IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 2, April 2025, pp. 1430~1440

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i2.pp1430-1440  1430

Detecting road damage utilizing retinanet and mobilenet models

on edge devices

Haniah Mahmudah1,2, Aulia Siti Aisjah2, Syamsul Arifin1, Catur Arif Prastyanto3
1
Department of Electrical Engineering, Politeknik Elektronika Negeri Surabaya, Surabaya, Indonesia
2
Department of Physics Engineering, Faculty of Industrial and Systems Engineering, Institut Teknologi Sepuluh Nopember,
Surabaya, Indonesia
3
Department of Civil Engineering, Faculty of Civil, Planning, and Geo Engineering, Institut Teknologi Sepuluh Nopember,
Surabaya, Indonesia

Article Info ABSTRACT

Article history: A particular form of road digitalization produces a system that detects road
damage automatically and in real time, employing the device to detect road
Received Mar 26, 2024 damage as an edge device. The application of RetinaNet152 and
Revised Oct 25, 2024 MobileNetV2 models for road damage detection on edge devices
Accepted Nov 14, 2024 necessitates a trade-off between high system performance and efficiency.
Currently, edge devices have limited storage. In this paper, we explore how
tuning hyperparameters with batch size and several optimizers improves
Keywords: system performance on RetinaNet152 and MobileNet models, as well as
how they are implemented on edge devices. After tuning hyperparameters in
Edge device the batch size of the optimizer, the Adam optimizer displayed enhanced
Inference time performance with mean average precision (mAP), average recall (AR), and
MobileNet model F1-score. This implies a positive impact on overall model performance. The
RetinaNet model MobileNetV2 model's hyperparameter tuning technique significantly
Tuning hyperparameters improves performance, resulting in faster inference times and overall system
performance. This demonstrates that the MobileNetV2 model could be used
directly on edge devices to identify road damage. However, the
RetinaNet152 model has a lower inference time, which cannot be deployed
directly to edge devices. The RetinaNet152 model can be deployed on edge
devices; however, a technique for speeding up inference time is essential.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Aulia Siti Aisjah
Department of Physics Engineering, Faculty of Industrial and Systems Engineering
Institut Teknologi Sepuluh Nopember
ITS Campus, Sukolilo, Surabaya, 60111, Indonesia
Email: [email protected]

1. INTRODUCTION
A particular form of road digitalization produces a system that detects road damage automatically
and in real time, employing the device to detect road damage as an edge device. The implementation of
digitalization on roads through automation, artificial intelligence (AI), and digital information flow results in
enhanced efficiency, lower operating costs, and improved road services. A particular application of
digitalization on roads is in road maintenance programs that support smart cities [1]–[3]. Several studies have
been conducted to decet road damage utilizing AI technologies, camera sensors and various image data
processing approaches. Some of the research with approach methodologies include image-based image
processing techniques [4]–[9], machine learning (ML) model [10]–[14], and deep learning method [15]–[27].
This investigation employs image processing techniques for a variety of applications, including the
detection of road defects. Potholes are a type of road defect that can be detected using classic image

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

Int J Artif Intell ISSN: 2252-8938  1431

processing techniques. This technique needs numerous stages to achieve high accuracy, including manually
extracting image features and tweaking image processing parameters. The speed of image frames varies
depending on road conditions [4]. The ML model includes a data processing stage that manually pulls
features to increase accuracy, which requires a significant number of computational resources. Several
studies have used deep convolutional neural network (CNN) approaches to automate the process of feature
extraction and categorization at the same time [18], [26].
A particular application of the CNN model for road problem identification on edge devices
necessitates a trade-off between high accuracy and high efficiency [13]. The CNN models are highly accurate
in terms of system performance recall, precision, high accuracy, and low loss. This necessitates CNN models
with massive architecture and competitive computing costs that are inefficient. As a result, it requires a
simpler CNN model configuration setting while maintaining good accuracy. The CNN models are highly
efficient in that they can detect road faults on edge devices with fast inference times. Because the CNN
model on the edge device has limited storage, inference time must be taken into consideration. The CNN
models have fast inference times, perform well on simpler model configurations, but are difficult to obtain
high accuracy. To achieve excellent system performance, CNN model parameters must be carefully selected.
When selecting detection models, there is no conclusive answer as to which model has the highest
performance, but one must make decisions based on demands [13], [14].
In order to produce a system that detects road damage automatically and in real time, CNN model
research must be developed. The development study in road damage detection enhances system performance
by utilizing one- and two-stage CNN models. Some research on road damage detection using CNN includes
detection techniques using CNN models with semi-supervised learning using pseudo-labels [15], CNN
one-stage detector model architecture namely InceptionV2 and MobileNet on personal computer (PC) [16],
YOLO namely Tiny-YOLOv2, darknet neural networks, YOLOv3, Tiny-YOLOv3 and YOLOv4 [20], [21],
YOLOv5 on smartphones [22], YOLOv3, YOLOv2 and TinyV3 models [23], YOLOv5, two variations of
YOLOR, and faster R-CNN with five different backbones namely ResNet50, VGG16, MobileNetV2,
InceptionV3, and proposed modified VGG16 (MVGG16) [24]. The results of road damage detection research
implementing the CNN model include system performance features such as accuracy, mean average
precision (mAP), average recall, and the F1-score.
One effort to increase accuracy in CNN models is tuning hyperparameters relating to network
structure and training. The tuning hyperparameters consider network structure, namely kernel size, width,
and depth. Researchers examined three network designs of a CNN one-stage detector model (small, medium,
and large), as well as a combination of hyperparameter tuning and activation function adjustments. System
performance can be improved by implementing a CNN model network architecture and adjusting
hyperparameters [17], [18], [26], [27]. Several research have achieved high accuracy by employing an
optimizer to change hyperparameters on CNN models. To obtain good performance, several researchers on
object identification employ one or two optimizers as tuning hyperparameters on pretrained CNN models
[28]–[31]. However, none of the CNN model research has been deployed on edge devices for detecting road
damage.
Several road defect investigations were conducted utilizing CNN models on edge devices, resulting
in systems that detect road faults automatically and in real time. Maeda et al. [16] employed MobileNet on a
smartphone, achieving a system performance of 71% and an inference time of 1.5 seconds. Other research
applies the single shot multi-box detector (SSD)-MobileNet model with a batch size of 64 to NVIDIA Jetson
Nano devices with accelerators [32] and unmanned aerial vehicles (UAV) using Raspberry Pi [33]. Other
researchers created MobileNetV2 with hyperparameter learning rate and batch size settings using NVIDIA
Jetson Nano, which has a higher mAP of 0.0869 and a lower total loss training of 0.6028 compared to SSD
Resnet50V1 [34], [35]. Based on the results of that investigation, the CNN model is merged into hardware
devices that have been developed but still require additional research to increase system performance and
accelerate inference time, as shown in Table 1.
The challenge of detecting road damage on edge devices is to do it automatically and in real time as
the edge device moves at a specific speed. This necessitates the deployment of a CNN mode on an edge
device capable of overcoming this limitation. The RetinaNet152 model detects objects in images by
combining anchor boxes with feature pyramid networks. RetinaNet152 refers to the number of layers in the
backbone network, which in this case is ResNet-152. ResNet-152 is well-suited for road damage detection
since it can recognize and categorize objects with high accuracy and speed, which is critical for real-time
applications. Based on the advantages of RetinaNet's research on road damage by modifying the backbone
[36], it employs RetinaNet152 with hyperparameter optimizer adjustment [37]. The RetinaNet152 model has
advantages; however, it has not been deployed on edge devices for road damage detection equipment.
According to Table 1, the MobileNetV2 model [34] must increase system performance by modifying
hyperparameters using an optimizer.

