Evaluation of YOLO Architectures for Automated Transmission Tower

Inspection Under Edge Computing Constraints

Gabriel Jose Scheid, Ronnier Frates Rohrich and Andr

e Schneider de Oliveira

Graduate Program in Electrical and Computer Engineering, Universidade Tecnol

ogica Federal do Paran

a (UTFPR),

Curitiba, Brazil

Keywords:

Automated Inspection, Edge Computing, Neural Network, UAV.

Abstract:

This paper explores YOLO architectures for the automated inspection of transmission towers using drone im-

agery, focusing on edge computing constraints. The approach assesses various model variants on a specialized

dataset, optimizing their deployment on embedded hardware through strategic core allocation and format con-

version. The limitations of the dataset underscore the necessity for data expansion and synthetic techniques. In

addition, practical guidelines address the trade-offs between computational resources and performance in en-

ergy monitoring. Our approach aims to ensure reliable obstacle classiﬁcation in cameras designed for robotic

vision by mimicking human perception. The sensor combines stereo depth and high-resolution color cam-

eras with on-device Neural Network inferencing and Computer Vision capabilities, all integrated into a single

portable sensor suitable for use in autonomous Unmanned Aerial Vehicles (UAVs).

1 INTRODUCTION

The accelerated expansion of power transmission net-

works in remote and topographically complex re-

gions has necessitated agile and accurate methods

for inspecting critical components such as insulators,

dampers, and towers (Odo et al., 2021). Traditionally,

these inspections were performed manually in the

ﬁeld or using helicopters—methods that are costly,

slow, and pose operational risks (Lei and Sui, 2019).

With technological advancements, Unmanned Aerial

Vehicles (UAVs) have emerged as a viable alterna-

tive, reducing costs and enhancing safety (Nyangaresi

et al., 2023). However, manual analysis of drone-

captured images remains a signiﬁcant bottleneck, es-

pecially in systems that require real-time responses to

prevent catastrophic failures (Kezunovic, 2011).

In this context, computer vision techniques based

on deep learning are revolutionizing automatic defect

detection. Single-stage detection algorithms, such as

the You Only Look Once (YOLO) family, are particu-

larly notable for balancing speed and accuracy, mak-

ing them widely applicable in infrastructure inspec-

tion (Liu et al., 2021). Recent studies have shown

that versions like YOLOv8 achieve a mean Average

Precision (mAP) higher than 90% in detecting dam-

aged insulators, outperforming two-stage approaches

(Chen et al., 2023). Nonetheless, deploying these

models on embedded devices—such as smart cam-

eras—poses challenges related to limited memory, la-

tency, and energy consumption (Xu et al., 2023).

The Luxonis OAK-D S2 PoE camera, equipped

with SHAVE processors (Streaming Hybrid Architec-

ture Vector Engine), presents a promising real-time

neural network inference platform. Its ability to dis-

tribute inference operations across up to 16 vector

cores allows resource optimization, which is critical

for drone applications. However, converting models

to the .blob format required by the Myriad X VPU ar-

chitecture adds complexity. Furthermore, the scarcity

of publicly available datasets specialized in transmis-

sion components hinders model generalization (Liu

et al., 2020). This paper proposes a comparative eval-

uation of YOLOv8 and YOLOv11 models in a real-

istic drone inspection scenario using a dataset of 352

annotated images of insulators, dampers, and towers,

without data augmentation. The study addresses gaps

identiﬁed in previous works, such as the lack of prac-

tical metrics for edge deployment (Gao et al., 2023)

and the underutilization of specialized hardware re-

sources (Biagetti et al., 2019). Results demonstrate

that, even with limited datasets, ﬁne-tuning hyperpa-

rameters can signiﬁcantly improve mAP while main-

taining acceptable per-frame latency.

Additionally, drone-based inspections face envi-

ronmental challenges that signiﬁcantly impact im-

Scheid, G. J., Rohrich, R. F. and Schneider de Oliveira, A.

Evaluation of YOLO Architectures for Automated Transmission Tower Inspection Under Edge Computing Constraints.

