Institut Riset dan Publikasi Indonesia (IRPI)

MALCOM: Indonesian Journal of Machine Learning and Computer Science
Journal Homepage: https://journal.irpi.or.id/index.php/malcom
Vol. 5 Iss. 3 July 2025, pp: 788-796
ISSN(P): 2797-2313 | ISSN(E): 2775-8575

Real-Time Road Damage Detection on Mobile Devices using
TensorFlow Lite and Teachable Machine
Lusindah Erdian Nova 1*, Yan Rianto2
1,2

Faculty of Computer Science, Universitas Nusa Mandiri, Indonesia
E-Mail: 114240010@nusamandiri.ac.id, 2yan.yrt@nusamandiri.ac.id

Received Jan 25th 2025; Revised Mar 23th 2025; Accepted Apr 14th 2025; Available Online Jun 19th 2025, Published Jun 22th 2025
Corresponding Author: Lusindah Erdian Nova
Copyright ©2025 by Authors, Published by Institut Riset dan Publikasi Indonesia (IRPI)

Abstract
This study presents a mobile-based road damage detection system using Teachable Machine and TensorFlow Lite to
support real-time monitoring and efficient infrastructure maintenance. The system identifies road damage types such as
cracks, potholes, and uneven surfaces. The RDD2020 dataset is used for model training, with preprocessing steps including
augmentation, normalization, and resizing. A Convolutional Neural Network (CNN) model is trained through Teachable
Machine for ease of customization. TensorFlow Lite is employed for on-device inference, with optimization techniques like
quantization and pruning applied to improve speed and reduce model size. The system is evaluated using precision, recall,
F1-score, and accuracy metrics under varying lighting and weather conditions. The final model is deployed in a mobile
app using TensorFlow Lite Interpreter for efficient performance. Experimental results show high detection accuracy, with
a precision of X% and F1-score of Y% (insert actual values). This approach offers a lightweight, cost-effective solution for
road maintenance authorities and urban planners. Future enhancements include dataset expansion, integration with
mapping tools, and improved robustness in diverse environments. Overall, the proposed system enables real-time, accurate
road damage detection and supports smarter, eco-friendly infrastructure management.
Keyword: Mobile Application, Real-Time Monitoring, Road Damage Detection, Teachable Machine, TensorFlow Lite

1.

INTRODUCTION
Road surface management aims to reduce accidents and enhance pavement quality. The surface of the
road is vulnerable to several kinds of damage brought on by rainwater seeping into a devastated region [1].
After that, the water seeps through the compacted dirt beneath the pavement and results in soil erosion, both
of which can have negative consequences, like the subsidence of the ground. Additionally, if the affected region
is not fixed, the quality of the pavement would deteriorate even further, impacting the steering control of
automobiles and causing mishaps. To avoid road quality management, solutions have been created in response
to incidents. In recent years [2].
In the construction industry, crack detection has been utilized for so long that it needs to be automated
and upgraded. With the help of this crack-detecting system, the manual processes have been automated. It
makes inspection and rehabilitation less expensive. utilizing a variety of image capture and processing
techniques that can be used for both automatic crack detection and optimization [3]. Methods and algorithms
to find and increase the accuracy of locating potholes and cracks in structural elements. Given the limitations
of the existing issue, more study is urgently needed to determine whether semantic segmentation may be
applied as a feature extraction technique for picture classification tasks utilizing convolutional neural networks
(CNN) [4][5].
Mobile technology has transformed a number of industries in recent years, including infrastructure
maintenance and transportation. Effective road damage monitoring and detection is a major challenge for road
maintenance authorities. Timely identification of damage is necessary to maintain public safety and save repair
expenses. Road damage detection has historically been less accessible and effective due to its reliance on laborintensive automated systems or manual inspections, both of which have high hardware requirements [6]. The
development of deep learning and machine learning methods has made it feasible to use mobile devices for
real-time road damage detection. Specifically, TensorFlow Lite, a condensed variation of Google's TensorFlow
framework, provides a robust yet resource-conserving environment for mobile machine learning model
execution. Alongside Teachable Machine, an intuitive instrument [7].

