
high-dimensional feature spaces make it a valuable tool – T be the number of decision trees in the Random
for enhancing road safety. Let: Forest.
– X be the feature vector extracted from eye move- The prediction for each tree t is given by:
ment behavior images. Ŷt = ft (X ) (8)
– Y be the binary label (1 for drowsiness, 0 for
alertness). The ensemble prediction is obtained by averaging
– βt represent the learned coefficients for each tree. individual tree predictions:
– ht (X ) be the prediction of the t-th tree. T
The log-odds of drowsiness are modeled as: 1X
ŶRF = Ŷt (9)
T
  X T t=1
P(Y = 1|X )
log = βt ht (X ) (5)
1 − P(Y = 1|X ) E. HYPER-PARAMETERS OPTIMIZATIONS
t=1
We conducted hyperparameter optimization for the applied
• Logistic Regression (LR): LR is a widely used binary
neural network techniques, and the selected parameters are
classification algorithm [27], [28]. In the context of
detailed in Table 2. Hyperparameter optimization signifi-
drowsiness detection, it analyzes features extracted from
cantly enhances the performance accuracy and generalization
eye movement behavior images. These features could
of the applied methods. We identified the optimal hyperpa-
include metrics like eye closure duration, blink rate, and
rameters through a recursive process of training and testing,
other relevant parameters. LR models the probability
utilizing a k-fold cross-validation mechanism. This analytical
of a driver being drowsy based on these features.
approach has proven effective in improving the performance
It estimates the log odds of drowsiness and applies
scores of driver drowsiness detection.
a sigmoid function to convert them into probabilities.
If the probability exceeds a predefined threshold, the
IV. RESULTS AND DISCUSSIONS
system alerts the driver. LR’s simplicity, interpretability,
In this section, we focused on driver drowsiness detection,
and efficiency make it suitable for real-time applications
experimental results are derived from the application of
in road safety. Let:
various methods analyzing eye movement images. To ensure
– X be the feature vector extracted from eye move- the reliability and accuracy of the findings, numerous
ment behavior images. model evaluation parameters are employed. These parameters
– Y be the binary label (1 for drowsiness, 0 for facilitated a comprehensive assessment of each method’s per-
alertness). formance, enabling the identification of the most promising
– β represent the learned coefficients. approaches in detecting driver drowsiness.
The logistic function (sigmoid) transforms the linear
combination: A. EXPERIMENTAL SETUP
1 The experimental setup for the detection of driver drowsiness
P(Y = 1|X ) = (6)
1 + e−βX through eye movement images is meticulously designed to

VOLUME 12, 2024 64771

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

TABLE 2. Hyperparameter optimizations analysis of applied neural

network models for image analysis.

ensure the reliability and accuracy of the results. Utilizing

the Google Colab platform, the study harnessed the power of
a GPU-accelerated backend, which is critical for processing
the extensive data involved in this research. This setup is
equipped with 13 GB of RAM and 90 GB of storage capacity,
providing the necessary computational resources to handle
the demanding tasks of image analysis and machine learning
model training. To rigorously evaluate the performance of
the applied machine learning models, a comprehensive suite
of metrics is employed. These included accuracy, precision,
recall, and F1 scores, which together offered a nuanced view
of the models’ effectiveness in correctly identifying signs of
drowsiness from eye movement images.

