IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 4, December 2024, pp. 4654~4666

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4654-4666  4654

Enhancing image quality using super-resolution residual

network for small, blurry images

Djarot Hindarto1,2, Mochammad Iwan Wahyuddin2, Andrianingsih Andrianingsih3,

Ratih Titi Komalasari2, Endah Tri Esti Handayani3, Mochamad Hariadi1
1
Department of Electrical Engineering, Faculty of Intelligent Electrical and Informatics Technology, Institut Teknologi Sepuluh Nopember,
Surabaya, Indonesia
2
Department of Informatics, Faculty of Communication and Information Technology, Nasional of University, Jakarta, Indonesia
3
Department of Information Systems, Faculty of Communication and Information Technology, Nasional of University,
Jakarta, Indonesia

Article Info ABSTRACT

Article history: In the background, when low-resolution images are utilized, image
identification tasks are frequently hampered. By employing the residual
Received Jan 3, 2024 network super-resolution framework, super-resolution techniques are used to
Revised Feb 29, 2024 enhance image quality, specifically in the detection and identification of small
Accepted Mar 21, 2024 and blurry objects. Improving resolution, decreasing blur, and enhancing
object detail are the main goals of the suggested approach. The novelty of this
research resides in its application of the activation exponential linear unit
Keywords: (ELU) to the super-resolution residual network (SR-ResNet) framework,
which has been demonstrated to enhance image sharpness. The experimental
Image processing findings demonstrate a substantial enhancement in the quality of the images,
Image quality as evidenced by the training data's structural similarity index (SSIM) of
Peak signal-to-noise ratio 0.9989 and peak signal-to-noise ratio (PSNR) of 91.8455. Furthermore, the
Small and blurry images validation data demonstrated SSIM 0.9990 and PSNR 92.5520. The results of
Structural similarity index this study indicate that the implementation of SR-ResNet significantly
enhances the capability of the detection system to detect and classify
diminutive and opaque entities precisely. The expected and projected
enhancement in image quality significantly influences image processing,
especially in situations where accuracy and object differentiation are vital.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Mochamad Hariadi
Department of Electrical Engineering, Faculty of Intelligent Electrical and Informatics Technology
Institut Teknologi Sepuluh Nopember
Surabaya 60111, Indonesia
Email: [email protected]

1. INTRODUCTION
Advancements in image processing have greatly enhanced the ability to recognize and analyze objects.
However, the detection of small and indistinct objects must be improved. The low image resolution poses a
constraint that impedes tasks such as security monitoring, medical applications, and autonomous technology.
It becomes difficult to discern intricate details in blurry images. In order to tackle the issue at hand, it is
necessary to go beyond the achievements of traditional image processing methods and instead focus on
fostering innovation and implementing a methodical approach. Super-resolution (SR) provides a promising
solution to enhance image resolution and overcome these limitations. Nevertheless, further investigation is
required to examine the capacity of SR to identify minute and indistinct entities. The integration of SR with
the super-resolution residual network (SR-ResNet) architecture has demonstrated the potential to improve

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  4655

image recognition. The limited image resolution poses significant challenges in accurately identifying and
distinguishing objects, resulting in profound implications in domains such as medicine, security, and
autonomous technology. Hence, the quest for solutions to detect diminutive and indistinct images continues to
be a vibrant field of study.
Computer vision and image processing assess image quality. Structural similarity index (SSIM) and
peak signal-to-noise ratio (PSNR) are used. SSIM takes structure, contrast, and luminance into account in order
to quantify the degree of structural similarity between two images [1]. This analysis is more closely aligned
with human perception compared to methods that solely rely on individual pixels. The SSIM values span from
-1 to 1, with a value of 1 indicating complete similarity. PSNR quantifies the relationship between the highest
signal intensity and the presence of noise in the signal representation. A high PSNR value, measured in
decibels, signifies excellent image or video quality following procedures such as compression or transmission.
Both metrics are commonly employed in research pertaining to image compression algorithms, image
restoration, and other computer vision applications. The utilization of both SSIM and PSNR enables a thorough
assessment of image quality, taking into account both perceptual and technical factors. Ensuring the delivery
of superior outcomes that align with user specifications is paramount in image processing applications.
This research focuses on the SSIM value aimed at measuring the effectiveness of image processing
and reconstruction compared to the original or reference image. SSIM is used as the primary metric due to its
sensitivity to structural changes recognized by human visual perception. This research also seeks to improve
image processing techniques to achieve higher structural similarity with the original image. This is very
important in image restoration and compression, as well as in security monitoring, medical applications, and
autonomous technology. Experiments conducted on training data show notable improvements in image quality
evaluation using the PSNR metric of 91.8455 and SSIM of 0.9990. These results show that the proposed
approach enhances the system's ability to improve image resolution and effectively identify small and blurry
objects in the training data environment. In addition, when conducting experiments at 4x scale, the use of
identical evaluation metrics showed an increase in PSNR value to 92.5520 and SSIM to 0.9990 on the
validation data. This validates that the suggested methodology not only achieves improved image quality but
also consistently improves the system's ability to detect small objects at high image resolutions.
The primary contribution of this research is the successful integration of SR and object detection
techniques. This integration has led to improvements in image resolution and increased accuracy in
distinguishing small and blurry objects in images. The obtained results demonstrate that the integration of
SR-ResNet yields substantial enhancements in image quality [2] and the system's capacity to identify objects
that are challenging to discern in low-resolution images. These findings are significant because they can
improve visual acuity in various image-processing applications that require accurate object detection, such as
in the fields of medicine, security, and autonomous technology. Through enhancements in image resolution
and enhanced detection capabilities for minute and indistinct objects, this method could pave the way for more
advanced image processing capabilities, enabling more precise identification of objects.
Prior research, as enumerated in Table 1 [3]-[16], tends to prioritize the PSNR value over the SSIM
value as the primary quality indicator. SSIM, an algorithm that assesses the similarity of two images by
analysing their brightness, contrast, and structure, is deemed highly pertinent; however, it has yet to be
exhaustively examined within the scope of this investigation. Hence, to achieve a more comprehensive
understanding, this study incorporates an assessment of the SSIM value on the reconstructed and enhanced
image, in addition to the PSNR value. Additionally, findings from prior research indicate that investigations
employing a 2x image scale exhibit a more pronounced augmentation in both PSNR and SSIM. As a result,
this study introduces an innovative approach by utilizing a fourfold increase in image scale. The impetus and
motivation for this research stem from the identification of gaps in prior studies. By integrating PSNR and
SSIM assessments and broadening the research scope to include images at a 4x scale, this study aims to offer
a more comprehensive understanding of methods for enhancing image resolution and reconstruction quality.
This method will likely generate a substantial contribution to our knowledge of image resolution enhancement
technologies and close a gap in the literature.
The novelty of this research lies in the significant improvement in SSIM values achieved through the use
of SR technology integrated with the SR-ResNet architecture. The main focus is on improving the system's ability
to identify fine details and distinguish small objects in images, which is reflected explicitly in the improved SSIM
values. This innovation marks an essential advancement in image processing, particularly in improving the
structural similarity of images, which is vital in applications that require accurate and high-detail object recognition.