Detecting road damage utilizing retinanet and mobilenet models on edge devices (Haniah Mahmudah)
1432  ISSN: 2252-8938

Table 1. The CNN models on road damage

Previous studies Model CNN Object detection Performance system Edge device
[16] MobileNet Eight types of road damage Recall, precision, inference time Smartphone
[32] MobileNetV2 Potholes, longitudinal cracks, mAP NVIDIA Jetson Nano
alligator cracks
[33] MobileNet V1 Pavement Accuracy, FPS Raspberry Pi
[34] MobileNetV2 Potholes, longitudinal cracks, mAP, AR NVIDIA Jetson Nano
alligator cracks
Proposed RetinaNet152 Potholes, longitudinal cracks, mAP, AR, F1-score, size model, NVIDIA Jetson Nano
system MobileNetV2 alligator cracks inference time, FPS

To address a research gap in this study, we present a system that improves the system performance
of the RetinaNet152 and MobilenetV2 models in a road damage detection system on edge devices, thereby
improving system performance and inference time. The research results with optimization produced good
system performance. Road damage detection systems for edge devices with limited memory and mobile data
retrieval use the RetinaNet152 and MobilenetV2 models. Larger batch sizes can help speed up the training
process by allowing the model to process more data concurrently. This can lead to faster convergence and
improved performance. A larger batch size can help reduce gradient variance, hence improving model
stability and accuracy [37]. However, it is particularly important to note that raising the batch size and
optimizer can result in higher memory needs and the need for more processing resources. To increase system
performance and implementation on edge devices with environmental variations, the road damage detection
system needs to simulate batch sizes and optimizer parameters on RetinaNet152 and MobileNet models.
The primary contributions of this research to the development of road defect detection systems are
described as follows:
‒ The development of RetinaNet152 [38] and MobileNetV2 [34] models for road damage is based on tuning
hyperparameters: batch size and optimizers to improve performance system mAP, AR, and F1-score.
‒ Edge-device deployment of the RetinaNet152 and MobileNetV2 models for real-time road damage
detection. This research uses a road damage system and a road damage detection system.
‒ Comparing the analysis systems performance of the Retinanet152 and MobileNetV2 models, inference
times, frames per second, and size model when deployed on edge devices.
The paper has four parts. Section 1 introduction focuses on the topic and research goal.
Section 2 covers methods. Section 3 contains results and discussions: i) the system's performance
tuning hyperparameter, ii) deploying the RetinaNet152 and MobileNetV2 models on edge device, and
iii) comparation system performance model CNN. Finally, section 4 presents ideas for future research.

2. METHOD
Several issues arise when gathering and identifying road defects when moving with edge devices.
Potholes, alligator cracks, and longitudinal fractures are produced by capturing road damage statistics on
edge devices in real time and while in motion. Under certain conditions, the recoverable cracks are very
small and unclear. This is due to the tiny size of the cracks and the fluctuation of road damage object
detection in the surrounding environment, which makes it difficult to discern road damage in the form of
cracks. Furthermore, road damage datasets in the form of fractures result in nearly identical color shifts
between the detected items and the surrounding surroundings. This is the challenge of gathering moving data
on road damage on an edge device. A CNN model is required for good performance and can be deployed on
edge devices.
The problem concerning road damage detection devices in the form of mobile edge devices exists
because they have limited computation and storage capabilities. Thus, the CNN model deployed on edge
devices must be considered. Some studies employ the CNN one-stage detector MobileNetV2 [16], [32], [34]
and RetinaNet152 [38] models installed on edge devices. The RetinaNet152 model has been selected for
current research to provide real-time detection and classification with hyperparameter tuning to increase
system performance, which includes mAP, recall, and F1-score. This research employs the RetinaNet152
model, with further advancements in road damage identification systems through batch size and real-time
video testing.
The RetinaNet152 and MobileNetV2 models are employed in a variety of steps, including dataset
retrieval, pre-processing, annotation, batch size and optimizer modeling, and system performance evaluation.
The data analysis parameters are mAP, AR, and F1-score. The RetinaNet152 and MobileNetV2 models were
converted to TFlite and deployed on an edge device using the NVIDIA Jetson Nano, allowing the inference
time to be measured. Validation testing of the RetinaNet152 and MobileNetV2 road damage categorization