B. RESULTS WITH CLASSICAL CNN METHOD

In our study, we initially implemented a classical CNN
model to analyze a specific time series dataset. The training
process and its impact on model performance are detailed in
Figure 5. Initially, the model exhibited high loss scores and
low accuracy, attributable to the random weight initializations
typical at the start of training. As training progressed,
the CNN learned to fine-tune its parameters, leading to a
gradual improvement in accuracy and a decrease in loss
scores. Despite these improvements, performance metrics
remained around 90%, indicating a relatively low-efficiency FIGURE 5. Time series performance of applied classical CNN method
during the training.
level. This suggests that while the model learns from the
data, there remains a significant need for performance
TABLE 3. Classification report results of CNN model for testing data.
enhancement strategies to further refine and improve its
predictive capabilities.
The study on driver drowsiness detection using a CNN
model, as detailed in Table 3, demonstrates that the model
can identify drowsiness with a 91% accuracy rate for unseen
data. However, the performance notably declines for the
specific task of recognizing the ‘eye closed’ class, indicating C. RESULTS WITH CLASSICAL VGG-16 METHOD
a critical area for improvement. Despite the high overall In our study, we initially implemented a classical VGG16
accuracy, these results underline the necessity for further model to analyze a specific time series dataset. The training
enhancements in the model’s capability to ensure more process and its impact on model performance are detailed in
reliable detection of driver drowsiness, particularly in critical Figure 6. Initially, the model exhibited high loss scores and
situations where the driver’s eyes are closed. This suggests low accuracy, attributable to the random weight initializations
that while CNN models hold promise for this application, typical at the start of training. As training progressed, the
optimization, and refinement are essential for achieving the VGG16 learned to fine-tune its parameters, leading to a
levels of performance required for real-world deployment. gradual improvement in accuracy and a decrease in loss

64772 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

TABLE 4. Classification report results of VGG16 model for testing data.

TABLE 5. Testing results with spatial features of VGG16.

situations where the driver’s eyes are closed. This suggests

that while VGG16 models hold promise for this application,
optimization, and refinement are essential for achieving the
levels of performance required for real-world deployment.

D. RESULTS WITH NOVEL PROPOSED TRANSFER

APPROACH
In this section, we examine the performance results of
our novel proposed transfer learning approach in detail.
Our approach leverages a transfer mechanism specifically
designed to extract salient features from driver eye images.
By utilizing this method, we created a new feature set that
was then comparatively evaluated. The comparative analysis
aimed to assess the effectiveness and efficiency of the
feature set generated through our transfer learning approach,
demonstrating its potential benefits in applications requiring
accurate eye movement and attention tracking, such as driver
FIGURE 6. Time series performance of applied classical VGG16 method
during the training. monitoring systems for enhanced road safety.

1) RESULTS WITH SPATIAL FEATURES

scores. Despite these improvements, performance metrics In this section, we aimed to evaluate the effectiveness
remained around 92%, indicating a relatively low-efficiency of Spatial Features extracted via the VGG16 model in
level. This suggests that while the model learns from the driver drowsiness detection. We fed these Spatial Features
data, there remains a significant need for performance into various machine-learning models and assessed their
enhancement strategies to further refine and improve its performance. According to the results presented in Table 5,
predictive capabilities. the Light Gradient Boosting Machine (LGBM) method
The study on driver drowsiness detection using a VGG16 stood out by achieving satisfactory performance scores,
model, as detailed in Table 4, demonstrates that the model while other machine learning approaches yielded moderate
can identify drowsiness with a 94% accuracy rate for unseen results. This outcome suggests that although Spatial Features
data. However, the performance is notably moderate for the extracted from eye images of drivers hold potential, the
specific task of recognizing the ‘eye closed’ class, indicating majority of tested machine-learning models fail to fully
a critical area for improvement. Despite the high overall leverage these features. Therefore, it is imperative to explore
accuracy, these results underline the necessity for further alternative transfer learning mechanisms that could enhance
enhancements in the model’s capability to ensure more the performance of utilizing prominent Spatial Features for
reliable detection of driver drowsiness, particularly in critical improved driver drowsiness detection.

VOLUME 12, 2024 64773

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

TABLE 6. Testing results with novel proposed transfer learning features.