2. METHOD
2.1. Proposed method
Figure 1 illustrates the proposed research methodology for enhancing image resolution. The resolution
enhancement process is accomplished by employing the SR-ResNet model [17], [18], uses two datasets:
Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4656  ISSN: 2252-8938

low-resolution and high-resolution. The ResNet architecture, which has been adopted, comprises fundamental
layers, including convolution layers (conv), batch normalization, activation functions utilizing exponential
linear unit (ELU) [19], and weight addition. The model trains the network to recognize patterns and
characteristics in low-resolution using a dataset. Upon completion of the training process, the model is
anticipated to possess the capability to comprehend the correlation between low-resolution and high-resolution
images. During this stage, the SR-ResNet model endeavors to generate high-resolution images from
low-resolution images. This approach entails an intricate procedure wherein the model acquires the ability to
comprehend intricate elements, formations, and patterns in low-resolution images, subsequently endeavoring
to reconstruct more distinct and elaborate high-resolution image.

Table 1. Research discussing super resolution

Author Algorithm Dataset Findings
[3] DMN, RDN MRI DMN-based eliminates false details in natural image tasks and improves
quantitative and qualitative. 0.62 DMN+RDN PNSR 22.78 SSIM.
[4] SR-radial fluctuations Microscopic A SRRF method reduces for large raw images, visualizes microtubule
(SRRF) images dynamics used in biomedicine.
[5] SR-CNN, PCA, Image Milling image Enhanced micro milling double dictionary. SR method outperforms others.
feature Mean PSNR: 43.62, 42.19, 43.20, 4
[6] SRCNN BSD100, The MBFSR method, extracts deep features using multiple feature extraction
Urban100 modules to create a high-resolution feature map, PNSR: 22.73, SSIM: 0.7427
[7] Two-branch crisscross BSD100, Two-branch crisscross generative adversarial network (TBCGAN) achieves
generative adversarial Urban100 realistic and accurate SR images. PSNR: 24.20, SSIM: 0.7518
network (TBCGAN)
[8] L2 gradient loss LPIPS Set5, Set4, Proposed an image SR algorithm based second-order gradient (SG) loss.
and FID Urban100 SSIM 24.98
[9] SR–cryoCLEM cryo-electron Effect of high-power laser illumination on cryosamples by apoferritin
tomography structure.
[10] SRGAN, SRRESNET Set14, BSD100 SSIM SRGAN = 0.7397, SSIM SR_RESNET = 0.8184
[11] ESRGAN ovine (Malta) SSIM ESRGAN = 0.7840
brucellosis
[12] ISFMFMNe COCO2017 SSIM ISFMFMNe = 0.8104
[13] DRMSFFN SET14, B100 DRMSFFN AVG SSIM = 0.83
[14] RCAN-SG SET4, SET5, Efficiency comparisons of upscale 4X RCAN-SG SSIM 0.8175
Urban100
[15] DBAGAN SRLP dataset DBAGAN, SSIM 0.824
[16] USRKDN Set5, Set14, USRKDN SSIM = 0.7990
BSD100
Author SR-RESNET Div2K Proposed an image SR-RESNET algorithm LR=0.000001, Optimisation
Adam, activation ELU. SSIM = 0.9989, PSNR = 91.8455

Figure 1. Proposed research method

The model uses filters to extract essential features from the imagery that is not high-quality during
convolution. The ELU activation function aids the network in learning more intricate information [20]. Batch
normalization is responsible for standardizing the output of the preceding layer, thereby accelerating the
training procedure and enhancing the stability of the network. This process is iterated by incorporating weights

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666

Int J Artif Intell ISSN: 2252-8938  4657

to fine-tune the network's weights, allowing them to gradually capture the correlation between low- and high-
resolution. The outcome of this procedure is anticipated to be an enhanced high-resolution image derived from
a low-resolution image. Figure 1 illustrates the iterative process of training the network using a low-resolution
dataset, optimizing parameters, and testing the model to generate an image that is sharper and more closely
resembles the high-resolution image as a whole. This method provides a robust approach to enhance the quality
of low-resolution images, resulting in higher-resolution images that closely resemble the original image.
Table 2 explains input for the Res-Net algorithm used in SR-ResNet is an RGB image with 256×256
pixels. After the image is convolution with 64 filters using a 9×9 kernel size, the leaky rectified linear unit
(LReLU) activation is applied to the convolution results. The output of the input image's convolution is the SR
variable, which needs to be initialized first. Each iteration entailed two steps: first, the SR was subjected to the
residual block process; second, the sum was computed between the SR and the residual block. This process
was reiterated twice. Afterward, SR underwent convolution with 64 filters, 9×9 kernels, and the same padding.
The process was repeated with 64 filters, 3×3 kernels, and the same padding. After that, the output of the
convolution is superimposed onto the original picture. We rerun the same loop, this time applying convolution
with 64 filters, 3×3 kernels, and the "same" padding to SR. Right after that, the SR undergoes a max pooling
operation, which is followed by an up-sampling operation. The last step is to perform back convolution on SR
using a 3×3 kernel, 64 filters, and the "same" padding as before. Finally, an image with three color channels
created by ELU a during the convolution stage, the model applies filters to the image to extract significant
features from the low-resolution image. The ELU activation function aids the network in learning more intricate
information. Batch normalization is responsible for standardizing the output of the preceding layer, thereby
accelerating the training procedure, and enhancing the stability of the network. This process is iterated by
incorporating weights to fine-tune the network's weights, enabling them to progressively capture the correlation
between low-resolution and high-resolution images. The outcome of this procedure is anticipated to be an
enhanced high-resolution image derived from a low-resolution image. Figure 1 illustrates the iterative process
of training the network using a low-resolution dataset, optimizing parameters, and testing the model to generate
an image that is sharper and more closely resembles the high-resolution image. This method provides a robust
approach to enhance the quality of low-resolution images, resulting in higher-resolution images that closely
resemble the original image.

