Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263

Intelligent Resource Optimization: Enhancing

Component Reuse through AI-Driven
Image Classification
Kunal G. Borase1; Dhanashree Meshram2; Sowmiya Radhakrishnan3;
Praveen Kumar Burra4; Bharani Kumar Depuru5
1;2;3;4;5
Aispry

Publication Date: 2025/04/11

Abstract: This paper presents an AI-powered system designed to automate the identification and cataloging of electric
switchgear components, improving inventory management and minimizing errors caused by manual classification.
Traditional identification methods rely on human efforts, which are labor-intensive and prone to misclassification, leading
to inefficiencies in warehouse operations. To overcome these challenges, we leveraged YOLO-based deep learning models to
classify switchgear components accurately while ensuring seamless integration with inventory records. Our approach
involved training YOLO models to classify switchgear components based on their unique visual features. The model matches
each identified component against a Master Data Sheet containing essential details such as part numbers, dimensions,
weight, and material specifications. By leveraging YOLO’s advanced feature extraction and classification capabilities, our
system achieves high precision in distinguishing visually similar components, ensuring reliable and real-time processing
suitable for industrial deployment. During model development, we addressed critical challenges such as variations in lighting
conditions, different orientations of components, and cluttered warehouse environments. Extensive data augmentation
techniques[10] and model fine-tuning were applied to enhance robustness and maintain high classification accuracy across
diverse scenarios. The final AI model achieves up to 95% accuracy, significantly reducing manual identification efforts by
70%, demonstrating its effectiveness in real-world applications. By automating switchgear component identification, our
system significantly enhances inventory tracking, minimizes human errors, and optimizes warehouse efficiency. This
research highlights the transformative potential of YOLO-based AI automation in industrial inventory management, paving
the way for future advancements in intelligent spare part classification and cataloging.

Keywords: Image Classification, Component Identification, YOLOv8n-cls, YOLOv8s-cls, YOLOv8m-cls, Deep Learning, Data
Augmentation, CRISP-ML(Q), Inventory Management, Resource Optimization, AI Automation, Warehouse Efficiency.

How to Cite: Kunal G. Borase; Dhanashree Meshram; Sowmiya Radhakrishnan; Praveen Kumar Burra; Bharani Kumar Depuru
(2025). Intelligent Resource Optimization: Enhancing Component Reuse through AI-Driven Image Classification.
International Journal of Innovative Science and Research Technology, 10(3), 2613-2622.
https://doi.org/10.38124/ijisrt/25mar1263

I. INTRODUCTION labeling, and delays in stock updates, and as the number of

components increases, these issues become harder to handle.
Electric switchgear is made up of different components, To solve this problem, this research focuses on developing an
each playing an important role in controlling and distributing AI-powered system that can automatically recognize and
electrical power safely. These components, such as circuit classify both assembled and leftover components. This
breakers, switches, contactors, fuses, and protective relays, system will enable real-time inventory tracking, helping
work together to prevent overloads, short circuits, and voltage warehouses keep accurate records, reduce manual effort, and
fluctuations. Once all the parts are assembled, they form a minimize errors. By using AI for classification, warehouse
complete switchgear unit, ensuring the safe and efficient operations[11] can become faster, more efficient, and better
operation of electrical systems. However, during the organized, leading to improved productivity and resource
assembly process, some extra or unused parts are left behind, management.
making it difficult to track and manage them properly.
Identifying these leftover components is necessary to reduce Artificial intelligence has become a powerful tool for
waste, improve inventory management, and make better use automating inventory processes, significantly reducing
of resources. Traditional methods like manual checking and human effort and errors. The proposed system employs
barcode scanning often lead to misplaced items, wrong YOLO(You Only Look Once) single shot image