2) RESULTS WITH NOVEL PROPOSED FEATURES

The evaluation of novel proposed features in applied machine
learning methods for driver drowsiness detection demon-
strated significant improvements in performance metrics,
as detailed in Table 6. The experiments highlighted the
effectiveness of the proposed transfer feature engineering
technique, with applied machine learning models like LGBM,
LR, and RFC achieving performance scores around 98%.
Notably, the proposed KNC method outperformed traditional
approaches, achieving an exceptional accuracy score of 99%.
This analysis underscores the significance and potential of the FIGURE 7. Histogram chart-based results comparisons of spatial and
novel proposed feature engineering approach in enhancing novel features.
machine learning applications, particularly in the context
of driver drowsiness detection, showcasing its superior
performance compared to classical methods.
In addition to our comprehensive analysis, we conducted
histogram-based comparisons, as illustrated in Figure 7,
to further substantiate our findings. The analysis explicitly
demonstrated that the utilization of solely spatial fea-
tures yielded suboptimal results in the context of driver
drowsiness detection. However, a significant enhancement
in performance was observed when we integrated both
spatial features and class probability features, as proposed
in our approach. This integration facilitated a more nuanced
and accurate detection mechanism, ultimately leading to
markedly higher performance scores. Therefore, the data
convincingly supports the superiority of our proposed method
for detecting driver drowsiness, highlighting the critical role
FIGURE 8. Confusion matrix-based performance validation of applied
of combining spatial and probabilistic features to achieve learning techniques.
optimal detection accuracy.
This analysis focused on driver drowsiness detection,
various predictive models are evaluated for their accuracy performance in the context of driver drowsiness detection.
and reliability. According to Figure 8, which presents The findings suggest that the KNC method is a highly reliable
the confusion matrix-based prediction error rates for each tool for identifying signs of drowsiness among drivers,
model, a comprehensive analysis reveals that all models thereby contributing to enhanced road safety measures.
demonstrated a commendable performance with notably
high rates of correct predictions and minimal instances E. NOVEL TRANSFER NEURAL LAYERS FEATURES
of incorrect predictions. Among these, the proposed KNC EXTRACTION
method emerged as particularly noteworthy, recording only In this section, we have successfully demonstrated the
12 instances of incorrect predictions while correctly identi- application of transfer learning and feature extraction using
fying the remaining cases. This striking outcome underlines our proposed approach, as illustrated in Figure 9. Our
the effectiveness of the KNC approach, validating its superior analysis focuses on the extraction of spatial features from

64774 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

TABLE 7. Results validations of applied machine learning methods. TABLE 9. State-of-the-art method comparisons with the proposed
approach.

TABLE 8. Runtime computational complexity analysis of applied

methods.