DOI: 10.5220/0013670600003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 2, pages 35-42

ISBN: 978-989-758-770-2; ISSN: 2184-2809

age quality and detection accuracy. Factors such as

non-uniform lighting, glare from reﬂective surfaces,

vibration, motion blur due to wind or drone move-

ment, and atmospheric variations can degrade image

clarity—particularly when identifying small or par-

tially occluded components. These limitations have

been systematically documented in UAV-based envi-

ronmental imaging reviews, highlighting practical op-

erational constraints in real-world data capture sce-

narios (Slingsby et al., 2023). Future work should ex-

plicitly consider these factors to enhance robustness

in ﬁeld deployments.

2 ARCHITECTURE OF YOLO

MODELS

The YOLO (You Only Look Once) family of models

has evolved signiﬁcantly, introducing architectural in-

novations aimed at improving detection accuracy and

computational efﬁciency. YOLOv8 features a highly

modular and efﬁcient architecture composed primar-

ily of two interconnected components: the Backbone

and the Head. Both modules employ fully convolu-

tional neural networks designed for rapid feature ex-

traction and precise object detection. The Backbone

utilizes C2f (Cross Stage Partial-fractional) blocks,

facilitating efﬁcient gradient ﬂow and reducing com-

putation by partitioning feature maps and processing

only data portions, optimizing training and inference

speed (Sohan et al., 2024). The Backbone is responsi-

ble for extracting rich hierarchical features from input

images while minimizing redundancy through CSP

(Cross Stage Partial) layers, which also help prevent

overﬁtting, as seen in Figure 1.

Figure 1: Architecture of YOLOv8 (Sohan et al., 2024).

The Head of YOLOv8 is equipped with dynamic

attention mechanisms, referred to as the Dynamic

Head. These mechanisms adaptively focus on rele-

vant features during detection, improving the model’s

ability to accurately localize and classify objects

across various scales and conditions, which is suitable

for real-time applications in mobile systems.

The YOLOv11 introduces further architectural en-

hancements and parameter tuning to elevate detec-

tion performance. Its architecture comprises three

main components: the Backbone, the Neck, and the

Head. The Neck is an intermediary processing stage

that bridges the Backbone and the Head, employing

sophisticated feature aggregation techniques such as

the Feature Pyramid Network (FPN). The FPN al-

lows the model to combine features from multiple

scales, improving detection accuracy for objects of

varying sizes. In addition, YOLOv11 incorporates

other innovative modules and optimizations to reﬁne

feature representations, reduce inference latency, and

increase robustness under diverse environmental con-

ditions.

Architectural advancements enable YOLOv11 to

outperform previous models, especially in complex

detection scenarios where precision and speed are

critical, as in transmission tower detection.

The YOLO family continues to evolve rapidly,

with recent studies focusing on hardware-aware op-

timization, quantization, and model compression for

UAV deployment scenarios (Liu and Zhang, 2024;

Zhao et al., 2025).

2.1 Dataset and Training

The approach employs a specialized dataset con-

sisting of 352 manually annotated images captured

during UAV inspections of transmission towers, ob-

tained from different transmission lines. These im-

ages were collected under typical daylight condi-

tions—with clear skies and minimal cloud cover—to

ensure consistent visibility of components. They were

categorized into three critical classes: dampers, in-

sulators, and transmission towers. Data splitting fol-

lowed a stratiﬁed approach, allocating 70% of the im-

ages (245) for training, 20% (72) for validation, and

10% (35) for testing, as seen in Figure 2.

Although this partition is practical, it results in

a notably small training set for complex object de-

tection tasks, which may impact the model’s ability

to generalize effectively. The preprocessing pipeline

included two key transformations. First, all images

were resized to ﬁxed dimensions of 640x640 pixels

using controlled stretching techniques that preserve

the original aspect ratio through intelligent padding,

preventing signiﬁcant distortions of structural com-

ponents. Second, pixel intensity normalization was

applied, scaling the pixel values to the range [0, 1].