IJISRT25MAR1263 www.ijisrt.com 2613

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
classification model to classify electric switchgear accuracy in component classification and seamless inventory
components even in challenging warehouse conditions, such tracking.
as poor lighting, and varying orientations. Unlike traditional
inventory tracking systems, AI-powered solutions provide By implementing this AI-powered vision system,
faster, more accurate, and scalable inventory management, warehouses can reduce manual stock verification efforts,
enabling seamless warehouse operations[11]. improve tracking accuracy, and optimize inventory
workflows. This study demonstrates how deep learning-
To build a robust and reliable classification system, a based classification can revolutionize warehouse
comprehensive dataset of switchgear component images was management [11]by providing businesses with automated,
collected and annotated. These images, captured from real-time component identification, minimal human errors,
multiple angles and under different environmental and enhanced operational efficiency. Through this approach,
conditions, were carefully labeled to train the AI model organizations can achieve cost savings, increased
effectively. To improve generalization, data augmentation productivity, and better resource utilization, making AI-
techniques[10] such as brightness normalization, contrast driven inventory automation an essential advancement in
adjustments, rotation, and noise filtering were applied, modern warehouse operations[11].
ensuring that the model maintains high classification
accuracy across diverse warehouse settings. The project methodology followed here is the open
source CRISP-ML(Q)[5] methodology from
A structured methodology, CRISP-ML(Q)[5], was 360DigiTMG(ak.1) [Fig 1], which stands for Cross Industry
followed to develop the AI-based inventory system. This Standard Process for Machine Learning with Quality
approach provides a systematic framework for building Assurance. This structured methodology ensures a systematic
machine learning models, ensuring efficient execution of approach to problem identification, data preprocessing,
each phase from data collection and preprocessing to model model training, evaluation, deployment and monitoring and
training, evaluation, and deployment. The system was maintenance. By following CRISP-ML(Q), the project
evaluated using key performance metrics, including follows a well-defined lifecycle from data collection to real-
precision, recall, confusion matrix analysis, and classification world implementation, ensuring robust and efficient model
accuracy. Testing different YOLO versions allowed us to performance.
select the most efficient model for deployment, ensuring high

Fig 1 This Figure Depicts the CRISP-ML(Q) Architecture that We have followed for this Research Study.
(Source: Mind Map - 360DigiTMG)

IJISRT25MAR1263 www.ijisrt.com 2614

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
II. METHODOLOGY AND TECHNOLOGY further enhance the dataset, images were also obtained from
secondary sources, including open-source platforms.
A. Data Collection Captured images exhibited varying resolutions, such as
The dataset utilized in this research was collected 2560x1920, 3024x4032, and 4032x3024 pixels, providing a
directly from the client's supplier location, capturing images wide range of scale and quality to ensure effective model
under various real-world environmental conditions to ensure training.
robust model performance. Over 1000+ images of different Below table [Fig 2] shows data description and sample
assembly components were gathered, featuring diverse component images.
orientations[9], backgrounds, and lighting scenarios. To

Fig 2 Data and Data Description

B. Data Preprocessing and Augmentation data, thereby enhancing the accuracy and reliability of the
Effective data preprocessing and augmentation classification model.
significantly enhanced the performance and accuracy of the
electric switchgear component[6] classification model. The  Dataset Balancing:
preprocessing pipeline involved several critical steps: To address the inherent class imbalance[12], images
across various switchgear component categories were
 Image Acquisition and Standardization: balanced by ensuring equal representation. Each component
Images of switchgear components were systematically class was standardized to a specific number of images,
collected from diverse warehouse environments. To maintain significantly improving model generalization and minimizing
dataset consistency and facilitate model training efficiency, potential bias.
all images were uniformly resized to dimensions of 320x320
pixels.  Data Augmentation:
Data augmentation techniques[10] were systematically
 Data Cleaning: applied to increase the variability and robustness of the
Images that were blurry, noisy, poorly illuminated, or of dataset. These augmentations simulated real-world
substandard quality were identified and either corrected or operational conditions encountered in warehouse
excluded from the dataset. This ensured high-quality input environments: Below table [Table 1] shows data
Augmentation techniques applied in training dataset.