Figure 10. Our analysis revealed that these novel transfer

features are highly linearly separable, indicating that our
proposed innovative approach successfully extracts features
sample images of drivers’ eyes. This process involves that significantly contribute to the model’s effectiveness.
identifying and extracting prominent features from each Consequently, this analysis supports the conclusion that the
image, which are critical in understanding the visual cues newly developed features play a crucial role in efficiently
indicative of the driver’s state, such as alertness or fatigue. detecting driver drowsiness, underscoring the value of our
By harnessing these extracted features, we have been able to proposed method in enhancing performance in this domain.
construct machine-learning models that exhibit high accuracy
scores. This achievement underscores the effectiveness of I. STATE OF THE ART STUDIES COMPARISONS
our method in not only accurately capturing essential visual For a fair comparison, we performed the state-of-the-art
information from the eye images but also in leveraging method comparisons as described in Table 9. Most previous
this information to enhance the predictive capability of the studies utilized classical machine learning and neural network
models. frameworks. We proposed a novel transfer of features.
Comparisons reveal that our novel approach outperformed
F. K-FOLD CROSS-VALIDATION ANALYSIS state-of-the-art studies with high-performance scores for
After achieving high-performance scores, we employed k- driver drowsiness detection.
fold validation mechanisms, specifically dividing the dataset In addition, we have compared our proposed method
into 10 folds, to validate the efficacy of each applied method with benchmark datasets and recent work such as YOLO
for driver drowsiness detection. The results, as detailed in models in this analysis. The comparison results also show
Table 7, underscore the robustness of our methods, with per- the superiority of our novel proposed approach or image
formance scores consistently exceeding 98% and exhibiting classification.
minimal standard deviation. This comprehensive analysis not
only confirms the high accuracy of our applied methods J. ABLATION STUDIES
but also demonstrates their generalizability across different This subsection demonstrates the performance effects of the
data segments, reinforcing their potential applicability in ablation study conducted to assess the proposed transfer
real-world scenarios for enhancing driver safety. learning-based system by systematically removing specific
components. This ablation analysis contributes to evaluating
G. COMPUTATIONAL COMPLEXITY ANALYSIS the overall performance results based on the significance of
The computational complexity analysis for driver drowsiness each component.
detection methods, as depicted in Table 8, highlights the Primarily, we evaluated the performance impact of our
efficiency of the proposed KNC detection algorithm. The novel proposed VGLG and KNC approach compared to
analysis reveals that the KNC method can identify critical traditional neural network models. The results indicate that
driver drowsiness in just 0.008 seconds, outperforming other upon removal of the proposed VGLG and KNC approach,
tested algorithms in terms of speed. This indicates that the we observed noteworthy results. However, by employing our
KNC detection algorithm not only achieves rapid detection invented transfer learning approach, the KNC method out-
times but also implies potential for real-time application in performed state-of-the-art methods with a high performance
driver safety systems, where quick response times are crucial accuracy of 99% in detecting driver drowsiness. This analysis
for preventing accidents caused by drowsy driving. underscores the superiority and effectiveness of the proposed
approach over classical methods.
H. ENTROPY FEATURE SPACE ANALYSIS The findings from this ablation study analysis highlight the
To investigate the cause of high-performance scores in significant importance of our proposed approach in detecting
detecting driver drowsiness, we conducted a feature space driver drowsiness. Overall, the results of the ablation study
analysis on newly created transfer features, as depicted in reinforce the robustness of our method.

VOLUME 12, 2024 64775

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

FIGURE 9. The proposed features extraction mechanism analysis from a driver eye image movement.

(RNNs), offer avenues for exploration. These models could

enhance the system’s ability to learn complex patterns in
eye movement behavior associated with drowsiness, thereby
improving detection accuracy.
Moreover, the computational complexity of our proposed
method is another area for discussion and improvement.
While we aim to develop a system that is both efficient and
effective, the balance between computational demands and
real-time performance is crucial, especially for in-vehicle
systems where processing power and energy consumption are
limited. Future research could focus on optimizing the model
to reduce computational complexity without compromising
detection accuracy. This might involve investigating lighter
models or techniques such as model pruning and quantiza-
tion, which can significantly reduce the computational load.
In our proposed transfer learning-based feature engineer-
ing approach from image data, specifically eye scans of
a driver during driving, several limitations and challenges
FIGURE 10. The feature space analysis of newly created transfer features. emerge as some disadvantages. Firstly, the variability in
lighting conditions within the vehicle presents a significant
challenge. The effectiveness of feature extraction can be
K. DISCUSSION AND STUDY LIMITATIONS drastically affected by changes in natural light throughout
This research aims to leverage eye movement behavior as a the day or artificial light sources at night, compromising the
significant indicator of drowsiness, utilizing transfer learning consistency and reliability of the derived features. Secondly,
to enhance detection accuracy. However, it’s essential to the quality of the camera capturing the eye scans plays a
recognize that while our findings are promising, the accuracy crucial role in the fidelity of the image data. Lower-quality
of the drowsiness detection system can still be improved. cameras may introduce noise and reduce the resolution of
Further exploration into more advanced neural network the images, thereby hampering the transfer learning model’s
architectures could potentially yield better performance and ability to accurately identify and extract relevant features.
more robust detection capabilities. This issue is exacerbated in dynamic driving environments
One of the significant discussions around our approach is where vibrations and movements can further degrade image
the potential for applying more sophisticated neural network quality.
models. Current advancements in deep learning and neural The scope of our research is primarily focused on eye
networks, such as CNNs and recurrent neural networks movement behavior as an indicator of drowsiness, which,