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1430-1440

Int J Artif Intell ISSN: 2252-8938  1433

models involved validating video and then evaluating system performance. Figure 1 shows the modeling
procedure for RetinaNet152 and MobileNetV2 models, including hyperparameter tuning, performance
analysis, and deployment on edge devices.
This dataset retrieval combines a collection of road damage images from the East Java Public Works
Office's Directorate of Highways survey findings with data obtained directly. Datasets were collected on
provincial highways in East Java utilizing the road damage tool system [15]. The dataset retrieval images
returned 11,176 images.
Pre-processing the dataset is required to equalize the image's size and format, allowing for faster
performance and improved model performance. The Windows utility converts image data into 640×640-pixel
images. The dataset is annotated to identify and describe characteristics based on item classification. The
labelImg tool annotates the entire dataset. This annotation process selects the type of road damage
classification. The current research utilizes four labeling datasets for road damage classification: potholes
(L00), longitudinal cracks (R02), alligator cracks (R03), and shadows (00). The dataset labeling method
addresses the demand for road damage data based on the asphalt road condition survey form provided by the
Department of Public Works, Directorate General of Highways in Indonesia.

Figure 1. The RetinaNet152 and MobileNetV2 models use hyperparameter tuning

The road damage detection procedure makes use of deep learning technology, specifically the
RetinaNet152 and MobileNetV2 models, which have already been trained. This decision is taken to obtain
peak performance while drastically reducing training time compared to starting from scratch.
The RetinaNet152 and MobileNetV2 models are developed using the TensorFlow object detection API, an
open-source framework based on TensorFlow. This API simplifies the development, training, and
deployment of object detection models. The TensorFlow object detection API includes the model Zoo, a
collection of pre-trained models with various architectures. These models were trained on many datasets,
including COCO, KITTI, and Open Image. Using these pre-trained models promotes transfer learning and
allows customization to meet a variety of operational objectives. This approach provides a solid foundation
for the road damage classification system.

Detecting road damage utilizing retinanet and mobilenet models on edge devices (Haniah Mahmudah)
1434  ISSN: 2252-8938

This research analyzes the use of optimizers like momentum, RMSprop, and Adam on the
RetinaNet152 and MobileNetV2 models. The momentum optimizer can help to overcome the issue of
delayed convergence and accelerate model training. The momentum optimizer assists in filling local gaps and
enhances converge in the right direction. momentum optimizer parameter equations as in (1) and (2) [38].

𝑣(𝑡) = 𝛽. 𝑣(𝑡−1) + (1 − 𝛽) ⋅ 𝛻𝐹. (𝑊𝑡 ) (1)

𝑊(𝑡−1) = 𝑊𝑡 − 𝛼 ⋅ 𝑣(𝑡) (2)

Where t is the parameter at the t-iteration, α is the learning rate, ∇F(Wt) is the gradient of the cost function
with respect to the parameter at iteration t. RMSprop is one of the optimization algorithms that maintains the
squared gradient average for each weight. The equation RMSprop is shown in (3) [38].

𝑀𝑒𝑎𝑛𝑆𝑞𝑢𝑎𝑟𝑒(𝑤,𝑡) = ρ x 𝑀𝑒𝑎𝑛𝑆𝑞𝑢𝑎𝑟𝑒(𝑤,𝑡−1) + 0.1( ∂E ∂w(𝑡) )2 (3)

Adam optimizer in (4) [38].

𝑚𝑡
𝑊𝑡 = 𝑊𝑡−1 − α( + ε)) (4)
√𝑣𝑡

Where g is gradient, m is first moment, v is second moment, 𝛽1, 𝛽2 are exponential decay rates, 𝛼 is learning
rate, W the parameter is the weight. During the optimization phase for the RetinaNet152 and MobileNetV2
models, trials are carried out with various learning rates and optimizer settings. To determine the test
outcomes, use the evaluation matrix defined by the two variables in (5) and (6) [19], [38].
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (5)
𝑇𝑃+𝐹𝑃

𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = (6)
𝑇𝑃+𝐹𝑁

The total loss is the result of an incorrect forecast, while the loss is a numerical value that measures
the degree of inaccuracy in the model's prediction. Precision is defined as the ratio of correct positive
predictions to the total number of positive instances predicted; it is simply the true positives (TP) divided by
total detections that are true positives (TP) and false positives (FP). Recall quantifies the ratio of successfully
predicted actual positives, specifically the true positives (TP), to all ground truth values that are true positives
(TP) and false negative (FN). The F1-score is an evaluation metric that determines the accuracy of the
RetinaNet152 and MobileNetV2 models. It brings together a model's precision and recall scores as shown in
(7) [19]. The accuracy statistic calculates the number of times the RetinaNet152 and MobileNetV2 models
predicted correctly over the full dataset.

(𝑃𝑟𝑒𝑐𝑖𝑠𝑜𝑛 𝑥 𝑅𝑒𝑐𝑎𝑙𝑙)
𝐹1 𝑠𝑐𝑜𝑟𝑒 = 2 𝑥 (7)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

Validation testing of the road damage classification model was performed using video samples
lasting around 40 seconds. Real-time video capture, which is based on the road damage detection system,
records the road segment in motion using a camera. During video capture, the camera is designed to be
mounted on the car's front bonnet at an angle of around 28 degrees to the road. The weather conditions
during data collection were described as partially cloudy. Video validation testing of the RetinaNet152 and
MobileNetV2 models was carried out on the Jupyter Notebook server.