IJISRT25MAR1263 www.ijisrt.com 2615

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
Table 1 Augmentation Techniques Applied in Training Dataset

These preprocessing and augmentation steps  Data Splitting:

collectively ensured high-quality, diverse data inputs, The raw dataset comprised 1000+ original images
significantly contributing to the robustness and accuracy of categorized into 34 distinct classes. After applying
the model’s predictive capability in operational warehouse comprehensive data augmentation[10], the dataset expanded
scenarios. substantially to a total of 10,336 images. The dataset was then
strategically partitioned into three subsets to facilitate
efficient model training and validation: Below table [Table 2]
shows Data splitting after Augmentation techniques applied.

Table 2 Data Splitting after Augmentation Techniques

This structured split enabled robust training, rigorous designed for seamless integration of multiple stages,
validation, and thorough evaluation, ensuring accurate and encompassing data collection, preprocessing, model training,
reliable classification outcomes across diverse real-world evaluation, deployment, and maintenance. This systematic
warehouse conditions. workflow ensures robust and efficient real-time performance
tailored specifically for industrial and operational
C. Model Architecture environments.
The component detection and classification system
leverages a comprehensive and structured architecture

Fig 3 End-to-End High-Level Architecture of the Component Classification System

This architecture demonstrates the end-to-end pipeline, The architecture begins with an initial phase of client
from image acquisition to model evaluation and deployment, interaction and extensive business analysis, which clarifies
ensuring that the system is scalable and efficient for real- the requirements and scopes the classification objectives.
world warehouse applications. Subsequent data collection utilizes sources such as CCTV
and local hardware systems to gather relevant visual data for
analysis.

IJISRT25MAR1263 www.ijisrt.com 2616

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
Data collected is then consolidated within an  Model Variants for Image Classification:
Operational Platform, leveraging tools and technologies like
Python scripts and database systems for structured data  YOLOv8n-cls:
handling. In the preprocessing phase, raw images undergo The nano (n) version of YOLOv8 is optimized for fast
essential transformations including resizing, normalization, inference with minimal computational requirements. It
augmentation, and enhancement techniques to optimize their consists of approximately 3.2 million parameters and 8.7
quality and suitability for model training. Python libraries like billion FLOPs(Floating Point Operations per Second). It is a
OpenCV and Pillow are extensively utilized to streamline and lightweight model suitable for real-time applications where
standardize the preprocessing procedures. processing speed is crucial. In this project, it was evaluated to
determine whether a compact architecture could still achieve
The preprocessed images are then fed into the model- accurate component classification while maintaining high
building stage, where advanced deep-learning frameworks processing speeds, making it viable for rapid operational
such as YOLO(You Only Look Once) from Ultralytics and decision-making.
PyTorch are employed. This step involves fine-tuning pre-
trained CNN(Convolutional Neural Network) architectures,  YOLOv8s-cls:
allowing the extraction of critical image features necessary The small(s) variant of YOLOv8 offers a balanced
for accurate detection and classification tasks. approach, featuring approximately 11.2 million parameters
and 28.6 billion FLOPs. This version aims to achieve a
Model evaluation follows a rigorous protocol to validate favorable trade-off between computational efficiency and
the effectiveness and accuracy of the developed model. classification accuracy. It was utilized to validate
Performance metrics such as precision, recall, F1-score, and performance across a broader range of scenarios, ensuring
mAP(mean Average Precision) are calculated to ensure the scalability and reliability for general industrial deployments.
model’s reliability using tools like PyTorch evaluation
modules.  YOLOv8m-cls:
The medium (m) variant of YOLOv8 provides enhanced
Once validated, the model is deployed using performance capabilities with around 25.9 million parameters
AWS(Amazon Web Services) cloud environment where and 78.9 billion FLOPs. It was selected for scenarios
streamlit Framework[7] for building AI/ML application, was requiring greater accuracy and robustness without sacrificing
used enabling easy integration and scalability. Post- significant inference speed. Its evaluation aimed to confirm
deployment, continuous monitoring and maintenance ensure its suitability for complex component classifications where
model performance remains optimal, with real-time analytics accuracy is paramount.
and feedback loops for continual improvement and adaptation
to new data or evolving scenarios.  ResNet50:
ResNet50[2], a residual neural network with 50 layers,
This complete and detailed pipeline, illustrated in [Table is renowned for its ability to efficiently handle complex
3], ensures an efficient, accurate, and highly adaptable image classification tasks by mitigating the vanishing
component classification system suitable for real-world gradient problem through residual connections. Its deep
deployment. architecture enables high accuracy, particularly suitable for
detailed component differentiation.
D. Model Building
For the classification of components, a comprehensive
 MobileNetV2:
approach was adopted using advanced deep learning models.
MobileNetV2[2] is optimized for mobile and embedded
The primary focus was on YOLO-cls(You Only Look Once
vision applications. Utilizing depthwise separable
classification)[1] models, specifically YOLOv8n-cls[8],
convolutions, it provides a highly efficient architecture with
YOLOv8s-cls[8], YOLOv8m-cls[8], ResNet50[2],
fewer parameters, making it ideal for scenarios demanding
MobileNetV2[2], MobileNetV3_small_100[3], ViT-base-
high-speed inference with minimal computational resources.
patch 16-224[4], and ViT_B_16[4], due to their proven
capability in real-time object detection and classification
 MobileNetV3_small_100:
tasks. These models were trained using all available
MobileNetV3_small_100[3] represents the smallest and
component classes, amounting to 34 distinct categories.
most efficient variant of the MobileNetV3[3] family,
YOLO's efficient architecture facilitated high accuracy and
designed for high efficiency and performance on resource-
low latency, making it ideal for real-world industrial
constrained devices. It combines innovative design
applications.
techniques with advanced architecture search for optimized
performance in industrial classification tasks.
The diverse selection of architectures, especially YOLO
variants trained across the full spectrum of 34 classes,
ensured a robust comparative study. This systematic  ViT-base-patch 16-224:
evaluation facilitated the identification of optimal models for Vision Transformer (ViT)[4] base variant, with a patch
deployment, balancing high accuracy, efficient inference, and size of 16 pixels and input image size of 224 pixels, utilizes
scalability for industrial component classification. self-attention mechanisms instead of traditional
convolutional approaches. It excels at capturing global