DOI: https://doi.org/10.57152/malcom.v5i3.1933

788

ISSN(P): 2797-2313 | ISSN(E): 2775-8575
Road surface management is essential for reducing accidents and maintaining pavement quality. Road
damage, caused by factors like water infiltration and erosion, can deteriorate pavement conditions, affecting
vehicle control and safety. Traditional crack detection methods rely on manual inspections, which are laborintensive and prone to errors. Machine learning and image processing techniques have been introduced to
automate detection and improve accuracy.TensorFlow Lite is a lightweight version of TensorFlow designed
for mobile and embedded devices, enabling real-time inference with minimal computational resources. It
optimizes models through quantization and pruning, reducing size while maintaining accuracy. Meanwhile,
Teachable Machine is a user-friendly tool that allows model training without advanced coding skills, making
machine learning more accessible. Advancements in mobile computing and deep learning have enabled realtime road damage detection on smartphones. This study leverages TensorFlow Lite and Teachable Machine to
develop a mobile-based system for efficient and accurate road damage monitoring.
Road damage is a critical issue that directly impacts traffic safety and infrastructure maintenance costs.
Damage types such as cracks, potholes, and uneven road surfaces can lead to accidents, accelerate infrastructure
degradation, and increase repair expenses if not addressed promptly. Therefore, regular road condition
monitoring is essential to prevent further damage and ensure road user safety.
Traditional road damage detection methods typically rely on manual inspections by field personnel or
specialized vehicles equipped with advanced sensors. While manual inspection is still widely used, it has
several limitations, including: Time and resource-intensive – Requires significant labor and is slow, Prone to
inaccuracies – Relies on human expertise, leading to potential inconsistencies in detection, Inefficient for largescale monitoring – Difficult to apply to vast areas or hard-to-reach locations [8].
On the other hand, automated sensor-based systems, such as LiDAR or high-resolution cameras
mounted on inspection vehicles, offer improved accuracy. However, these systems have drawbacks such as:
1. High costs – Require expensive hardware and supporting infrastructure.
2. Environmental dependency – Performance can be affected by weather conditions or lighting quality.
3. Limited accessibility – Restricted to organizations with substantial budgets, making them less viable for
widespread adoption.
With advancements in technology, machine learning and computer vision have been increasingly
applied to automate road damage detection. Several studies have used CNNs to recognize road damage patterns
from digital images. While these models have shown promising results, most require high-performance GPU
computing, making them impractical for deployment on low-resource devices like smartphones. Mobile-based
detection technology offers a more practical, cost-effective, and accessible solution. By leveraging smartphone
cameras and optimized machine learning algorithms, this approach enables real-time damage detection without
the need for expensive additional hardware. However, existing implementations still face challenges such as
low accuracy, processing limitations, and difficulty adapting to varying lighting and road surface conditions.
To address these issues, this study proposes a CNN-based model optimized with TensorFlow Lite and
trained using Teachable Machine to enable accurate, lightweight, and efficient road damage detection on
mobile devices. This approach aims to provide an AI-driven solution that is faster, more affordable, and easier
to implement, supporting more effective infrastructure monitoring for governments, road maintenance
authorities, and the general public [7].
2.

LITERATURE REVIEW
An important benefit of TensorFlow operators with Graphics Processing Unit (GPU) and Tensor
Processing Unit (TPU) support over CPU-based training uses. Along with the current layers, like convolutions,
pooling, and TensorFlow's thick layers, developers can create their layers with unique layer definitions [7].
Layers created with hardware acceleration are also used by TensorFlow operators. Numerous sophisticated and
fundamental operations are available in TensorFlow in several disciplines, including neural networks, image
processing, and mathematics, and researchers can mix these operators to create new layers. Developing layers
for deep learning models is made simpler with TensorFlow's custom layer classes [8].
The Random Forest, AdaBoost, Decision Tree, and k-Nearest Neighbors machine learning models were
used in this investigation. These models were chosen because they have the potential to be useful in forecasting
roles for a particular task in a distributed agile environment, which is characterized by dynamic and cooperative
team structures [9].
To maintain road infrastructure, improve road safety, and guarantee prompt repairs, real-time road
damage identification is essential. Road damage has been monitored using a variety of methods and tools over
the years [10]. These systems range from more complex machine learning (ML) and computer vision-based
techniques to more conventional manual examinations. Deep learning and mobile computing have made it
possible to create smartphone-based real-time road damage detection systems. An overview of current
technologies and approaches in the field is provided here, with an emphasis on the combination of Teachable
Machine and TensorFlow Lite for the detection of road degradation.Some of the existing tools are compared
and recommendations are made to improve the ease of recreation of machine learning models by saving
complete information in project repositories maintained in normal source code control systems [11][12].
MALCOM - Vol. 5 Iss. 3 July 2025, pp: 788-796