3. RESULTS AND DISCUSSION

This section presents the research achievements of the proposed system from the RetinaNet152 and
MobileNetV2 models. The system performance analysis results of RetinaNet152 and MobileNetV2 models
include hyperparameter tuning system performance analysis, video validation, and deployment results on
edge devices, specifically NVIDIA Jetson 4 GB. In addition, comparative analysis of mAP, AR, and F1-score
system performance, inference time, and previous work and the proposed system. Specifically, the stages are
stated as follows.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1430-1440

Int J Artif Intell ISSN: 2252-8938  1435

3.1. The system's performance tuning hyperparameter

The influence of batch size and optimizer could affect total loss and system performance. Table 2
shows the relationship between batch size and optimizer for the proposed RetinaNet152 and MobileNetV2
models. The simulation results of the RetinaNet152 model show that increasing the batch size could develop
a more stable gradient estimate, allowing the momentum optimizer to produce a more consistent gradient
descent direction, resulting in less loss. The RetinaNet152 model uses the momentum optimizer, and batch
size 4 loses 0.573, whereas batch size 8 loses 0.225. The RMSProp optimizer adjusts the learning rate for
each parameter based on the most recent observation and the average gradient. However, the RMSProp
optimizer has a method for managing important gradient variance, resulting in a loss of 0.802 for batch size 4
to 1.088 for batch size 8, as compared to momentum.

Table 2. The system performance of RetinaNet152 and MobileNetV2 models

Proposed Parameter mAP mAP F1-score F1-score
Loss (AR)
Sytem (Optimizer, Batch size) (@IoU 0.5-0.95) (@IoU 0.5) (@IoU 0.5-0.95) (@IoU 0.5)
Momentum, 4 0.573 0.277 0.406 0.449 0.343 0.426
RMSprop, 4 0.802 0.235 0.454 0.457 0.310 0.455
RetinaNe
Adam, 4 0.264 0.351 0.457 0.499 0.412 0.477
t152
Momentum, 8 0.225 0.334 0.455 0.423 0.373 0.438
Model
RMSprop, 8 1.088 0.256 0.477 0.446 0.325 0.461
Adam, 8 0.220 0.349 0.469 0.410 0.377 0.438
Momentum, 4 0.326 0.292 0.563 0.467 0.349 0.451
RMSprop, 4 1.970 0 0 0 0 0
MobileN
Adam, 4 0.656 0.208 0.578 0.408 0.338 0.462
etV2
Momentum, 8 0.668 0.201 0.561 0.368 0.342 0.450
Model
RMSprop, 8 2.010 0 0 0 0 0
Adam, 8 0.320 0.298 0.606 0.476 0.322 0.481

Based on the second estimate of the gradient moments, Adam's optimizer adjusts the learning rate
adaptively to modify the batch size. Simulation results for the RetinaNet152 model show that a very small
batch size of 4 could produce an unstable estimate of the gradient moment, resulting in a loss of 0.264, but a
batch size of 8 makes the Adam optimizer more stable, resulting in a loss of just 0.220. The batch size 8 and
Adam optimizer scenario had a higher performance system with mAP 0.469 and AR 0.410, resulting in the
highest F1-score value of 0.438. The scenario had a less overall loss than the other possibilities, which is
advantageous.
Simulation results for the MobileNetV2 model demonstrate that raising the batch size can lead to
more stable gradient estimation, allowing the momentum optimizer to lower the loss of 0.326 at batch size 4
and 0.668 at batch size 8. However, the RMSProp optimizer adapts the learning rate to the average gradient.
However, the RMSProp optimizer generates a loss higher than one; therefore, it cannot attain system
performance. Adam's optimizer with adaptive produces a more constant gradient descent direction, which
leads to less loss. Adam's optimizer has a lower loss of 0.656 at batch size 4 and 0.320 at batch size 8 than
momentum and RMSprop. The results of system performance according to the MobileNetV2 model, batch
size 8, and Adam optimizer scenario produced a higher performance system with mAP 0.606 and AR 0.476,
resulting in the highest F1-score value of 0.481.
The batch size indicates the number of samples utilized in a single training iteration before the
Retinanet152 and MobileNetV2 model weights are adjusted. The RetinaNet152 model illustrates that raising
the batch size reduces loss in the momentum and Adam optimizers since the resulting gradient is an average
of more data, except in RMSprop, where increasing the batch size increases loss. The MobileNetV2 model
reveals that increasing the batch size on the momentum optimizer and RMSprop increases the loss. The
MobileNetV2 model minimizes loss on the Adam optimizer while maintaining acceptable system
performance (mAP, AR, and F1-score). Overall, the RetinaNet152 and MobileNetV2 models have the lowest
loss and the best system performance (mAP, AR, and F1-score) in batch size 8 and Adam optimizer.
When implementing the Retinanet152 and MobileNetV2 models on edge devices with limited
memory, it is important to keep in mind that if the batch size exceeds the available memory capacity,
memory overflow and performance challenges may occur. In road defect detection systems, the Retinanet152
and MobileNetV2 models are used on edge devices with limited memory. This is an important issue to
consider while selecting the RetinaNet152 and MobileNetV2 models on settings such as batch size and
optimizer. It is critical to undertake procedures and cross-validation to determine the best combination of the
RetinaNet152 and MobileNetV2 models based on parameters such as batch size and optimizer.
Our findings suggest that the proposed systems use RetinaNet152 and MobileNetV2, and that
adding an adaptive optimizer, specifically Adam, can result in improved system performance. Some of the
Detecting road damage utilizing retinanet and mobilenet models on edge devices (Haniah Mahmudah)
1436  ISSN: 2252-8938

issues faced by this system are road damage datasets with almost equal variance between detection objects
and the surrounding environment, small road damage detection objects, and object retrieval while in motion.
This results in inadequate system performance, enhancing this issue of real-time road damage detection
research using the device. As a result, more investigation is required to increase the performance of a more
optimal system by including many approaches into the one-stage CNN model used on edge devices.