Table 2. Algorithm SR-ResNet

SR-ResNet
Input: image size 256×256, RGB
Operation:
image convolution, with filter 64, kernel 9×9, padding 'same'
ELU activation in convolution results
Define Model:
Initialize the SR variable, the result of convolution of the input image
Loop 2 times:
a. Residual Block to SR.
b. the sum between SR and Residual Block.
c. 64 convolution filters, 9×9 kernels, and 'same' padding on SR.
d. 64 filter convolution, 3×3 kernel, and 'same' padding on SR.
e. the summation between the SR convolution results and the input image.
Loop 1 time:
a. 64 filter convolution, 3×3 kernel, and 'same' padding on SR.
b. Max Pooling on SR.
c. up sampling operation on SR.
d. convolution with 64 filters, 3×3 kernels, and 'same' padding on SR.
Output: 3 color channel image using ELU activation and 'same' padding.

2.2. Dataset
Table 3 is a dataset containing numerical data, text, images, and other data for analysis and modeling
[21]. Train, evaluate, and test machine learning models using training, validation, and test datasets. Dataset for
image and video restoration (Div2K) is a collection of data intended for evaluation and development purposes
in the field of image restoration [22]. The dataset is widely recognized for its exceptional quality and extensive
assortment of images. Div2K comprises a collection of 1,760 low-resolution and high-resolution images
captured under diverse conditions, such as low-light conditions, with a range of textures and levels of detail.
Every image in this dataset has a high resolution, with the majority having a resolution of 2K (hence the
dataset's name, Div2K). Due to its exceptionally high-resolution, it is a preferred option among researchers
who are primarily concerned with image resolution enhancement or restoration techniques. Div2K empowers

Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4658  ISSN: 2252-8938

users to conduct a comprehensive assessment of image processing algorithms and methodologies, with a
particular focus on image restoration and SR techniques. By utilizing the variety of images supplied by Div2K,
users can evaluate the efficacy of their algorithms across a spectrum of realistic scenarios. Div2K has emerged
as a highly valuable dataset in the field of image processing research, particularly in the domains of image
restoration and resolution enhancement. Its extensive collection of high-quality images has significantly
contributed to the advancement of image processing technology.

Table 3. Dataset Div2K composition

Dataset Number of datasets Literature
Training high-resolution 685 [23]
Training low-resolution 685 [24]
Validation high-resolution 170 [25]
Validation low-resolution 170 [26]
Testing low-resolution 50 [27]

There are several low-resolution image datasets available for free on the internet. Here are some links
that can be used to download low resolution image datasets: Div2k-dataset - A dataset consisting of
low-resolution and high-resolution images used for object detection and segmentation tasks. In this research,
the author collects datasets by mixing datasets from several image providers for free, so that the image datasets
become diverse, so that later you can find out the strengths and weaknesses if the dataset is processed using
the SR method. This research uses the SR-ResNet.

2.3. Super-resolution
SR is a technology to increase the resolution of images or videos from low resolution to high
resolution by utilizing image processing techniques and machine learning [28]. SR technology is very useful
when high resolution images or videos are needed [29], but only low-resolution images or videos are available.
The super resolution approach can be done in two ways: traditional methods and deep learning methods.
Traditional methods usually use interpolation techniques or signal processing to increase the resolution of
images or videos. The deep learning method uses a neural network model to make predictions on the missing
parts of an image or video and produce a higher resolution image or video [30]. There are several techniques
used in super resolution, including,
− Interpolation: an interpolation technique [31] is used to increase the resolution of an image by adding
pixels between the pixels already in the image. However, interpolation techniques often produce images
or videos with poor quality, especially if the magnification is very large.
− Upscaling: an upscaling or upscaling technique is a technique used to increase the resolution of an image
or video by doubling the pixels that are already in the image or video [32]. Upscaling techniques usually
produce better quality images or videos than interpolation techniques [33].
SR is a technique for enhancing image quality by improving resolution and visual sharpness. This
technology is applied in various fields, including photography, video streaming, medical image processing,
and the movie industry. In photography, SR allows images taken at a low resolution to be enhanced to become
more vivid and detailed. In video streaming, this technology helps to improve video quality by reducing blur
and increasing sharpness. In medical image processing, SR is used to enhance medical images, enabling more
accurate diagnosis. In the movie industry, SR is essential for improving the resolution of images or videos so
that critical visual details can be seen more sharply. They are providing a more immersive and satisfying
viewing experience.
Upscaling is a technique that enhances the resolution of an image or video by adding additional pixels.
This method is beneficial for improving the visual appearance of low-resolution media [34]. Various
algorithms, such as nearest neighbor, bilinear, bicubic, and lanczos, are employed in the upscaling process to
improve the details of an image or video. Upscaling can increase resolution but may only sometimes enhance
visual quality significantly and can appear poor if magnification is excessive. To address this issue, upscaling
methods are frequently integrated with additional image processing techniques like denoising, deblurring, and
deep learning to enhance the quality of images or videos by producing higher resolution and visually pleasing
results. A more detailed mathematical formula for upscaling is as follows:
Let's say I_low is a low-resolution image, and I_high is the high-resolution image you want to
produce. Upscaling is done by multiplying the I_low matrix by the T transformation matrix to get the I_high
matrix:

I_high (x, y) = T * I_low (x, y) (1)

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666

Int J Artif Intell ISSN: 2252-8938  4659

Where (x, y) are the pixel coordinates in the image, and T is the transformation matrix. The transformation
matrix T is found through different interpolation techniques, such as nearest neighbour, bilinear, bicubic, and
lanczos.