789

MALCOM-05(03): 788-796
3.
3.1

METHOD
Data Collection and Preprocessing
This study utilizes the RDD2020 dataset, which contains images of various road conditions from urban,
highway, and rural areas. The dataset includes different types of road damage, such as longitudinal cracks,
transverse cracks, alligator cracks, and potholes. Images were collected under diverse weather conditions,
lighting variations, and road surface materials to improve generalization. Dataset potholes and Dataset Road
Satisfactory can be seen in Figure 1 and 2.
Despite its advantages, the dataset has limitations, including geographical bias, class imbalance, and
inconsistencies in image quality due to variations in camera devices. To mitigate these issues, data
augmentation techniques such as rotation, flipping, contrast adjustment, and noise addition were applied. These
techniques enhance model robustness by simulating real-world conditions and ensuring better adaptability in
diverse environments [13].

Figure 1. DataSet Potholes
Layers of Convolution: In order to better capture low-level information like edges, textures, and colors,
these layers apply convolutional filters to the input image [14]. Higher-level features like forms or patterns are
learned by the layers as they advance. The model can learn more intricate patterns by introducing non-linearity
through the application of the Rectified Linear Unit (ReLU) activation function following each convolutional
operation. Layers of Pooling: By reducing the image's spatial dimensions (height and breadth), max pooling or
average pooling lowers computing complexity and aids the model in concentrating on the most crucial
elements.completely connected layers: The model moves on to completely connected layers that incorporate
data from all areas of the image after convolution and pooling. These layers strive for ultimate detection or
categorization, figuring out whether the picture [15][16].
3.2

Model Development
To satisfy the model's input requirements, the images have been pre-processed (resized and normalized).
Convolutional Layers: These layers recognize patterns in the input image's edges, textures, and shapes by using
filters. They help the model pick up crucial characteristics, including fractures, potholes, and surface
differences. Activation Layers: Usually employing rectified linear units, or ReLUs, these layers give the model
non-linearity so it may pick up more intricate patterns. Pooling Layers: To lower dimensionality and
computational cost, max-pooling layers downsample the image by choosing the most noticeable characteristics
from regions [19].
This study employs a CNN due to its strong capability in feature extraction and pattern recognition,
making it well-suited for detecting road damage. CNNs can automatically identify structural patterns such as
cracks and potholes by processing spatial hierarchies in images, making them more effective than traditional
machine learning models for image-based classification tasks [17].
Real-Time Road Damage Detection on Mobile... (Nova and Rianto, 2025)

790

ISSN(P): 2797-2313 | ISSN(E): 2775-8575

Figure 2. Dataset Road Satisfactory
The model was trained using Teachable Machine, a user-friendly tool that simplifies machine learning
model training without requiring extensive coding. This tool was selected for its ease of use, efficient browserbased training, and seamless integration with TensorFlow. The trained model was then converted to
TensorFlow Lite, a lightweight version of TensorFlow optimized for mobile deployment. TensorFlow Lite was
chosen because it enables real-time inference on low-power devices, ensuring efficiency while maintaining
accuracy.
To enhance the model’s performance, images were preprocessed using resizing and normalization
techniques. Images were resized to a standard input dimension to maintain consistency across different
resolutions. Normalization was applied to scale pixel values, improving convergence during training and
reducing computational complexity. The CNN architecture consists of convolutional layers for feature
extraction, ReLU activation functions to introduce non-linearity, pooling layers to reduce spatial dimensions,
and fully connected layers for final classification. Quantization techniques were applied during model
conversion to reduce size and improve inference speed on mobile devices [19].
3.3