3.2. Video validation and deploying RetinaNet152 and MobileNetV2 models on edge device
The evaluation of the RetinaNet152 and MobileNetV2 models is based on classification results,
which have a considerable impact on road damage calculations. The test seeks to differentiate the model's
predictions when running on video samples to see if they match the intended objects. Confusion evaluators
use metrics such as TP, FP, and FN. TP indicates that objects within the bounding box were accurately
identified based on their anticipated label. FP implies that objects were mistakenly identified within the
bounding box, whereas FN shows that things should have been discovered but were not. Precision, recall, and
F1-score are among the measures obtained from confusion metrics.
Precision assesses the ability to make accurate positive predictions. Recall (sensitivity or true
positive rate) measures the model's ability to detect all cases that should be positive. The F1-score illustrates
the trade-off between precision and recall, demonstrating the balance of these two criteria. Table 3 shows the
results of the validation testing of the RetinaNet152 and MobileNetV2 model's detection with video samples.
Table 3 presents the detection results for the RetinaNet152 and MobileNetV2 models based on
video samples. The average values for the system performance parameters precision, recall, and F1 score are
the same for the RetinaNet152 and MobileNetV2 models. Based on the video validation results, Figures 2
exhibit the classification and detection capabilities of the RetinaNet152 and MobileNetV2 models to detect
road damage systems.

Table 3. The performance of video validation

Proposed sytem Type TP FP FN Ground truth Precision Recall F1-score
RetinaNet152 model L00 29 14 1 30 0.67 0.97 0.79
R02 3 2 1 4 0.60 0.75 0.67
R03 12 7 1 13 0.63 0.92 0.75
00 3 1 4 7 0.75 0.43 0.55
Average 0.66 0.77 0.67
MobileNetV2 model L00 32 10 2 28 0.76 0.94 0.84
R02 5 2 1 3 0.71 0.83 0.76
R03 13 5 2 12 0.72 0.86 78
00 2 2 4 7 0.5 0.33 0.4
Average 0.67 0.74 0.69

Figure 2. Prediction results RetinaNet152 and MobileNetV2 models, parameter batch size 8, and
Adam optimizer

The implementation testing of the road fault detection system focuses on analyzing the CNN
model's performance on NVIDIA Jetson Nano 4 GB used as an edge device [32]. In this research, utilizing
the RetinaNet152 [38] and MobileNetV2 models, adjusting hyperparameters such as batch size and an
optimizer are built to increase system performance. Performance parameters include average frame per
second (FPS) and average inference time. Table 4 shows the results of deploying the RetinaNet-152 and
MobileNetV2 models on the edge device with modifications in hyperparameter tuning parameters such as
batch size and optimizer.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1430-1440

Int J Artif Intell ISSN: 2252-8938  1437

Table 4. Comparison of inference time RetinaNet152 and MobileNetV2 models

Model FPS Inference Time (s) Model (Mb)
RetinaNet152 0.05 19.48 70
MobileNetV2 3.50 0.28 3.7

Table 4 shows that the performance of the RetinaNet-152 model on the NVIDIA Jetson Nano 4 GB
has direct field-testing results with a very low FPS of 0.05. This demonstrates that adjusting the batch size
and optimizer hyperparameters can increase system performance (mAP, recall, and F1-score) but also speed
up inference time. Inference time with smaller values and model sizes demonstrates that the RetinaNet 152
model performs better at detecting road damage. However, deployment findings on the edge device model
RetinaNet152 showed a lower FPS on the Adam optimizer with a batch size of 8 versus 4. Larger batch sizes
have the potential to accelerate the training process, resulting in faster inference time. However, big batches
require resources, particularly when implemented on edge devices with limited memory. Among
the cases, Adam optimizer with batch size 8 produced better parameter values: FPS 0.05, inference time
19.4780 seconds, model size 70 MB.
Based on research conducted with the MobileNetV2 model [34], this research developed the
MobileNetV2 model, which was deployed utilizing an NVIDIA Jetson 4 GB. The MobileNetV2 model has
an inference time 0.28 seconds faster than RetinaNet152 and an FPS of 3.5, giving it better performance than
RetinaNet152. Table 4 demonstrates how the MobileNetV2 model, which has faster inference, may be
utilized to detect road damage on edge devices.
The general RetinaNet152 model cannot be employed directly on edge devices; however, methods
are required to reduce inference time. This is due to the RetinaNet152 model's far more sophisticated
architecture than MobileNetV2. RetinaNet152 contains more layers and parameters to accommodate the need
for high precision in object detection tasks involving multiple objects at varying scales and features.
This means that RetinaNet152 needs more processing resources to perform this inference. If the
RetinaNet152 model is used in research on edge devices, methods such as quantization, prunning, and
knowledge distillation must be used to reduce inference time.

3.3. Comparation system performance model CNN

Table 2 shows the system performance results of the RetinaNet152 and MobileNetV2 models with
respect to all the minimum loss parameters. The results of mAP of 0.469; AR of 0.410 and F1-score of 0.377
for the RetinaNet152 model parameters batch size 8 and Adam optimizer are obtained. Table 4 shows the
efficiency of the RetinaNet152 model with batch size 8 and Adam optimizer in terms of inference time of
19.48 seconds, FPS of 0.05, and model size of 70 MB. The MobileNetV2 model parameters batch size 8 and
Adam optimizer yielded the highest F1-score value of 0.481, with mAP 0.606 and AR 0.476, respectively.
The MobileNetV2 model with batch size 8 and Adam optimizer is efficient in terms of inference time of 0.28
seconds, FPS of 3.50, and model size of 3.7 MB. The RetinaNet152 and MobileNetV2 models with test
results have the best system performance when compared to the findings of another investigation, as shown
in Table 5.