IJISRT25MAR1263 www.ijisrt.com 2617

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
context, making it highly effective for detailed component effectiveness in accurately classifying 34 distinct industrial
classification tasks requiring global spatial relationships. components while maintaining moderate resource demands.

 ViT_B_16: MobileNet variants, including MobileNetV2[2] and

Similar to ViT-base, ViT_B_16[4] employs transformer MobileNetV3_small_100, showcased notable efficiency and
architectures for image recognition, providing robust were optimal for scenarios involving constrained
performance and efficient context understanding through computational resources, such as edge or mobile devices.
attention mechanisms. This model is particularly suitable for However, they displayed comparatively lower accuracy,
complex scenarios involving detailed visual component limiting their suitability for critical deployments requiring
classification. very high precision.

E. Model Evaluation ResNet50[2], although robust in general image

A rigorous model evaluation strategy was implemented recognition tasks, showed moderate performance in this
to assess the performance, scalability, and applicability of application due to relatively high computational costs and
various deep learning architectures. Models evaluated lower achieved accuracy. Vision Transformer(ViT) models
included YOLOv8 (variants: nano, small, and medium)[8], offered promising accuracy, particularly ViT_B_16[4], but
ResNet50[2], MobileNetV2[2], MobileNetV3_small_100[3], their computational overhead and inference latency were
and Vision Transformer variants (ViT-base-patch 16-224 and significantly higher, restricting practical deployment to cloud
ViT_B_16)[4]. These models were systematically compared environments with ample computational resources.
across multiple critical metrics: accuracy, inference speed,
resource utilization, ease of deployment, and scalability. Based on this comprehensive evaluation, YOLOv8m-
cls[8] emerged as the most suitable model for practical
The YOLOv8 models consistently outperformed other implementation, offering the optimal balance among
architectures in terms of achieving a balance between accuracy, inference speed, resource efficiency, and
accuracy and computational efficiency, demonstrating their scalability. Its performance aligns precisely with the
suitability for real-time industrial image classification tasks. objectives of enhancing component reuse through efficient
Specifically, the YOLOv8m-cls[8] variant exhibited the and precise classification, thereby supporting intelligent
highest validation accuracy of 95.20%, clearly indicating its resource optimization in industrial applications.