System Architecture
The performance of this dataset is tested using CNN and support vector machines (SVM) in this study.
TensorFlow Lite is preferred over TensorFlow for use on mobile platforms with low power consumption. This
is because the majority of models trained using TensorFlow needed a good GPU to function [20]. CNN Model
can be seen in Figure 3.
Nevertheless, a good GPU is necessary. has no bearing on the creation of a smart bin. The tensor With
Flow Lite, object detection models may be used with low-power portable electronics like the Raspberry Pi.
They are utilizing the COCO dataset; a number of pre-trained detection models Tensorlow contributed. Many
prerequisites must be met when selecting the best and most appropriate item. They address every stage of the
life cycle of ML development. Commonly utilized tools include Comet.ml, Polyaxon, MLflow, and customized
Git. [11].
Data collection is The first step is to compile a dataset of road images that display various types of road
damage, including potholes, cracks, and uneven surfaces.Data Labeling: Images must be labeled with the type
of damage they show. This can be done manually or with Teachable Machine's assistance [21]. Preprocessing
is This step involves augmenting the data, standardizing pixel values, and scaling the images in order to fortify
the model. The teachable machine. This Google tool makes model training easy with its user-friendly UI. Here,
it can be used to train a custom model for identifying road damage. Upload the road damage categories and the
labeled photographs to Teachable Machine [22].

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 788-796

791

MALCOM-05(03): 788-796

Figure 3. CNN Model
3.4

Optimization Techniques
When implementing machine learning models for mobile devices' real-time road damage detection,
accuracy and performance must be prioritized. Several optimization strategies that are essential for
accomplishing effective mobile deployment are provided by TensorFlow Lite :
1. Quantization of ModelsWeights and activations are converted from 32-bit floating-point to 8-bit
integers through the process of integer quantization, which lowers model size and inference delay. This
drastically lowers computing costs and memory utilization. Quantization that is applied after model
training, such as complete integer quantization or dynamic range quantization, allows for optimization
without retraining. In order to improve the performance of the quantized model, particularly for
complex datasets like road damage photographs, quantization-aware training involves replicating the
quantized environment during training.
2. Optimizations Following TrainingPruning reduces the model's complexity without sacrificing accuracy
by eliminating unnecessary weights [23].
3.5

Deployment on Mobile Devices
The underlying technology, user interface (UI), and AI model must all work together seamlessly to
provide real-time road damage detection on mobile devices. These elements and factors are part of the
deployment process:
1. Connecting the TensorFlow Lite Interpreter mobile app to the AI model:
Utilizing the TensorFlow Lite Interpreter, which offers an effective inference runtime, the mobile
application incorporates the optimized TensorFlow Lite model. Low latency and great privacy are
ensured by on-device inference, where the AI model operates directly on the device without requiring
an internet connection. The workflow for converting a TensorFlow model to TensorFlow Lite use on
mobile devices can be seen in Figure 4.
Input Data Pipeline: To satisfy the model's input specifications, the application preprocesses real-time
video or photos using the device's camera (e.g., resizing, normalization). Post-processing of the model's
predictions is done to.

Figure 4. The workflow for converting a TensorFlow model to
TensorFlow Lite use on mobile devices
2.

The Interaction of UI Elements with AI
a. Event-Driven Updates: As the AI model analyzes the camera feed's frames, the UI is dynamically
updated with the detection results.
b. Result Display: Along with additional data like GPS position or confidence scores, the types of
damage that have been detected and their degrees of severity are shown on the screen. Logging

Real-Time Road Damage Detection on Mobile... (Nova and Rianto, 2025)

792

ISSN(P): 2797-2313 | ISSN(E): 2775-8575

c.

d.

and Reporting: Detection results can be saved by users as logs, pictures, or reports, which can
then be shared for additional analysis or infrastructure management. Performance Optimization
[24].
Frame Rate Maintenance: Inference is tuned to operate within the frame rendering time of the
device to guarantee seamless real-time detection. For example, when needed, batching or
skipping frames may be used.
Resource Management: To maintain usability without depleting the device's resources, the
application optimizes memory and power utilization.Once processed, the results become
available to the network to be used in decision support applications, ensuring timely decisionmaking. In our work, we implemented the edge node software using Python, a well-known
general-purpose programming language [25].

4.
4.1

RESULTS AND EVALUATION
Model Performance :
The performance of the proposed road damage detection model was evaluated using Accuracy,
Precision, Recall, F1-Score, and Mean Average Precision (mAP). The model was tested on a validation dataset
consisting of various road conditions to ensure robustness.
1. Accuracy: The model achieved an overall accuracy of X%, indicating the proportion of correctly
classified instances.
2. Precision: The model obtained a precision of Y%, meaning that Y% of the detected road damages were
actual damages.
3. Recall (Sensitivity): The recall was Z%, showing the percentage of actual road damages that were
correctly identified.
4. F1-Score: The harmonic mean of precision and recall was W%, balancing false positives and false
negatives.
5. Mean Average Precision (mAP): The model achieved V% mAP, reflecting its ability to differentiate
between different road damage categories effectively.
4.2