Table 5. The system performance on edge device

Previous studies Model CNN System performance Edge device
[16] MobileNet Recall 0.71; precision 0.77; inference time 1.5 s. Smartphone
[32] MobileNetV2 mAP 0.22 NVIDIA Jetson Nano
[33] MobileNet V1 Accuracy 0.60; frames per second 1.2 s Raspberry Pi
[34] MobileNetV2 mAP 0.086; AR 0.241 NVIDIA Jetson Nano
Proposed system RetinaNet152 mAP 0.469; AR 0.410; F1-score 0.377; size model 70 MB; NVIDIA Jetson Nano
inference time 19,4780 second; FPS 0.05
Proposed system MobileNetV2 mAP 0.606; AR 0.476; F1-score 0.481; size model 3.7 MB; NVIDIA Jetson Nano
inference time 0.28 second; FPS 3.5

This study investigates the effects of batch size and the optimizer on CNN models, specifically the
RetinaNet152 and MobileNetV2 models, which are deployed on edge devices to detect road defects.
Although earlier research investigated the influence of hyperparameter CNN models on system performance,
little has been performed to specifically examine the effect of batch size and the optimizer. This result study
is compared to the MobileNetV2 model [32], [34] research using learning rate and batchsize hyperparameter
tuning. In this study, based on Table 5, the RetinaNet152 model using batchsize and optimizer
hyperparameter tuning methods proposed in this study show a proportional improvement in system
performance results for mAP and AR. Based on the system performance comparison in Table 5, the
Detecting road damage utilizing retinanet and mobilenet models on edge devices (Haniah Mahmudah)
1438  ISSN: 2252-8938

hyperparameter tuning strategy in the MobileNetV2 model improves performance significantly, resulting in
good system performance and rapid inference time. This demonstrates that the MobileNetV2 model could
potentially be implemented directly on edge devices for road damage detection. The RetinaNet152 model,
which employs the batchsize hyperparameter tuning method and the optimizer developed in this paper,
achieves a proportional improvement in system performance results for mAP and AR. However, the
RetinaNet152 model cannot be implemented directly on edge devices; it requires post-training procedures
like quantization and prunning, among others.
However, according to the results of this study, RetinaNet152 models need to improve system
performance in several respects when compared to the MobileNetV2 model [16]. The first method is to
improve system performance by expanding the dataset by boosting image capture resolution with a
high-resolution camera. There is no comparable dataset for road damage research [16]; hence, it presents
unique issues. Furthermore, adding many augmenting scenarios will increase system performance on the
RetinaNet152 and MobileNetV2 models. In addition, to improve system performance, the RetinaNet152 and
MobileNetV2 models should be optimized for hyperparameter selection utilizing grid search, random search,
bayesian optimization, and so on. Further research into the deployment of CNN models on edge devices should
focus on the structure and size of the models. This is a crucial issue to consider because CNN models with large
model structures and sizes demand more processing resources during the inference phase to detect objects.
To reduce inference time, a CNN model with a large model structure and size should be employed in edge
device research, along with methods such as quantization, prunning, and knowledge distillation [39], [40].

4. CONCLUSION
There is a need for a device as an edge device that recognizes road damage using technology and
AI, as a sort of road digitalization creates a system to identify road damage automatically, efficiently, and in
real time. The use of CNN models for road damage detection on edge devices requires a trade-off between
high accuracy and high efficiency. Currently, edge devices have limited storage. Results comparing the
effects of increasing batch size and different optimizers on RetinaNet152 and MobileNetV2 models: i) the
Adam optimizer performed better, with a mAP, AR, and F1-score for batch size 8. This means that tweaking
the optimizer improves overall model performance; ii) the MobileNetV2 model's hyperparameter tuning
technique considerably enhances performance, resulting in fast inference time and good system performance
overall. This shows that the MobileNetV2 model might be deployed directly on edge devices to detect road
damage; and iii) the RetinaNet152 model, which uses the batchsize hyperparameter tuning approach and the
optimizer described in this research, achieves a proportionate gain in system performance for mAP and AR.
However, the RetinaNet152 model cannot be deployed directly to edge devices using the Jetson Nano 4 GB.
The RetinaNet152 model can be deployed on edge devices; however, a technique for speeding up inference
time is essential, as are post-training procedures such as quantization and prunning. Several
recommendations are made for further research or implementation. First, consider augmenting and improving
the dataset's quality to attain better CNN model training outcomes. Second, system implementation testing
should be carried out on a high-end computer server, such as a NVIDIA RTX 3080 with 32 GB of RAM. In
addition, consider using the RetinaNet50 model, which has a lighter design and model size than the
RetinaNet152. Finally, investigate speeding up the inference time exploring methods like post-training
quantization and converting Tensorflow Lite models to TensorRT to improve implementation efficiency.

ACKNOWLEDGEMENTS
The authors would like to thank Politeknik Elektronika Negeri Surabaya and Institut Teknologi
Sepuluh Nopember for financial support, laboratory equipment and facilities for this work.

REFERENCES
[1] I. Widyatmoko, “Digital transformation to improve quality, efficiency and safety in construction of roads incorporating recycled
materials,” IOP Conference Series: Earth and Environmental Science, vol. 599, no. 1, Nov. 2020, doi: 10.1088/1755-
1315/599/1/012093.
[2] V. Hegde, D. Trivedi, A. Alfarrarjeh, A. Deepak, S. Ho Kim, and C. Shahabi, “Yet another deep learning approach for road
damage detection using ensemble learning,” in 2020 IEEE International Conference on Big Data (Big Data), IEEE, Dec. 2020,
pp. 5553–5558, doi: 10.1109/BigData50022.2020.9377833.
[3] A. Pernestål, A. Engholm, M. Bemler, and G. Gidofalvi, “How will digitalization change road freight transport? scenarios tested
in sweden,” Sustainability, vol. 13, no. 1, Dec. 2020, doi: 10.3390/su13010304.
[4] G. M. Jog, C. Koch, M. Golparvar-Fard, and I. Brilakis, “Pothole properties measurement through visual 2d recognition and 3d
reconstruction,” in Computing in Civil Engineering (2012), Reston, VA: American Society of Civil Engineers, Jun. 2012, pp.
553–560, doi: 10.1061/9780784412343.0070.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1430-1440