Table 3 Model Selection and Performance Evaluation using Decision Analysis and Resolution (DAR)

This analysis visually highlights how each model requirements, ease of deployment, and broad scalability, thus
compares regarding accuracy, resource efficiency, ease of confirming its suitability for real-time industrial applications
use, system requirements, and scalability. YOLOv8m-cls[8] aimed at component reuse and intelligent resource
is clearly identified as the optimal model choice due to its optimization.
superior balance of high accuracy, moderate computational

IJISRT25MAR1263 www.ijisrt.com 2618

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
F. Model Hyperparameter Tuning much regularization is applied, and how image enhancements
To improve the performance of the YOLOv8m-cls like flipping and brightness adjustments are used. These
model, hyperparameter tuning was carried out by adjusting tuning methods are summarized in Table 4.
key settings such as learning rate, weight decay, optimizer
type, and data augmentation techniques[10]. The goal was to The objective of this hyperparameter tuning was to
increase the model’s accuracy while keeping it stable and achieve the best possible trade-off between accuracy,
efficient for real-world industrial use. Different tuning computational efficiency, and model stability for practical
strategies were tested by changing how the model learns, how deployment in industrial applications

Table 4 Hyperparameter Tuning for the Best Model (YOLOv8m-cls)

After testing various tuning approaches, the best model from 0.926 to 0.964, reducing incorrect classifications, while
setup showed significant improvements``in accuracy and the recall improved from 0.911 to 0.959, ensuring fewer
detection performance. As seen in Table 5, the validation missed detections. Although training time rose slightly by
accuracy increased from 95.20% to 97.35%, making the 12%, the model’s inference speed improved, allowing faster
model more reliable in correctly identifying switchgear real-time classification. The GPU usage remained moderate,
components. The mAP50(Mean average precision calculated meaning it still runs efficiently without requiring excessive
at an intersection over union (IoU) threshold of 0.50) score computing power. Additionally, scalability improved from
improved from 0.963 to 0.995, meaning the model became good to excellent, making the model more adaptable for
more precise in detecting objects. The precision increased large-scale industrial applications.

Table 5 Model Performance Metrics after Hyperparameter Tuning

IJISRT25MAR1263 www.ijisrt.com 2619

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
These improvements show that fine-tuning the model’s III. MODEL DEPLOYMENT
parameters significantly enhances its performance, making it
highly accurate, efficient, and reliable for warehouse After testing and fine-tuning different models, the
inventory tracking[11]. By reducing human errors and YOLOv8m-cls (Set 1) model was deployed in an AWS cloud
automating the classification process, this optimized AI environment to ensure scalability, real-time processing, and
model helps industries manage their components more easy accessibility. The Streamlit framework[7] was used to
effectively, improving productivity and resource utilization. build an interactive AI-powered web application, allowing
users to easily upload images of switchgear components and
The optimized YOLOv8m-cls model (Set 1) achieved receive instant classification results. Figure 4 shows
remarkable improvements in accuracy, precision, and recall, streamlit’s lightweight and efficient design enabled seamless
demonstrating its efficacy in real-time industrial applications. integration with the trained YOLO model, ensuring smooth
This tuning provided a robust model, capable of precise and real-time inference.
rapid component classification, critical for enhancing
component reuse and intelligent resource optimization.

Fig 4 A Streamlit Framework for Automated Component Identification and Attribute Extraction