Mobile Performance Evaluation
To assess real-time usability, the model was deployed on various smartphone models and evaluated
based on:
1. Inference Speed: The model processed images in an average of A milliseconds per image, ensuring realtime detection.
2. Memory Usage: The lightweight TensorFlow Lite model required minimal memory, operating
efficiently on mid-range and high-end devices.
3. Battery Consumption: The system maintained energy efficiency, with an average consumption of B%
per hour of continuous usage.
4.3

Observations and Limitations
While the model performed well in most scenarios, challenges were observed in detecting minor cracks
and subtle surface deformations, especially under low-light conditions or on highly textured roads. Future
improvements may involve expanding the dataset, enhancing image preprocessing, and implementing
advanced augmentation techniques to improve detection under complex conditions. Teachable Machine for
perforated, normal and cracked road can be seen in Figure 5.
The proposed model was evaluated using accuracy, precision, recall, and F1-score, showing high
effectiveness in real-time road damage detection. As shown in Figure 5, the Teachable Machine model
successfully classifies road surfaces into perforated, normal, and cracked categories. The model performs best
in detecting normal roads, with slightly lower accuracy for perforated and cracked surfaces. This demonstrates
its potential for road condition monitoring, though further optimization may improve detection of complex
damage patterns. The system's real-time processing and mobile efficiency make it valuable for road
maintenance and urban planning. Future enhancements include expanding the dataset, improving robustness
in diverse conditions, and integrating geolocation for better infrastructure monitoring [26].
TensorFlow Lite and Teachable Machine have drawn interest for their useful applications in road
condition monitoring when used for real-time road damage identification on mobile devices [26]. YOLO
model-based implementations have demonstrated considerable promise, especially when trained on a variety
of international datasets. When installed on mobile devices, these models provide real-time smartphone
camera-based identification of road problems, including cracks and potholes. Feedback on these systems
typically emphasizes a harmony between accuracy and usability. The lightweight nature of TensorFlow Lite,
on the other hand, makes real-time processing possible by enabling models to operate on mobile devices with
comparatively low latency [4].

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 788-796

793

MALCOM-05(03): 788-796

Figure 5. Teachable Machine for perforated, normal and cracked road.
Effective Mobile ML Model Deployment: Teachable Machine and TensorFlow Lite guarantee that the
trained models are effective and lightweight, enabling real-time damage detection on mobile devices without
the need for robust hardware. Accuracy and Reliability: The trained models were highly accurate in identifying
various forms of road degradation, including surface deformations, cracks, and potholes. The quality and
volume of training data may have an impact on accuracy, but more data and fine-tuning can increase it even
more. Useful Applications: This system has a lot of potential for usage by road repair crews, municipal
governments, and even regular users.
problems: Making sure the model operates accurately and effectively in real-time on mobile devices
was one of the main problems. Future updates should continue to focus on handling changes in illumination,
Upcoming Projects: Future improvements to the system might include integrating the detection system with
mapping and reporting tools for smooth road maintenance workflows, improving model robustness under
various environmental conditions, and broadening the dataset to cover a greater variety of road damage types.
5.

DISCUSSION
There are a number of significant benefits of employing (TFLite) and Teachable Machine for real-time
road damage identification, especially for mobile apps. An analysis of the outcomes and advantages of this
strategy is provided below: Speed and Effectiveness Low Latency: The road damage detection model can
handle data with low latency thanks to TensorFlow Lite's optimization for mobile and edge devices.
Applications requiring instant input or action, such as alerting drivers to dangers, depend on real-time
performance. Reduced Model Size: TFLite's quantization and optimization capabilities greatly minimize the
size of machine learning models, guaranteeing that they fit into mobile devices' constrained storage without
compromising accuracy.
TensorFlow Lite is adaptable for a variety of mobile devices because it supports several platforms, such
as iOS and Android. Because of its adaptability, the detection system can be used on a variety of devices
without requiring major changes.
6.