Int J Artif Intell ISSN: 2252-8938  1439

[5] E. Buza, S. Omanovic, and A. Huseinovi, “Pothole detection with image processing and spectral clustering,” in 2nd International
Conference on Information Technology and Computer Networks (ITCN ’13), 2013, pp. 48–53.
[6] L. Huidrom, L. K. Das, and S. K. Sud, “Method for automated assessment of potholes, cracks and patches from road surface
video clips,” Procedia - Social and Behavioral Sciences, vol. 104, pp. 312–321, Dec. 2013, doi: 10.1016/j.sbspro.2013.11.124.
[7] M. B. S. G. Naik and V. Nirmalrani, “Detecting potholes using image processing techniques and real-world footage,” in Cognitive
Informatics and Soft Computing-Advances in Intelligent Systems and Computing, Springer, Singapore, 2021, pp. 893–902, doi:
10.1007/978-981-16-1056-1_72.
[8] R. Sharma, K. Singh, and L. Chand, “Analysis of image processing techniques for road anomalies detection,” International
Journal of Emerging Research in Management &Technology, vol. 5, no. 1, 2016, doi: 10.13140/RG.2.2.15513.54886.
[9] M. Gao, X. Wang, S. Zhu, and P. Guan, “Detection and segmentation of cement concrete pavement pothole based on image
processing technology,” Mathematical Problems in Engineering, vol. 2020, pp. 1–13, Jan. 2020, doi: 10.1155/2020/1360832.
[10] M. H. Yousaf, K. Azhar, F. Murtaza, and F. Hussain, “Visual analysis of asphalt pavement for detection and localization of
potholes,” Advanced Engineering Informatics, vol. 38, pp. 527–537, Oct. 2018, doi: 10.1016/j.aei.2018.09.002.
[11] N.-D. Hoang, “An artificial intelligence method for asphalt pavement pothole detection using least squares support vector
machine and neural network with steerable filter‐based feature extraction,” Advances in Civil Engineering, vol. 2018, no. 1, Jan.
2018, doi: 10.1155/2018/7419058.
[12] H. Song, K. Baek, and Y. Byun, “Pothole detection using machine learning,” in Advanced Science and Technology Letters,
SERSC, Feb. 2018, pp. 151–155, doi: 10.14257/astl.2018.150.35.
[13] L. Liu et al., “Deep learning for generic object detection: a survey,” International Journal of Computer Vision, vol. 128, no. 2,
pp. 261–318, Feb. 2020, doi: 10.1007/s11263-019-01247-4.
[14] N.-D. Hoang, T.-C. Huynh, and V.-D. Tran, “Computer vision‐based patched and unpatched pothole classification using machine learning
approach optimized by forensic‐based investigation metaheuristic,” Complexity, vol. 2021, no. 1, Jan. 2021, doi: 10.1155/2021/3511375.
[15] C. Chun and S.-K. Ryu, “Road surface damage detection using fully convolutional neural networks and semi-supervised
learning,” Sensors, vol. 19, no. 24, Dec. 2019, doi: 10.3390/s19245501.
[16] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata, “Road damage detection and classification using deep neural
networks with smartphone images,” Computer-Aided Civil and Infrastructure Engineering, vol. 33, no. 12, pp. 1127–1141, Dec.
2018, doi: 10.1111/mice.12387.
[17] S. Zhou and W. Song, “Deep learning-based roadway crack classification using laser-scanned range images: a comparative study
on hyperparameter selection,” Automation in Construction, vol. 114, Jun. 2020, doi: 10.1016/j.autcon.2020.103171.
[18] S. Faghih-Roohi, S. Hajizadeh, A. Nunez, R. Babuska, and B. De Schutter, “Deep convolutional neural networks for detection of
rail surface defects,” in 2016 International Joint Conference on Neural Networks (IJCNN), IEEE, Jul. 2016, pp. 2584–2589, doi:
10.1109/IJCNN.2016.7727522.
[19] L. Pauly, H. Peel, S. Luo, D. Hogg, and R. Fuentes, “Deeper networks for pavement crack detection,” in 34th International
Symposium on Automation and Robotics in Construction, International Association for Automation and Robotics in Construction,
Jul. 2017, pp. 479–485, doi: 10.22260/ISARC2017/0066.
[20] A. Zhang et al., “Deep learning–based fully automated pavement crack detection on 3d asphalt surfaces with an improved
cracknet,” Journal of Computing in Civil Engineering, vol. 32, no. 5, Sep. 2018, doi: 10.1061/(ASCE)CP.1943-5487.0000775.
[21] Y. Cha, W. Choi, and O. Büyüköztürk, “Deep learning‐based crack damage detection using convolutional neural networks,”
Computer-Aided Civil and Infrastructure Engineering, vol. 32, no. 5, pp. 361–378, May 2017, doi: 10.1111/mice.12263.
[22] D. Wang, Z. Liu, X. Gu, W. Wu, Y. Chen, and L. Wang, “Automatic detection of pothole distress in asphalt pavement using
improved convolutional neural networks,” Remote Sensing, vol. 14, no. 16, Aug. 2022, doi: 10.3390/rs14163892.
[23] M. Faramarzi, “Road damage detection and classification using deep neural networks (yolov4) with smartphone images,” SSRN
Electronic Journal, 2020, doi: 10.2139/ssrn.3627382.
[24] M. A. Benallal and M. S. Tayeb, “An image-based convolutional neural network system for road defects detection,” IAES
International Journal of Artificial Intelligence (IJ-AI), vol. 12, no. 2, Jun. 2023, doi: 10.11591/ijai.v12.i2.pp577-584.
[25] R. Ishimwe, P. Iradukunda, and J. B. Kwizera, “Real-time road damage detection using deep convolutional neural networks and a
smartphone: project report,” Carnegie Mellon University, pp. 1-5, 2021, doi: 10.13140/RG.2.2.26696.44801.
[26] J. Dharneeshkar, V. S. Dhakshana, S. A. Aniruthan, R. Karthika, and L. Parameswaran, “Deep learning based detection of
potholes in indian roads using yolo,” in 2020 International Conference on Inventive Computation Technologies (ICICT), IEEE,
Feb. 2020, pp. 381–385, doi: 10.1109/ICICT48043.2020.9112424.
[27] K. R. Ahmed, “Smart pothole detection using deep learning based on dilated convolution,” Sensors, vol. 21, no. 24, Dec. 2021,
doi: 10.3390/s21248406.
[28] E. M. Dogo, O. J. Afolabi, and B. Twala, “On the relative impact of optimizers on convolutional neural networks with varying
depth and width for image classification,” Applied Sciences, vol. 12, no. 23, Nov. 2022, doi: 10.3390/app122311976.
[29] H. Samma, S. A. Suandi, N. A. Ismail, S. Sulaiman, and L. L. Ping, “Evolving pre-trained cnn using two-layers optimizer for road
damage detection from drone images,” IEEE Access, vol. 9, pp. 158215–158226, 2021, doi: 10.1109/ACCESS.2021.3131231.
[30] Y. Wang, Z. Xiao, and G. Cao, “A convolutional neural network method based on adam optimizer with power-exponential
learning rate for bearing fault diagnosis,” Journal of Vibroengineering, vol. 24, no. 4, pp. 666–678, Jun. 2022, doi:
10.21595/jve.2022.22271.
[31] Vidushi, M. Agarwal, A. Rajak, and A. K. Shrivastava, “Assessment of optimizers impact on image recognition with
convolutional neural network to adversarial datasets,” Journal of Physics: Conference Series, vol. 1998, no. 1, Aug. 2021, doi:
10.1088/1742-6596/1998/1/012008.
[32] I. D. Pratama, H. Mahmudah, and R. W. Sudibyo, “Design and implementation of real-time pothole detection using convolutional
neural network for iot smart environment,” in 2021 International Electronics Symposium (IES), IEEE, Sep. 2021, pp. 675–679,
doi: 10.1109/IES53407.2021.9594038.
[33] M. Al Qurishee, “Low-cost deep learning uav and raspberry pi solution to real time pavement condition assessment,” M.Sc.
Thesis, Department of Science and Engineering, University of Tennessee, Chattanooga, US, 2019.
[34] Z. S. Hernanda, H. Mahmudah, and R. W. Sudibyo, “CNN-based hyperparameter optimization approach for road pothole and
crack detection systems,” in 2022 IEEE World AI IoT Congress (AIIoT), IEEE, Jun. 2022, pp. 538–543, doi:
10.1109/AIIoT54504.2022.9817316.
[35] A. C. Aqsa, H. Mahmudah, and R. W. Sudibyo, “Detection and classification of road damage using cnn with hyperparameter
optimization,” in 2022 6th International Conference on Informatics and Computational Sciences (ICICoS), IEEE, Sep. 2022, pp.
101–104, doi: 10.1109/ICICoS56336.2022.9930607.
[36] L. Ale, N. Zhang, and L. Li, “Road damage detection using retinanet,” in 2018 IEEE International Conference on Big Data (Big
Data), IEEE, Dec. 2018, pp. 5197–5200, doi: 10.1109/BigData.2018.8622025.
Detecting road damage utilizing retinanet and mobilenet models on edge devices (Haniah Mahmudah)
1440  ISSN: 2252-8938