64776 VOLUME 12, 2024

H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

while significant, does not account for all possible drowsiness [5] M. H. Smolensky, L. Di Milia, M. M. Ohayon, and P. Philip, ‘‘Sleep
indicators. Other physiological or behavioral signals, such as disorders, medical conditions, and road accident risk,’’ Accident Anal.
Prevention, vol. 43, no. 2, pp. 533–548, Mar. 2011.
heart rate variability or yawning, could also be integrated into [6] T. Zhang and A. H. S. Chan, ‘‘Sleepiness and the risk of road accidents
the detection system to improve its accuracy and robustness. for professional drivers: A systematic review and meta-analysis of
retrospective studies,’’ Saf. Sci., vol. 70, pp. 180–188, Dec. 2014.
[7] J. Chen, H. Wang, and E. He, ‘‘A transfer learning-based CNN deep
V. CONCLUSION AND FUTURE DIRECTIONS
learning model for unfavorable driving state recognition,’’ Cognit.
This study aimed to detect driver drowsiness through the Comput., vol. 16, no. 1, pp. 121–130, Jan. 2024.
imagery of eye movement behavior. We utilized a standard [8] M. I. B. Ahmed, H. Alabdulkarem, F. Alomair, D. Aldossary, M. Alahmari,
image dataset reflecting drivers’ eye movement behavior M. Alhumaidan, S. Alrassan, A. Rahman, M. Youldash, and G. Zaman,
‘‘A deep-learning approach to driver drowsiness detection,’’ Safety, vol. 9,
for this research experiment. We proposed a novel transfer no. 3, p. 65, 2023.
learning-based feature generation method that combines [9] R. Florez, F. Palomino-Quispe, R. J. Coaquira-Castillo,
the strengths of the VGG-16 and LGBM methods. The J. C. Herrera-Levano, T. Paixão, and A. B. Alvarez, ‘‘A CNN-based
approach for driver drowsiness detection by real-time eye state
proposed VGLG approach first extracts spatial features from
identification,’’ Appl. Sci., vol. 13, no. 13, p. 7849, Jul. 2023.
input eye image data and then generates salient transfer [10] M. K. Gatea, S. K. Gharghan, and A. H. Ali, ‘‘Deep learning neural network
features using LGBM. Experimental evaluations revealed that for driver drowsiness detection using eyes recognition,’’ in Proc. AIP Conf.,
the k-neighbors classifier outperformed the state-of-the-art vol. 2591, 2023, Art. no. 020008.
[11] Y. Albadawi, A. AlRedhaei, and M. Takruri, ‘‘Real-time machine learning-
approach with a high-performance accuracy of 99%. The based driver drowsiness detection using visual features,’’ J. Imag., vol. 9,
computational complexity analysis shows that our proposed no. 5, p. 91, Apr. 2023.
approach detects driver drowsiness in 0.00829 seconds. [12] J. S. Bajaj, N. Kumar, R. K. Kaushal, H. L. Gururaj, F. Flammini, and
We enhanced the performance through hyperparameter R. Natarajan, ‘‘System and method for driver drowsiness detection using
behavioral and sensor-based physiological measures,’’ Sensors, vol. 23,
tuning and validations using k-fold validation. no. 3, p. 1292, Jan. 2023.
[13] J. Singh, R. Kanojia, R. Singh, R. Bansal, and S. Bansal, ‘‘Driver drowsi-
A. FUTURE WORK ness detection system: An approach by machine learning application,’’
2023, arXiv:2303.06310.
In future research, we aim to further improve performance [14] A. Amidei, S. Spinsante, G. Iadarola, S. Benatti, F. Tramarin, P. Pavan, and
scores by refining our proposed machine-learning model. L. Rovati, ‘‘Driver drowsiness detection: A machine learning approach on
A significant focus will be on the design and deployment skin conductance,’’ Sensors, vol. 23, no. 8, p. 4004, Apr. 2023.
of our model within a camera-based application specifically [15] S. Arif, S. Munawar, and H. Ali, ‘‘Driving drowsiness detection using
spectral signatures of EEG-based neurophysiology,’’ Frontiers Physiol.,
tailored for real-time driver drowsiness detection. This vol. 14, Mar. 2023, Art. no. 1153268.