2.4. Exponential linear unit

Neuronal networks are activated with the ELU [35]. ELUs [36], [37] unlike rectified linear unit (ReLUs),
have negative values, pushing mean unit activations closer to zero, like batch normalization, with less
computational complexity. A reduced bias shift effect brings the average gradient closer to the unit's natural
gradient, promoting zero-speed learning. However, negative LReLUs and parametrized rectified linear unit
(PReLUs) values do not guarantee noise-robust deactivation. Smaller inputs saturate ELUs to a negative value,
reducing forward propagated variation and information. The ELU aids deep neural network learning and improves
classification accuracy. ELUs mitigate the vanishing gradient problem like ReLUs, LReLUs, and PReLUs by
identifying positive values. ELUs learn better than other activation functions. Formula ELU with 0 < α is:

𝑎, 𝑖𝑓 𝑎 > 0
k(a) = { (2)
𝛼(exp(𝑎) − 1), 𝑖𝑓 𝑎 ≤ 0

1, 𝑖𝑓 𝑎 > 0
k′(a) = { (3)
𝑘(𝑎) + 𝛼, 𝑖𝑓 𝑎 ≤ 0

3. RESULTS AND DISCUSSION

3.1. Super-resolution residual network
This research aimed to conduct training experiments to enhance image resolution by employing the
super resolution technique with a 4X scale. Upscaling is the process of improving the resolution of a
low-resolution image to four times its original resolution. This experiment entails a training process of a SR
model, which enables the conversion of low-resolution images into higher-resolution images. The objective of
this experiment is to enhance the clarity, intricacy, and overall excellence of low-resolution images,
transforming them into more transparent images with a higher resolution. By implementing a sequence of
training procedures at a scale four times larger than the original, the super resolution model, specifically
SR-ResNet, is educated to discern the underlying patterns and structures present in low-resolution images.
Consequently, it generates images with a resolution that is four times greater than the initial input
image. This procedure entails iteratively optimizing the artificial neural network to enhance the model's
capacity to estimate and restore intricate details that are diminished in low-resolution imagery. The utilization
of 4X scaling in this experiment showcases the capability of the super resolution model to execute more
extensive transformations, thereby posing additional difficulties in achieving precise reconstruction and
enhancing the overall quality of the image. This research seeks to investigate the capabilities of the super
resolution technique in addressing the problems associated with enhancing image resolution on a larger scale.
It is anticipated that this technique will yield more satisfactory outcomes in the restoration of high-resolution
images from low-resolution images. During the training process, the use of SR-ResNet to enhance image
resolution from low to high has resulted in a collection of PSNR and SSIM values at each epoch. The outcomes
of these ten epochs demonstrate substantial enhancements in the caliber of image reconstruction. In the
beginning, during the first epoch, the PSNR and SSIM values were 21.2507 and 0.6619, respectively. At a low
PSNR, the reconstructed image is significantly different from the original, indicating poor image
reconstruction. Nevertheless, as the training iterations advanced, there was a steady enhancement in both
evaluation metrics.
The PSNR graph in Figure 2 demonstrates a steady upward trend from the initial epoch to the fifth
epoch, culminating in the highest value of 24.8211 in the fifth epoch. This signifies a notable enhancement in
the resemblance difference between the original and reconstructed image. However, during the sixth epoch,
while it continued to be at a relatively high level, there was a slight reduction in the PSNR value to 24.9382.
Furthermore, the SSIM graph exhibits a comparable pattern. Commencing at a value of 0.6619 in the initial
epoch, the SSIM progressively rises until it reaches its zenith in the tenth epoch, attaining a value of 0.8255.
This signifies that the resemblance between the reconstructed image and the original structure, brightness, and
contrast is approaching a state of near-perfect similarity.
However, it is essential to note that there are variations in some periods, exemplified by the seventh
period, where the PSNR value decreased substantially to 22.1056. This shows that there may be fluctuations
in the results due to the complex nature of the images or specific attributes of the training procedure. In
Figure 3, graphs depicting the fluctuations in PSNR and SSIM values at each epoch offer a more in-depth
understanding of the model performance. The image reconstruction improved from low resolution to high

Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4660  ISSN: 2252-8938

resolution as the SR-ResNet training progressed, despite occasional fluctuations. This demonstrates the
capacity of this technique to enhance image resolution significantly.

Figure 2. Training model SSIM and PSNR

Figure 3. Training model SSIM and PSNR

For ten epochs, Adam's optimizer trained the model at 0.00001, improving image quality metrics like
SSIM and PSNR. This improvement was seen in training and validation datasets. The SSIM, which measures
structural similarity between images, increased from 0.8371 to 0.8474 in the training dataset and 0.8445 to
0.8543 in the validation dataset. This indicates that the model is becoming more effective at maintaining the
visual integrity of the original images as the training progresses. The PSNR, a metric that quantifies the noise
level in the reconstructed image in relation to the original image, demonstrates notable enhancements. In the
training data, it increased from 25.2095 dB to 25.5490 dB, and in the validation data, it rose from 25.7978 dB
to 26.2971 dB. This rise suggests a reduction in the image reconstruction error rate, leading to an improvement
in the image reconstruction quality, approaching that of the original image.
The research proves that Adam's optimizer is helpful for training and validating picture reconstruction
models. PSNR and SSIM were two measures of picture quality that the model consistently improved. On the
training dataset, the model's improvement went from 91.8385 dB in the first epoch to 91.8455 dB in the tenth.
On the validation dataset, it went from 92.5448 dB to 92.5520 dB, showing a steady improvement in
reconstruction ability. SSIM, an essential measure of how healthy images are perceived, stayed high all the
way through validation and training, rising slightly from 0.9990 in the first epoch to 0.9990 in the tenth. The
model's consistent performance on hidden data suggests that it is balanced, as evidenced by its low variance
and good convergence during training and validation. The modest gains in PSNR and SSIM, however, point
to either untapped potential or a learning plateau with the provided data. The cautious conditions to prevent

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666

Int J Artif Intell ISSN: 2252-8938  4661

overshooting gradient descent, which could slow down convergence, could be indicated by the low learning
rate. With a stable SSIM value at a high threshold and a consistent improvement in PSNR, the final training
results demonstrate excellent performance. Additional trials with varying learning rates are being contemplated
to evaluate the model's efficacy without lowering the bar for generalizability or reconstruction quality. The
explanation above can be seen in Figure 4.

Figure 4. Training model SSIM and PSNR

Figure 5 shows tested the super resolution model, two metrics assessed image quality after
reconstruction: SSIM and PSNR. We employed a variety of low-resolution images for this purpose. The first
step in the testing process involves processing a red car image three times. The PSNR value increased
consistently in both the first and second tests, reaching 22.14 dB and 23.35 dB, respectively, with an SSIM
value of approximately 0.82. This proves that the model can enhance red car images with higher resolution.
Nevertheless, despite a rise in PSNR to 23.35 dB in the third test using a blue car image, SSIM decreased.
When an image's structure is altered, the reconstructed and original versions will look different from a structural
perspective. A drop in PSNR to 23.35 dB and an increase in SSIM to 0.83 were observed in the most recent
test on the white car image. Although the PSNR has decreased, the structural similarity between the original
and reconstructed images has significantly improved.