[37] Z. Xue et al., “Large-batch optimization for dense visual predictions,” arXiv-Computer Science, pp. 1-23, Oct. 2022.
[38] H. Mahmudah, S. Arifin, A. S. Aisjah, and C. A. Prastyanto, “Optimizers impact on retinanet model for detecting road damage on
edge device,” in 2024 IEEE International Conference on Artificial Intelligence and Mechatronics Systems (AIMS), IEEE, Feb.
2024, pp. 1–6, doi: 10.1109/AIMS61812.2024.10512864.
[39] H. Mahmudah, S. Arifin, and A. S. Aisyah, “A survey of the optimization system road damage detection on cnn model for edge
device,” in 2023 International Conference on Information Technology and Computing (ICITCOM), IEEE, Dec. 2023, pp. 323–
328, doi: 10.1109/ICITCOM60176.2023.10442409.
[40] M. M. H. Shuvo, S. K. Islam, J. Cheng, and B. I. Morshed, “Efficient acceleration of deep learning inference on resource-
constrained edge devices: a review,” Proceedings of the IEEE, vol. 111, no. 1, pp. 42–91, Jan. 2023, doi:
10.1109/JPROC.2022.3226481.

BIOGRAPHIES OF AUTHORS

Haniah Mahmudah completed bachelor’s degree in Department of Engineering

Physics, Institut Teknologi Sepuluh Nopember and master's degree in Department of
Electrical Engineering, Institut Teknologi Sepuluh Nopember. She is currently working as a
lecturer in Department of Electrical Engineering, Politeknik Elektronika Negeri Surabaya. She
is also active in conducting research especially in signal processing, wire and wireless
communication, internet of things, applied of machine learning. She can be contacted at email:
[email protected].

Aulia Siti Aisjah received doctoral degree Institut Teknologi Sepuluh Nopember.
She is currently working as a lecturer in Department of Engineering Physics, Institut
Teknologi Sepuluh Nopember. Her research interests are in apllied of machine learning in
industrial, marine engineering and instrumentation. She can be contacted at email:
[email protected].

Syamsul Arifin is currently working as a lecturer in Department of Engineering

Physics, Institut Teknologi Sepuluh Nopember. His research interests are in instrumentation
and apllied of machine learning. He can be contacted at email: [email protected].

Catur Arif Prastyanto completed bachelor’s degree, master's degree and

doctor’s degree in Department of Civil Engineering, Institut Teknologi Sepuluh Nopember.
He is currently working as a lecturer in Department of Civil Engineering, Institut Teknologi
Sepuluh Nopember. Her research interests are highways and transportation. He can be
contacted at email: [email protected].

Int J Artif Intell, Vol. 14, No. 2, April 2025: 1430-1440