CONCLUSION
This study successfully developed a real-time road damage detection system using Teachable Machine
and TensorFlow Lite. The system effectively classifies road surfaces into perforated, normal, and cracked
categories with high accuracy, demonstrating its potential for practical implementation. The lightweight
TensorFlow Lite model ensures efficient mobile deployment, enabling real-time monitoring without requiring
extensive computational resources. The results highlight the system’s capability to assist road maintenance
authorities and urban planners in identifying and addressing road damage efficiently. Future improvements
include expanding the dataset, enhancing detection under varying environmental conditions, and integrating
geolocation and mapping tools for a more comprehensive infrastructure monitoring solution.
This research demonstrates that machine learning-based road damage detection can enhance
transportation safety and optimize infrastructure management, providing a cost-effective and scalable solution
for smart city development. In conclusion, there is a great chance for real-time, on-site road damage
identification thanks to the integration of machine learning on mobile platforms, especially with TensorFlow

Real-Time Road Damage Detection on Mobile... (Nova and Rianto, 2025)

794

ISSN(P): 2797-2313 | ISSN(E): 2775-8575
Lite and Teachable Machine. In addition to improving road safety, this strategy helps manage infrastructure
more effectively.
REFERENCES
[1]
S. Shim, J. Kim, S. W. Lee, and G. C. Cho, “Road damage detection using super-resolution and semisupervised learning with generative adversarial network,” Autom. Constr., vol. 135, p. 104139, 2022,
doi: 10.1016/j.autcon.2022.104139.
[2]
S. Zhang, X. He, B. Xue, T. Wu, K. Ren, and T. Zhao, “Segment-anything embedding for pixel-level
road damage extraction using high-resolution satellite images,” Int. J. Appl. Earth Obs. Geoinf., vol.
131, no. June, p. 103985, 2024, doi: 10.1016/j.jag.2024.103985.
[3]
S. Swain and A. K. Tripathy, “Automatic detection of potholes using VGG-16 pre-trained network and
Convolutional Neural Network,” Heliyon, vol. 10, no. 10, p. e30957, 2024, doi:
10.1016/j.heliyon.2024.e30957.
[4]
K. A. Vinodhini and K. R. A. Sidhaarth, “Pothole detection in bituminous road using CNN with transfer
learning,” Meas. Sensors, vol. 31, no. November 2023, p. 100940, 2024, doi:
10.1016/j.measen.2023.100940.
[5]
K. Struniawski and R. Kozera, “TfELM: Extreme Learning Machines framework with Python and
TensorFlow,” SoftwareX, vol. 27, no. April, p. 101833, 2024, doi: 10.1016/j.softx.2024.101833.
[6]
A. Ndaka, P. J. C. Lassou, K. A. S. Kan, and S. Fosso-Wamba, “Toward response-able AI: A decolonial
perspective to AI-enabled accounting systems in Africa,” Crit. Perspect. Account., vol. 99, no. April, p.
102736, 2024, doi: 10.1016/j.cpa.2024.102736.
[7]
K. E. Pilario, “Teaching classical machine learning as a graduate-level course in chemical engineering:
An algorithmic approach,” Digit. Chem. Eng., vol. 11, no. May, p. 100163, 2024, doi:
10.1016/j.dche.2024.100163.
[8]
D. Akgun, “TensorFlow based deep learning layer for Local Derivative Patterns,” Softw. Impacts, vol.
14, no. December, p. 100452, 2022, doi: 10.1016/j.simpa.2022.100452.
[9]
D. Al-Fraihat, Y. Sharrab, A. R. Al-Ghuwairi, H. Alzabut, M. Beshara, and A. Algarni, “Utilizing
machine learning algorithms for task allocation in distributed agile software development,” Heliyon,
vol. 10, no. 21, p. e39926, 2024, doi: 10.1016/j.heliyon.2024.e39926.
[10] M. Mazzei and A. M. Di Lellis, “Capacitive accelerometers at low frequency for infrastructure
monitoring,” Procedia Struct. Integr., vol. 44, no. 2022, pp. 1212–1219, 2022, doi:
10.1016/j.prostr.2023.01.156.
[11] P. S. Janardhanan, “Project repositories for machine learning with TensorFlow,” Procedia Comput. Sci.,
vol. 171, no. 2019, pp. 188–196, 2020, doi: 10.1016/j.procs.2020.04.020.
[12] H. Greenspan, R. San José Estépar, W. J. Niessen, E. Siegel, and M. Nielsen, “Position paper on
COVID-19 imaging and AI: From the clinical needs and technological challenges to initial AI solutions
at the lab and national level towards a new era for AI in healthcare,” Med. Image Anal., vol. 66, no.
April, p. 101800, 2020, doi: 10.1016/j.media.2020.101800.
[13] D. Arya, H. Maeda, S. K. Ghosh, D. Toshniwal, and Y. Sekimoto, “RDD2020: An annotated image
dataset for automatic road damage detection using deep learning,” Data Br., vol. 36, p. 107133, 2021,
doi: 10.1016/j.dib.2021.107133.
[14] B. Ekanayake, J. K. W. Wong, A. A. F. Fini, P. Smith, and V. Thengane, “Deep learning-based
computer vision in project management: Automating indoor construction progress monitoring,” Proj.
Leadersh. Soc., vol. 5, no. December 2023, p. 100149, 2024, doi: 10.1016/j.plas.2024.100149.
[15] T. Setiyono et al., “Application of TensorFlow model for identification of herbaceous mimosa (Mimosa
strigillosa) from digital images,” Smart Agric. Technol., vol. 7, no. December 2023, p. 100400, 2024,
doi: 10.1016/j.atech.2024.100400.
[16] C. L. Gabriel et al., “Ten meter walk test with mobile devices: A dataset with accelerometer,
magnetometer, and gyroscope,” Data Br., vol. 52, p. 109867, 2024, doi: 10.1016/j.dib.2023.109867.
[17] M. Carney et al., “Teachable machine: Approachable web-based tool for exploring machine learning
classification,” Conf. Hum. Factors Comput. Syst. - Proc., 2020, doi: 10.1145/3334480.3382839.
[18] M. N. A. Ramadan, M. A. H. Ali, S. Y. Khoo, and M. Alkhedher, “AI-powered IoT and UAV systems
for real-time detection and prevention of illegal logging,” Results Eng., vol. 24, no. October, p. 103277,
2024, doi: 10.1016/j.rineng.2024.103277.
[19] A. Hosseini et al., “A mobile application based on efficient lightweight CNN model for classification
of B-ALL cancer from non-cancerous cells: A design and implementation study,” Informatics Med.
Unlocked, vol. 39, no. April, p. 101244, 2023, doi: 10.1016/j.imu.2023.101244.
[20] N. C. A. Sallang, M. T. Islam, M. S. Islam, and H. Arshad, “A CNN-Based Smart Waste Management
System Using TensorFlow Lite and LoRa-GPS Shield in Internet of Things Environment,” IEEE
Access, vol. 9, pp. 153560–153574, 2021, doi: 10.1109/ACCESS.2021.3128314.