advanced application will not only detect signs of driver [16] L. Yang, H. Yang, H. Wei, Z. Hu, and C. Lv, ‘‘Video-based driver
fatigue more effectively but will also incorporate a timely drowsiness detection with optimised utilization of key facial features,’’
IEEE Trans. Intell. Transp. Syst., early access, Jan. 5, 2024, doi:
alert system designed to prevent accidents. 10.1109/TITS.2023.3346054.
The tests on human subjects will also be performed to [17] A. Turki, O. Kahouli, S. Albadran, M. Ksantini, A. Aloui, and
validate the proposed method for real-time driver drowsiness M. B. Amara, ‘‘A sophisticated drowsiness detection system via deep
detection. transfer learning for real time scenarios,’’ AIMS Math., vol. 9, no. 2,
pp. 3211–3234, 2024.
[18] H. Fahmy. Drowsiness Detection Dataset. Accessed: Feb. 2, 2024.
CONFLICTS OF INTEREST [Online]. Available: https://www.kaggle.com/datasets/hazemfahmy/
The authors declare that there are no conflicts of interest openned-closed-eyes
[19] E. Hassan, M. S. Hossain, A. Saber, S. Elmougy, A. Ghoneim, and
regarding the publication of this manuscript. Any affiliations, G. Muhammad, ‘‘A quantum convolutional network and ResNet (50)-
or relationships with organizations or entities that might pose based classification architecture for the MNIST medical dataset,’’ Biomed.
a conflict of interest with the subject matter discussed in this Signal Process. Control, vol. 87, Jan. 2024, Art. no. 105560.
work are hereby disclosed. [20] X. Luo, W. Wang, Y. Xu, Z. Lai, X. Jin, B. Zhang, and D. Zhang, ‘‘A deep
convolutional neural network for diabetic retinopathy detection via mining
local and long-range dependence,’’ CAAI Trans. Intell. Technol., vol. 9,
REFERENCES no. 1, pp. 153–166, Feb. 2024.
[21] J. Biju and D. Shanthi, ‘‘Performance analysis of diabetic retinopathy and
[1] S. A. El-Nabi, W. El-Shafai, E.-S.-M. El-Rabaie, K. F. Ramadan, F. E. Abd
hemorrhages segmentation system using enhanced visual geometry group
El-Samie, and S. Mohsen, ‘‘Machine learning and deep learning techniques
(EVGG) architecture,’’ J. Electr. Eng. Technol., pp. 1–12, Feb. 2024.
for driver fatigue and drowsiness detection: A review,’’ Multimedia Tools
Appl., vol. 83, no. 3, pp. 9441–9477, Jan. 2024. [22] M. Gulhane and S. Kumar, ‘‘Oriental method to predict land cover and land
[2] R. Sathya, D. S. S. Harsha, G. P. S. Reddy, and M. G. Krishna, ‘‘IoT-based usage using Keras with VGG16 for image recognition,’’ Adv. Aerial Sens.
driver drowsiness detection and alerting system using Haar cascade and eye Imag., pp. 33–46, 2024.
aspect ratio algorithms,’’ in Integration of AI-Based Manufacturing and [23] A. Naseer, M. Amjad, A. Raza, K. Munir, N. A. Samee, and M. A. Alohali,
Industrial Engineering Systems with the Internet of Things. Boca Raton, ‘‘A novel transfer learning approach for detection of pomegranates growth
FL, USA: CRC Press, 2024, pp. 245–259. stages,’’ IEEE Access, vol. 12, pp. 27073–27087, 2024.
[3] N. Zhang, M. Fard, J. Xu, J. L. Davy, and S. R. Robinson, ‘‘Road [24] A. Raza, K. Munir, M. S. Almutairi, and R. Sehar, ‘‘Novel transfer learning
safety: The influence of vibration frequency on driver drowsiness, reaction based deep features for diagnosis of down syndrome in children using
time, and driving performance,’’ Appl. Ergonom., vol. 114, Jan. 2024, facial images,’’ IEEE Access, vol. 12, pp. 16386–16396, 2024.
Art. no. 104148. [25] B. Abu-Salih, S. Alotaibi, R. Abukhurma, M. Almiani, and M. Aljaafari,
[4] D. Wang, Q. Liu, L. Ma, Y. Zhang, and H. Cong, ‘‘Road traffic accident ‘‘DAO-LGBM: Dual annealing optimization with light gradient boosting
severity analysis: A census-based study in China,’’ J. Saf. Res., vol. 70, machine for advocates prediction in online customer engagement,’’ Cluster
pp. 135–147, Sep. 2019. Comput., pp. 1–27, Jan. 2024.