Figure 5. Testing low resolution to high resolution

The results show that the SR model's performance varies with the image's characteristics. Increasing
PSNR does not necessarily ensure better structural similarity despite PSNR's widespread use as a critical
Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4662  ISSN: 2252-8938

metric. To evaluate changes in reconstructed images, it is essential to have a better grasp of structural
similarities; SSIM evaluation offers this. Because it paints a more complete picture of structural similarity—
which may differ from PSNR values-SSIM should be prioritized as an evaluation metric. Regardless of the
PSNR value, the key to comprehending the potential structural changes in the reconstructed image lies in the
application of SSIM in SR evaluation. To guarantee consistently improved image quality, it is crucial to
evaluate SR performance holistically. This research demonstrates that when SSIM is used in conjunction with
PSNR, a more comprehensive and accurate picture of the reconstructed image's quality can be obtained. In this
comprehensive analysis, the structural similarity measured by SSIM values is also considered so that the rise
in PSNR does not solely dictate the increase in picture resolution. This lends credence to the idea that
developing super resolution techniques necessitates a thorough evaluation to attain more consistent and
satisfying outcomes.

3.2. Discussion, limitations and future work

In this research experiment, a Core I7 laptop featuring 16 GB of RAM and a 1 TB hard drive disk was
used to carry out the SR-ResNet experiment. Opencv, tqdm, tensorflow, and numpy libraries, in addition to the
python programming language in operating system Windows 10. Necessary library packages used in
computational mathematics and data science include numpy, tqdm, tensorflow, and opencv. Constantly, twelve
hours are devoted to training. The SR-ResNet model is generated through this procedure.

3.2.1. Dataset analysis

Div2K is a dataset that includes five primary subsets, one for each of the three main types of data:
training, validation, and testing. The goal of using 685 high- and low-resolution images pairs in the training
data is to help the artificial neural network grasp the connection between the two types of graphics. As a result,
this aids the training process for transforming low-resolution images into high-resolution ones that are both
clear and detailed. resolution for the sake of fair evaluation of the model's performance and to guarantee that it
can generalize to new data as well as the training data, verification like this is crucial. At the same time, 50
low-resolution images make up the testing portion of this dataset; they are utilized to assess the model's
robustness on novel data. Researchers can test the efficacy of restoring images of high resolution from
low-resolution ones. The Div2K dataset offers enough variety in terms of both composition and subset size to
conduct comprehensive and trustworthy training, validation, and testing of algorithms and models for the task
of solving the challenge of re-creating high-resolution images from low-resolution ones.

3.2.2. Process training super-resolution residual network

The results of the training process using the SR-ResNet for ten epochs or iterations are shown in
Tables 4 to 6. Every row in the table represents a training epoch and a model performance metric. The "epoch"
column displays the precise epoch number during the training process. Additionally, multiple columns
encompass evaluation metrics, including "loss," which represents the loss function value computed by the
model during each epoch. This loss function measures the model's predictive accuracy in determining the target
value.
Additionally, there is a column labeled "Acc," which indicates the accuracy of the model at each
epoch. Accuracy measures the level of precision exhibited by the model in its predictions of targets. Following
that, there is the PSNR column, which serves as a metric for evaluating the quality of image reconstruction,
specifically in the context of high-level image restoration. PSNR quantifies the level of similarity between the
reconstructed image and the original image in terms of both clarity and accuracy. A column called SSIM
measures the similarity of the reconstructed image's structure or texture to the original. The column labeled
"val" represents the validation process and displays the same metrics as loss, acc, PSNR, and SSIM, which are
measured on validation data. Validation data is data that is not utilized during the training process to prevent
overfitting. In general, this table presents a summary of how the model's performance changes over time as it
undergoes training. As time passes, it is evident that the loss value decreases, while accuracy, PSNR, and SSIM
show an upward trend in both training and validation data. This suggests the model can reconstruct
high-resolution from low-resolution ones better. Despite occasional fluctuations in loss values during specific
periods, the evaluation metric values on the validation data consistently demonstrated improvement, suggesting
the model's strong ability to generalize.

3.2.3. Limitations
The SR-ResNet method, 4X image scale experiments, and ELU activation in this research have several
drawbacks. With exploring lower scales like 2X or 3X, 4X experiments are unlimited. Experiments at a specific
scale may lose important insights into SR-ResNet method performance and generalization at lower resolutions.
Consider multiple scales in experiments to better understand how the method adapts to resolution changes.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666

Int J Artif Intell ISSN: 2252-8938  4663

This limitation may limit SR-ResNet's applicability in contexts with different resolution requirements. In
addition, ELU activation, while helpful in vanishing gradient problems, can limit you if misused, especially in
situations where other activation methods are better.

3.2.4. Future work

Given the limitations of the current study, extending the experiments to different resolution scales
would help determine the SR-ResNet method's efficacy. In addition, investigating various activation function
options may help better understand the complexity of this model. Although this research has provided a solid
foundation, there is an urgent need to conduct a more comprehensive follow-up study. The main focus of future
research should be on analyzing the broader resolution context and parameter variations in SR-ResNet. By
addressing these limitations, it is expected that the SR-ResNet method can be further optimized for various SR
applications. This follow up research could improve image resolution in photography, film, and medical image
processing.

Table 4. Process training with optimizer=Adam (learning_rate=0.0001)

Epoch Loss Acc PSNR SSIM Val loss Val Acc Val PSNR Val SSIM
1 0.0094 0.7109 21.2507 0.6619 0.0053 0.7660 23.6326 0.7566
2 0.0059 0.7762 23.2355 0.7667 0.0043 0.8075 24.6491 0.7964
3 0.0050 0.8032 23.9899 0.7983 0.0037 0.8462 25.3208 0.8221
4 0.0047 0.8074 24.2967 0.8130 0.0037 0.8575 25.3848 0.8317
5 0.0042 0.8316 24.8211 0.8277 0.0034 0.8600 25.8173 0.8359
6 0.0041 0.8344 24.9382 0.8331 0.0032 0.8653 26.0199 0.8460
7 0.0193 0.7401 22.1056 0.7372 0.0047 0.7834 24.0669 0.7709
8 0.0049 0.8021 23.9617 0.7938 0.0038 0.8255 25.1737 0.8155
9 0.0043 0.8232 24.6017 0.8159 0.0037 0.8112 25.1761 0.8244
10 0.0042 0.8338 24.8172 0.8255 0.0036 0.8287 25.3313 0.8334