IJISRT25MAR1263 www.ijisrt.com 2620

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
The deployed system allows warehouse staff to upload By implementing this AI-driven system, warehouses
images, which are processed in real time to classify the can reduce manual effort, enhance tracking accuracy, and
component. Along with classification, the system retrieves optimize workflows. The findings of this research
and displays key attributes such as Material Description, demonstrate the transformative impact of AI in industrial
Length, Width, Thickness, Coating Type, and Part Weight to inventory management, paving the way for future
provide comprehensive inventory insights. advancements in automated component classification,
predictive analytics, and mobile-friendly AI applications.
This cloud-based deployment ensures that warehouses
can automate component classification, reduce manual FUTURE SCOPE
efforts, and improve inventory accuracy while maintaining
seamless and efficient operations. By leveraging AWS The model can also be used on tablets or mobile phones
services, the system supports real-time tracking, minimizes by turning this AI application into a mobile app. This would
classification errors, and enhances overall warehouse allow workers to scan and classify switchgear components
management[11], making AI-driven inventory automation an instantly, making inventory management easier and more
essential advancement in modern industrial operations. efficient. The mobile version would run smoothly using
lightweight AI models, ensuring fast results even without an
IV. CONCLUSION internet connection.

This research successfully developed an AI-powered In the future, the system could also be improved to
system for classifying and identifying electric switchgear identify multiple components in a single image. This would
components, significantly improving warehouse inventory help warehouse staff scan entire shelves at once instead of
management. Traditional manual classification methods classifying parts one by one, saving time and reducing
often result in misidentification, misplaced inventory, and manual work. Additionally, AI-powered image enhancement
operational inefficiencies. By leveraging YOLO-based deep techniques could be integrated to improve detection accuracy
learning models, the proposed system automates component in low-light or cluttered environments.
identification with high precision, reducing human errors and
increasing efficiency. Another potential enhancement is the incorporation of
augmented reality (AR) to overlay real-time classification
The model was trained on a comprehensive dataset, and inventory details directly on the device’s screen. This
incorporating diverse environmental conditions to ensure would help workers quickly locate and verify components
robustness. Data augmentation techniques further enhanced without needing to cross-check with manual records.
the model’s generalization, making it effective in real-world
warehouse settings. Among the tested models, YOLOv8m- Furthermore, integrating cloud connectivity and
cls(Set1) emerged as the best-performing model after predictive analytics could enable real-time stock monitoring,
hyperparameter tuning, significantly improving classification automated restocking alerts, and data-driven decision-making
performance. for inventory optimization. Future versions might also
support voice commands, making hands-free operation
Before tuning, the YOLOv8m-cls baseline model possible for increased efficiency in fast-paced warehouse
achieved a validation accuracy of 95.20%, with a precision of environments.
0.926 and a recall of 0.911. The mAP50 score was 0.963,
indicating strong detection performance. The model These advancements will make the system more
exhibited moderate inference speed and scalability, making it powerful, user-friendly, and indispensable for modern
suitable for deployment but with room for optimization. warehouse operations.

After tuning (Set 1), the model demonstrated notable ACKNOWLEDGMENTS

improvements, achieving a validation accuracy of 97.35%,
precision of 0.964, and recall of 0.959. The mAP50 score We sincerely thank 360 DigiTMG for providing us with
increased to 0.995, showcasing enhanced detection the opportunity to work on this project. We also appreciate
performance. Although training time increased by 12%, the guidance and support of our partners throughout this
inference speed improved, and scalability was enhanced from research. Additionally, we acknowledge the use of the
good to excellent. CRISP-ML(Q) framework and ML Workflow, which are
openly available on the official 360 DigiTMG website and
To ensure scalability and accessibility, the model was used with their explicit consent.
deployed in an AWS cloud environment using the Streamlit
framework, allowing warehouse staff to upload images and REFERENCES
receive instant classification results. This cloud-based
approach enables real-time tracking, seamless inventory [1]. Naqif Fared Nor, Hazlyna Harun. An Experiment on
updates, and smooth integration with warehouse management Lung Disease Classification using YOLOv8. DOI:
systems. https://doi.org/10.58915/amci.v13i3.626.
[2]. Mohammad Rafka Mahendra Ariefwan, I Gede
Susrama Mas Diyasa, Kartika Maulidya Hindrayani.