VOLUME 12, 2024 64777

 H. A. Madni et al.: Novel Transfer Learning Approach for Driver Drowsiness Detection

[26] A. Babu and S. A. Jerome, ‘‘ICMFKC with optimize XGBoost classifica- RUKHSHANDA SEHAR received the B.Sc.
tion for breast cancer image screening and detection,’’ Multimedia Tools degree in computer science and the Master of
Appl., pp. 1–28, Jan. 2024. Computer Science (M.C.S.) degree from the
[27] M. S. Chowdhury, M. N. Rahman, M. S. Sheikh, M. A. Sayeid, Islamia University of Bahawalpur, in 2017 and
K. H. Mahmud, and B. Hafsa, ‘‘GIS-based landslide susceptibility map- 2021, respectively, and the M.S. degree in com-
ping using logistic regression, random forest and decision and regression puter science from the Khawaja Fareed University
tree models in chattogram district, Bangladesh,’’ Heliyon, vol. 10, no. 1, of Engineering and Technology (KFUEIT). Her
Jan. 2024, Art. no. e23424.
research interests include machine learning and
[28] P. Krishnamoorthy, M. Sathiyanarayanan, and H. P. Proença, ‘‘A novel
deep learning.
and secured email classification and emotion detection using hybrid deep
neural network,’’ Int. J. Cognit. Comput. Eng., vol. 5, pp. 44–57, Jul. 2024.
[29] G. Aziz, N. Minallah, A. Saeed, J. Frnda, and W. Khan, ‘‘Remote sensing
based forest cover classification using machine learning,’’ Sci. Rep.,
vol. 14, no. 1, p. 69, Jan. 2024.
[30] A. Chunduru, A. R. Kishore, B. K. Sasapu, and K. Seepana, ‘‘Multi chronic
disease prediction system using CNN and random forest,’’ Social Netw.
Comput. Sci., vol. 5, no. 1, p. 157, Jan. 2024.
[31] Y. Du, N. Pan, Z. Xu, F. Deng, Y. Shen, and H. Kang, ‘‘Pavement distress
detection and classification based on Yolo network,’’ Int. J. Pavement Eng., NISREAN THALJI received the bachelor’s and
vol. 22, no. 13, pp. 1659–1672, Nov. 2021. master’s degrees in computer science from
[32] N. Aishwarya, K. Manoj Prabhakaran, F. T. Debebe, M. S. S. A. Reddy, and Yarmouk University, Jordan, Irbid, and the
P. Pranavee, ‘‘Skin cancer diagnosis with YOLO deep neural network,’’ Ph.D. degree in computer science, specializing
Proc. Comput. Sci., vol. 220, pp. 651–658, 2023. in artificial intelligence and data science from
[33] G. Li, Z. Song, and Q. Fu, ‘‘A new method of image detection for small University Malaysia Perlis (UNIMAP), Perlis,
datasets under the framework of YOLO network,’’ in Proc. IEEE 3rd Malaysia, with an excellent academic record. She
Adv. Inf. Technol., Electron. Autom. Control Conf. (IAEAC), Oct. 2018, achieved outstanding academic performance and
pp. 1031–1035. ranked first among her colleagues for her master’s
degree. With a strong educational background,
she is currently an Assistance Professor with Jadara University. She is
an accomplished individual in the field of computer science. Her diverse
expertise spans various domains, including artificial intelligence, machine
learning, deep learning, algorithm engineering, swarm intelligence, and
natural language processing. Her research interests continue to thrive in these
cutting-edge areas, further solidifying her reputation as a dedicated scholar
and a researcher in the field.
HAMZA AHMAD MADNI received the B.S.