Table 5. Process training with optimizer=Adam (learning_rate=0.00001)

Epoch Loss Acc PSNR SSIM Val loss Val Acc Val PSNR Val SSIM
1 0.0039 0.8515 25.2095 0.8371 0.0034 0.8243 25.7978 0.8445
2 0.0039 0.8545 25.2795 0.8389 0.0033 0.8703 25.9431 0.8460
3 0.0038 0.8558 25.3142 0.8406 0.0032 0.8705 26.0831 0.8462
4 0.0038 0.8579 25.3580 0.8415 0.0032 0.8363 26.0554 0.8486
5 0.0038 0.8511 25.2991 0.8417 0.0032 0.8606 26.1499 0.8491
6 0.0037 0.8624 25.4504 0.8434 0.0032 0.8637 26.1724 0.8503
7 0.0037 0.8630 25.4675 0.8448 0.0031 0.8776 26.2104 0.8514
8 0.0037 0.8617 25.4825 0.8457 0.0031 0.8735 26.2445 0.8527
9 0.0037 0.8572 25.4916 0.8463 0.0031 0.8763 26.2723 0.8528
10 0.0037 0.8631 25.5490 0.8474 0.0031 0.8688 26.2971 0.8543

Table 6. Process training with optimizer=Adam (learning_rate=0.000001)

Epoch Loss Acc PSNR SSIM Val loss Val Acc Val PSNR Val SSIM
1 0.0036 0.8694 91.8385 0.9989 0.0031 0.8746 92.5448 0.9990
2 0.0036 0.8692 91.8393 0.9989 0.0031 0.8744 92.5459 0.9990
3 0.0036 0.8692 91.8400 0.9989 0.0031 0.8757 92.5464 0.9990
4 0.0036 0.8694 91.8409 0.9989 0.0031 0.8745 92.5474 0.9990
5 0.0036 0.8696 91.8417 0.9989 0.0031 0.8746 92.5485 0.9990
6 0.0036 0.8692 91.8424 0.9989 0.0031 0.8750 92.5492 0.9990
7 0.0036 0.8693 91.8432 0.9989 0.0031 0.8744 92.5495 0.9990
8 0.0036 0.8694 91.8440 0.9989 0.0031 0.8747 92.5505 0.9990
9 0.0036 0.8695 91.8447 0.9989 0.0031 0.8752 92.5513 0.9990
10 0.0036 0.8694 91.8455 0.9989 0.0031 0.8752 92.5520 0.9990

4. CONCLUSION
The SR-ResNet model has successfully improved image resolution by applying SR techniques at a
4X scale. It effectively identifies patterns and structures in low-resolution images, resulting in images with four
times higher resolution than the originals. However, challenges remain in achieving accurate reconstruction
and improving image quality. The model showed significant improvement in image quality during training,
with peak values of PSNR and SSIM at the fifth epoch. However, results varied over the training period,
suggesting that image characteristics and certain training elements can impact the model's performance.
Evaluation of SSIM is essential for a more comprehensive assessment of image quality. The study highlights
the need for a comprehensive approach when developing SR techniques. Future research should explore the

Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4664  ISSN: 2252-8938

use of the SRThis research has successfully demonstrated that the SR-ResNet model is capable of significantly
improving the resolution of images through the application of SR techniques at a scale of 4 times. The model
shows a unique ability to identify patterns and structures in low-resolution images and produce four times
higher quality output. This represents a significant advance in efforts to improve image resolution. It lays the
groundwork for broad practical applications in visual quality enhancement. In this study, the main findings are
the significant increase in PSNR and SSIM values achieved by the SR-ResNet model, trial 1, learning rate
0,0001, ADAM optimizing, 5th epoch, PSNR=24,9382, SSIM=0,8331. Experiment 2, learning rate 0.00001,
ADAM optimization, tenth epoch, PSNR=25.5490, SSIM=0.8474. Experiment 3, learning rate 0.000001,
ADAM optimization, tenth epoch, PSNR=91.8455, SSIM=0.9989. This indicates that the model has great
potential to produce images with much better quality than the original. However, the variability of the results
over the training period indicates the significant influence of image characteristics and training parameters on
the performance of the model. In this context, SSIM evaluation becomes very important to obtain a more
accurate assessment of image quality. The results of this research have significant implications for the theory
and practice of information technology and image processing. By demonstrating the effectiveness of the SR
technique implemented by the SR-ResNet model, this research paves the way for the development of more
advanced image resolution enhancement methods. This not only enriches the theory in the field of image
processing but also provides practical solutions for applications that require image quality enhancement, such
as medicine, security surveillance, and entertainment. This research has limitations, especially in terms of the
variation in results caused by image characteristics and training elements. This suggests the need for further
research to optimize the model by taking these factors into account. Future research can explore the
applicability of the SR-ResNet model on a larger scale and different image types and seek the optimal balance
between PSNR and SSIM improvement. A deeper investigation of the influence of image characteristics on
model performance is expected to provide a deeper and more comprehensive understanding of SR techniques.
This research provides new and essential insights into the potential and challenges of developing SR
techniques. It provides a window of opportunity for future studies that push the boundaries of understanding
and provide valuable solutions in the domains of computer vision and data processing.-ResNet model with
different scaling scales and image types, optimize the balance between PSNR and SSIM enhancement, and
investigate the influence of image characteristics on model performance. This will contribute to more advanced
and versatile SR techniques for various applications.

ACKNOWLEDGMENTS
Part of the funding for this project came from the Research and Community Service (PPM) Grant
049/SP3K/Ka. Admin Bureau. PPM / X / 2023 at Nasional University of Jakarta.