IJISRT25MAR1263 www.ijisrt.com 2621

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1263
InceptionV3, ResNet50, ResNet18 and MobileNetV2 and Data Engineering, vol. 21, no. 9, pp. 1263-1284,
Performance Comparison on Face Recognition 2009. DOI: 10.1109/TKDE.2008.239.
Classification. [13]. Chawla, Nitesh V., et al. "SMOTE: Synthetic
DOI:https://doi.org/10.56480/jln.v4i1.990. Minority Over-sampling Technique." Journal of
[3]. Mohamed Abd Elaziz 1,2,*ORCID,Abdelghani Artificial Intelligence Research, vol. 16, pp. 321-357,
Dahou 3,Naser A. Alsaleh 4ORCID,Ammar H. 2002. DOI: 10.1613/jair.953.
Elsheikh 5,*ORCID,Amal I. Saba 6 andMahmoud [14]. Japkowicz, Nathalie, and Shaju Stephen. "The Class
Ahmadein 4,5 ORCID. Boosting COVID-19 Image Imbalance Problem: A Systematic Study." Intelligent
Classification Using MobileNetV3 and Aquila Data Analysis, vol. 6, no. 5, pp. 429-449, 2002. DOI:
Optimizer Algorithm. DOI: 10.3233/IDA-2002-6504.
https://doi.org/10.3390/e23111383. [15]. He, Haibo, and Yunqian Ma. "Imbalanced Learning:
[4]. Chu Myaet Thwal1 , Ye Lin Tun1 , Minh N. H. Foundations, Algorithms, and Applications." IEEE
Nguyen2 , Eui-Nam Huh1 , Choong Seon Hong1* Press/Wiley, 2013. DOI: 10.1002/9781118646106.
1Kyung Hee University 2Vietnam-Korea University
of Information and Communication Technology.
CLIP-PING: Boosting Lightweight Vision-Language
Models with Proximus Intrinsic Neighbors Guidance.
DOI: https://doi.org/10.48550/arXiv.2412.03871.
[5]. Stefan Studer , Thanh Binh Bui , Christian Drescher ,
Alexander Hanuschkin , Ludwig Winkler , Steven
Peters and Klaus-Robert Müller Towards CRISP-
ML(Q): A Machine Learning Process Model with
Quality Assurance Methodology .
DOI:https://doi.org/10.3390/make3020020.
[6]. R. Umamaheswari a & R. Sarathi a a Department of
Electrical Engineering , Indian Institute of
Technology Madras , Chennai, India Published online:
07 Oct 2011. Identification of Partial Discharges in
Gas-insulated Switchgear by Ultra-high frequency
Technique and Classification by Adopting Multi-class
Support Vector Machines. DOI:
https://doi.org/10.1080/15325008.2011.596506.
[7]. Himangi Dani, Pooja Bhople, Hariom Waghmare,
Kartik Munginwar, Prof. Ankush Patil Review on
Frameworks Used for Deployment of Machine
Learning Model
DOI:https://doi.org/10.22214/ijraset.2022.40222.
[8]. Utsha Saha∗† , Imtiaj Uddin Ahamed‡, Md Ashique
Imran§, Imam Uddin Ahamed¶, Al-Amin Hossain‡,
Ucchash Das Gupta YOLOv8-Based Deep Learning
Approach for Real-Time Skin Lesion Classification
Using the HAM10000 Dataset.
DOI:10.1109/HealthCom60970.2024.10880715.
[9]. Ilia V. Safonov, Ilya V. Kurilin, Michael N.
Rychagov, Ekaterina V. Tolstaya, "Content-Based
Image Orientation Recognition," in Adaptive Image
Processing Algorithms for Printing, 2018. DOI:
10.1007/978-981-10-6931-4_12.
[10]. Marcus D. Bloice, Christof Stocker, Andreas
Holzinger, "Augmentor: An Image Augmentation
Library for Machine Learning," arXiv preprint
arXiv:1708.04680, 2017. DOI:
10.48550/arXiv.1708.04680.
[11]. Niels Faber, René de Koster, and Ale Smidts,
"Organizing Warehouse Management," International
Journal of Physical Distribution & Logistics
Management, vol. 43, no. 9, pp. 764-784, 2013. DOI:
10.1108/IJPDLM-11-2011-0203.
[12]. Haibo He and Edwardo A. Garcia, "Learning from
Imbalanced Data," IEEE Transactions on Knowledge

IJISRT25MAR1263 www.ijisrt.com 2622