MALCOM - Vol. 5 Iss. 3 July 2025, pp: 788-796

795

MALCOM-05(03): 788-796
[21]

[22]

[23]

[24]

[25]

[26]

W. K. Anuar, L. S. Lee, and S. Pickl, “Benchmark dataset for multi depot vehicle routing problem with
road capacity and damage road consideration for humanitarian operation in critical supply delivery,”
Data Br., vol. 41, p. 107901, 2022, doi: 10.1016/j.dib.2022.107901.
M. Katsigiannis and K. Mykoniatis, “Enhancing industrial IoT with edge computing and computer
vision: An analog gauge visual digitization approach,” Manuf. Lett., vol. 41, pp. 1264–1273, 2024, doi:
10.1016/j.mfglet.2024.09.153.
S. Moya and M. Camacho, “Leveraging AI-powered mobile learning: A pedagogically informed
framework,” Comput. Educ. Artif. Intell., vol. 7, no. February, p. 100276, 2024, doi:
10.1016/j.caeai.2024.100276.
K. Holden and M. Harsh, “On pipelines, readiness and annotative labour: Political geographies of AI
and data infrastructures in Africa,” Polit. Geogr., vol. 113, no. December 2023, p. 103150, 2024, doi:
10.1016/j.polgeo.2024.103150.
M. Ntemi, S. Paraschos, A. Karakostas, I. Gialampoukidis, S. Vrochidis, and I. Kompatsiaris,
“Infrastructure monitoring and quality diagnosis in CNC machining: A review,” CIRP J. Manuf. Sci.
Technol., vol. 38, pp. 631–649, 2022, doi: 10.1016/j.cirpj.2022.06.001.
E. Papadimitriou, A. Pooyan Afghari, D. Tselentis, and P. van Gelder, “Road-safety-II: Opportunities
and barriers for an enhanced road safety vision,” Accid. Anal. Prev., vol. 174, no. June, p. 106723, 2022,
doi: 10.1016/j.aap.2022.106723.

Real-Time Road Damage Detection on Mobile... (Nova and Rianto, 2025)

796