REFERENCES
[1] I. I. -Azcarate, R. D. Acemel, J. J. Tena, I. Maeso, J. L. G. -Skarmeta, and D. P. Devos, “4Cin: A computational pipeline for 3D
genome modeling and virtual Hi-C analyses from 4C data,” PLoS Computational Biology, vol. 14, no. 3, 2018, doi:
10.1371/journal.pcbi.1006030.
[2] Z. Lu, X. Jiang, and A. Kot, “Deep coupled ResNet for low-resolution face recognition,” IEEE Signal Processing Letters, vol. 25,
no. 4, pp. 526–530, 2018, doi: 10.1109/LSP.2018.2810121.
[3] X. Liu, H. Guo, H. Liu, and J. Li, “Domain migration representation learning for blind magnetic resonance image super-resolution,”
Biomedical Signal Processing and Control, vol. 86, 2023, doi: 10.1016/j.bspc.2023.105357.
[4] J. Chen et al., “Deep-learning accelerated super-resolution radial fluctuations (SRRF) enables real-time live cell imaging,” Optics
and Lasers in Engineering, vol. 172, 2024, doi: 10.1016/j.optlaseng.2023.107840.
[5] S. Li, Z. Ling, and K. Zhu, “Image super resolution by double dictionary learning and its application to tool wear monitoring in
micro milling,” Mechanical Systems and Signal Processing, vol. 206, 2024, doi: 10.1016/j.ymssp.2023.110917.
[6] D. Li, S. Yang, X. Wang, Y. Qin, and H. Zhang, “Multi-branch-feature fusion super-resolution network,” Digital Signal Processing:
A Review Journal, vol. 145, 2024, doi: 10.1016/j.dsp.2023.104332.
[7] Q. Yang, Y. Liu, and J. Yang, “Two-branch crisscross network for realistic and accurate image super-resolution,” Displays, vol. 80,
2023, doi: 10.1016/j.displa.2023.102549.
[8] S. Lin, C. Zhang, and Y. Yang, “A pluggable single-image super-resolution algorithm based on second-order gradient loss,”
BenchCouncil Transactions on Benchmarks, Standards and Evaluations, vol. 3, no. 4, 2023, doi: 10.1016/j.tbench.2023.100148.
[9] M. G. F. Last, W. E. M. Noteborn, L. M. Voortman, and T. H. Sharp, “Super-resolution fluorescence imaging of cryosamples does
not limit achievable resolution in cryoEM,” Journal of Structural Biology, vol. 215, no. 4, 2023, doi: 10.1016/j.jsb.2023.108040.
[10] C. Ledig et al., “Photo-realistic single image super-resolution using a generative adversarial network,” Proceedings - 30th IEEE
Conference on Computer Vision and Pattern Recognition, CVPR 2017, vol. 2017, pp. 105–114, 2017, doi: 10.1109/CVPR.2017.19.
[11] X. Wang, “An epidemiologic study on brucellosis in the vicinity of hohhot in China,” American Journal of Biomedical and Life
Sciences, vol. 6, no. 3, 2018, doi: 10.11648/j.ajbls.20180603.15.
[12] L. Fu, H. Jiang, H. Wu, S. Yan, J. Wang, and D. Wang, “Image super-resolution reconstruction based on instance spatial feature
modulation and feedback mechanism,” Applied Intelligence, vol. 53, no. 1, pp. 601–615, 2023, doi: 10.1007/s10489-022-03625-x.
[13] F. Liu, X. Yang, and B. De Baets, “A deep recursive multi-scale feature fusion network for image super-resolution,” Journal of
Visual Communication and Image Representation, vol. 90, 2023, doi: 10.1016/j.jvcir.2022.103730.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666

Int J Artif Intell ISSN: 2252-8938  4665

[14] W. Ying, T. Dong, and J. Fan, “An efficient multi-scale learning method for image super-resolution networks,” Neural Networks,
vol. 169, pp. 120–133, 2024, doi: 10.1016/j.neunet.2023.10.015.
[15] S. Pan, S. B. Chen, and B. Luo, “A super-resolution-based license plate recognition method for remote surveillance,” Journal of
Visual Communication and Image Representation, vol. 94, 2023, doi: 10.1016/j.jvcir.2023.103844.
[16] N. Yuan, B. Sun, and X. Zheng, “Unsupervised real image super-resolution via knowledge distillation network,” Computer Vision
and Image Understanding, vol. 234, 2023, doi: 10.1016/j.cviu.2023.103736.
[17] M. Duan et al., “Learning a deep ResNet for SAR image super-resolution,” 2021 SAR in Big Data Era, BIGSARDATA 2021 -
Proceedings, 2021, doi: 10.1109/BIGSARDATA53212.2021.9574228.
[18] W. S. Jeon and S. Y. Rhee, “Single image super resolution using residual learning,” 2019 International Conference on Fuzzy Theory
and Its Applications, iFUZZY 2019, pp. 310–313, 2019, doi: 10.1109/iFUZZY46984.2019.9066214.
[19] D. Kim, J. Kim, and J. Kim, “Elastic exponential linear units for convolutional neural networks,” Neurocomputing, vol. 406, pp.
253–266, 2020, doi: 10.1016/j.neucom.2020.03.051.
[20] Y. Li, C. Fan, Y. Li, Q. Wu, and Y. Ming, “Improving deep neural network with multiple parametric exponential linear units,”
Neurocomputing, vol. 301, pp. 11–24, 2018, doi: 10.1016/j.neucom.2018.01.084.
[21] D. Hindarto and A. Djajadi, “Android-manifest extraction and labeling method for malware compilation and dataset creation,”
International Journal of Electrical and Computer Engineering, vol. 13, no. 6, pp. 6568–6577, 2023, doi:
10.11591/ijece.v13i6.pp6568-6577.
[22] Y. Wan, M. Shao, Y. Cheng, D. Meng, and W. Zuo, “Progressive convolutional transformer for image restoration,” Engineering
Applications of Artificial Intelligence, vol. 125, 2023, doi: 10.1016/j.engappai.2023.106755.
[23] S. Gao et al., “Implicit diffusion models for continuous super-resolution,” in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pp. 10021–10030, 2023, doi: 10.1109/cvpr52729.2023.00966.
[24] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee, “Enhanced deep residual networks for single image super-resolution,” IEEE
Computer Society Conference on Computer Vision and Pattern Recognition Workshops, pp. 1132–1140, 2017, doi:
10.1109/CVPRW.2017.151.
[25] S. H. Park, Y. S. Moon, and N. I. Cho, “Perception-oriented single image super-resolution using optimal objective estimation,” in
2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1725–1735, 2023, doi:
10.1109/cvpr52729.2023.00172.
[26] A. Lugmayr, M. Danelljan, L. Van Gool, and R. Timofte, “SRFlow: learning the super-resolution space with normalizing flow,” in
Computer Vision – ECCV 2020, pp. 715–732, 2020, doi: 10.1007/978-3-030-58558-7_42.
[27] J.-E. Yao, L.-Y. Tsao, Y.-C. Lo, R. Tseng, C.-C. Chang, and C.-Y. Lee, “Local implicit normalizing flow for arbitrary-scale image
super-resolution,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1776–1785, 2023, doi:
10.1109/cvpr52729.2023.00177.
[28] J. Wu, Z. He, and L. Zhuo, “Video satellite imagery super-resolution via a deep residual network,” International Geoscience and
Remote Sensing Symposium (IGARSS), vol. 2019, pp. 2762–2765, 2019, doi: 10.1109/IGARSS.2019.8900265.
[29] K. Karthik, S. S. Kamath, and S. U. Kamath, “Automatic quality enhancement of medical diagnostic scans with deep neural image
super-resolution models,” 2020 IEEE 15th International Conference on Industrial and Information Systems, ICIIS 2020 -
Proceedings, pp. 162–167, 2020, doi: 10.1109/ICIIS51140.2020.9342715.
[30] M. S. Habeeb, B. Aydin, A. Ahmadzadeh, M. Georgoulis, and R. A. Angryk, “Solar line-of-sight magnetograms super-resolution
using deep neural networks,” Proceedings - 2020 IEEE International Conference on Big Data, Big Data 2020, pp. 4586–4593,
2020, doi: 10.1109/BigData50022.2020.9378480.
[31] S. H. Nguyen and D. H. Phan, “Selective element domain interpolation technique for assumed rotations and shear strains in
polygonal finite element thick/thin plate analysis,” Thin-Walled Structures, vol. 186, 2023, doi: 10.1016/j.tws.2023.110677.
[32] W. Tan, N. Qin, Y. Zhang, H. McGrath, M. Fortin, and J. Li, “A rapid high-resolution multi-sensory urban flood mapping framework
via DEM upscaling,” Remote Sensing of Environment, vol. 301, 2024, doi: 10.1016/j.rse.2023.113956.
[33] T. Hunter, S. Hulsoff, and A. Sitaram, “SuperAdjoint: super-resolution neural networks in adjoint-based output error estimation,”
in XI International Conference on Adaptive Modeling and Simulation (ADMOS 2023), pp. 1-8, 2023, doi:
10.23967/admos.2023.058.
[34] J. Panda and S. Meher, “An improved image interpolation technique using OLA e-spline,” Egyptian Informatics Journal, vol. 23,
no. 2, pp. 159–172, 2022, doi: 10.1016/j.eij.2021.10.002.
[35] K. Zhang, X. Yang, J. Zang, and Z. Li, “FeLU: a fractional exponential linear unit,” Proceedings of the 33rd Chinese Control and
Decision Conference, CCDC 2021, pp. 3812–3817, 2021, doi: 10.1109/CCDC52312.2021.9601925.
[36] B. Grelsson and M. Felsberg, “Improved learning in convolutional neural networks with shifted exponential linear units (ShELUs),”
2018 24th International Conference on Pattern Recognition (ICPR), Beijing, China, 2018, pp. 517-522, doi:
10.1109/ICPR.2018.8545104.
[37] H. Shen, Z. Wang, J. Zhang, and M. Zhang, “L-Net: A lightweight convolutional neural network for devices with low computing
power,” Information Sciences, vol. 660, 2024, doi: 10.1016/j.ins.2024.120131.