degree in computer system engineering from
NFCIET (BZU), Multan, Pakistan, in 2010, the
M.S. degree in electrical and electronic engineer-
ing from the University of Bradford, U.K., in 2012,
and the Ph.D. degree in electromagnetic field and
microwave technology from Zhejiang University,
China, in June 2017. He was an Assistant Professor LAITH ABUALIGAH received the degree from the
with the Khwaja Fareed University of Engineering Department of Computer Information Systems,
and Information Technology, Rahim Yar Khan, Al-Albayt University, Jordan, in 2011, the mas-
Pakistan. From December 2017 to September 2019, he held a postdoctoral ter’s degree from the Department of Computer
position at Southeast University (SEU), Nanjing, China. His current Science, Al-Albayt University, in 2014, and the
research interests include transformation optics, coding metamaterials, Ph.D. degree from the School of Computer Sci-
antennas, invisibility cloaking, and theoretical and numerical methods of ence, Universiti Sains Malaysia (USM), Malaysia,
electromagnetics, sensors, and controllers. in 2018. He is currently an Associate Professor
with the Prince Hussein Bin Abdullah College for
Information Technology, Al Al-Bayt University.
He is also a Distinguished Researcher with the School of Computer Science,
USM. He has published more than 350 journal articles and books, which
collectively have been cited more than 12500 times (H-index of 53).
His research interests include arithmetic optimization algorithm (AOA),
bio-inspired computing, nature-inspired computing, swarm intelligence,
artificial intelligence, meta-heuristic modeling, and optimization algorithms,
ALI RAZA received the Bachelor of Science evolutionary computations, information retrieval, text clustering, feature
and M.S. degrees in computer science from the selection, combinatorial problems, optimization, advanced machine learn-
Department of Computer Science, Khwaja Fareed ing, big data, and natural language processing. According to the report
University of Engineering and Information Tech- published by Clarivate, he is one of the Highly Cited Researchers, in
nology (KFUEIT), Rahim Yar Khan, Pakistan, in 2021 and 2022, and the 1% influential researchers, which depicts the 6,938
2021 and 2023, respectively. He is currently a top scientists in the world. He was the first researcher in the domain of
Lecturer with the Faculty of Information Tech- computer science in Jordan, in 2021. According to the report published
nology, Department of Software Engineering, The by Stanford University in 2020, he is one of the 2% influential scholars,
University of Lahore, Pakistan. He has published which depicts the 100,000 top scientists in the world. He serves as an
several articles in reputed journals. His current Associate Editor of Journal of Cluster Computing (Springer), the Journal
research interests include data science, artificial intelligence, data mining, of Soft Computing (Springer), and the Journal of Engineering Applications
natural language processing, machine learning, deep learning, and image of Artificial Intelligence (Elsevier).
processing.

64778 VOLUME 12, 2024