BIOGRAPHIES OF AUTHORS

Djarot Hindarto received the B.Eng. degree in computer engineering from

Sepuluh Nopember Institute of Technology (ITS), Indonesia, in 1994 and the Master of
Information Technology Pradita University, in 2022, respectively. Currently, he is a lecture at
the Department of Information Technology, University Nasional Indonesia. His research
interests include security, artificial intelligence, deep learning, machine learning, data mining,
game technology, internet of things, and blockchain. He can be contacted at email:
[email protected].

Enhancing image quality with SR-ResNet: super-resolution techniques for small … (Djarot Hindarto)
4666  ISSN: 2252-8938

Mohammad Iwan Wahyuddin received the B.Eng. degree in electrical

engineering from, University of Northern Sumatra, Indonesia, Master of Gajah Mada
University, Indonesia, in and Doctoral in electrical engineering from Indonesia University,
Indonesia. Currently, he is a lecturer at the Department of Information Technology, University
Nasional. His research interests include artificial intelligence, machine learning, and
networking. He can be contacted at email: [email protected].

Andrianingsih Andrianingsih completed a doctoral degree in Information

Technology at the Doctoral Program 2023. The Magister Program in Information Systems was
completed between 2003 and 2006, while the bachelor’s program in the same discipline
commenced between 1998 and 2002 at Gunadarma University in Jakarta. She is affiliated with
various international and national organizations, including the IEEE, the position of Head of the
Information Systems Study Program. She has been working as a lecturer since 2003 at the
National University. Conducting field study on the topics of information systems technology,
geospatial analysis, business intelligence, communication data, and LoRa WAN. She can be
contacted at email: [email protected].

Ratih Titi Komalasari is a lecturer and researcher at the Informatics Study

Program, Faculty of Communication and Information Technology, National University of
Jakarta since 2010. She has an undergraduate education background in Informatics, Gunadarma
University, master’s in information systems management at Gunadarma University and is
currently pursuing doctoral studies in information systems at Diponegoro University. The
teaching areas artificial intelligence, game programming, and data science. Meanwhile, the
research field she is pursuing is data science, artificial intelligence, and geoinformatics. She can
be contacted at email: [email protected].

Endah Tri Esti Handayani is a lecturer and researcher at the Information Systems
Study Program, Faculty of Communication and Information Technology, National University
of Jakarta since 2017. She has an undergraduate education background in Food Technology,
Brawijaya University, Masters in Information Systems Management at Gunadarma University
and is currently pursuing doctoral studies information systems at Diponegoro University. The
teaching areas taught are statistics and probability, discrete mathematics, and management
information systems. Meanwhile, the research field he is pursuing is data science. She can be
contacted at email: [email protected].

Mochamad Hariadi is a lecturer and researcher at the Information Technology

Department Computer Engineeering, Faculty of Electrical Engineering and Intelligent
Informatics, Institut Teknologi Sepuluh Nopember, since 1996. He has an undergraduate
education background in Bachelor Electrical Engineering (Institut Teknologi Sepuluh
Nopember), Magister Information System (Tohoku University), Doctoral Computer Science
and Mathematics (Tohoku University). He can be contacted at email: [email protected].

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4654-4666