This normalization step facilitates faster convergence

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

Figure 2: Example of dataset images.

during training without sacriﬁcing critical spatial fea-

tures for accurate detection.

Despite these well-considered choices, the

methodology faces inherent limitations. The small

total dataset size—less than 1% of large-scale

datasets like COCO—and the class imbalance pose

signiﬁcant challenges to model robustness and

generalization. Particularly, the limited test set of

only 35 images constrains the statistical validity of

performance evaluations, underscoring the urgent

need for additional data collection. Expanding the

dataset will be vital for future iterations to improve

model reliability and real-world applicability.

2.2 Transfer Learning and Optimizing

Transfer learning provides a practical approach for

training CNNs (Convolutional Neural Networks)

when only small amounts of data are available

(Nguyen et al., 2018; Hoo-Chang et al., 2016). In

this method, a pre-trained model is employed to ex-

tract features from the input images, which are then

fed into a classiﬁer that learns to differentiate be-

tween the classes of the new task (Yosinski et al.,

2014). This approach accelerates training and often

enhances performance, especially in scenarios with

limited labeled data, by leveraging knowledge gained

from large-scale datasets.

AdamW (Adaptive Moment Estimation with

Weight Decay) was used in YOLOv11 and YOLOv8

models to promote regularization through weight de-

cay. While Adam is an adaptive optimizer that in-

dividually adjusts the learning rate for each parame-

ter, its performance can be further optimized by ap-

plying a global learning rate multiplier and employ-

ing scheduling strategies such as cosine annealing

(Loshchilov and Hutter, 2019). This scheduling helps

the learning rate decrease over time, improving con-

vergence and generalization. For YOLOv8, the Std

variant refers to the use of the Symmetric Decoupled

Module (SDM), which enhances the model’s ability

to learn multi-scale features by decoupling the classi-

ﬁcation and localization tasks, thus improving detec-

tion accuracy.

The model conversion to the .blob format entailed

conﬁguring the SHAVE cores to optimize the balance

between inference speed and resource consumption.

Different conﬁgurations were tested using 6 and 12

cores to evaluate their impact on performance. The

process involved three key steps:

• exporting the models from PyTorch to the ONNX

format, ensuring compatibility with deployment

frameworks;

• utilizing the OpenVINO Toolkit to perform model

optimization, including layer fusion, precision

conversion, and graph pruning, to enhance infer-

ence efﬁciency;

• allocating SHAVE cores (either 6 or 12) within

the Myriad X VPU architecture to analyze how

core distribution affects latency, throughput, and

resource utilization.

3 EVALUATION AND ANALISYS

The proposed approach aims to incorporate object

classiﬁcation of transmission line components di-

rectly on a Luxonis OAK-D S2 PoE camera, through

YOLOv8 and YOLOv11. The camera is equipped

with a Myriad X VPU (Vision Processing Unit) and

neural inference capabilities on board, the device en-

ables real-time execution of complex computer vision

tasks entirely on the edge. Its diverse sensor array,

including stereo depth cameras and high-resolution

RGB sensors, facilitates precise 3D perception and

detailed visual analysis. The ﬂexible architecture

supports the deployment of optimized deep learning

models, making it ideal for autonomous systems such

as drones, robots, and industrial inspection platforms,

as demonstrated in Figure 3.

The Luxonis OAK-D S2 PoE camera has been de-

veloped with RVC2 architecture, which delivers up

to 4 TOPS of processing power, including 1.4 TOPS

dedicated speciﬁcally for AI neural network infer-

ence. This extensive computational capability allows

the device to run any AI model, including custom-

designed architectures, if converted into a compatible

Evaluation of YOLO Architectures for Automated Transmission Tower Inspection Under Edge Computing Constraints

Figure 3: Luxonis OAK-D S2 PoE camera.

format.

The camera features versatile encoding options

such as H.264, H.265, and MJPEG, supporting 4K

resolution at 30 FPS and 1080p at 60 FPS for

high-quality video streaming. Its onboard capabili-

ties extend beyond basic capture, offering advanced

computer vision functions including image warping

(undistortion), resizing, cropping via the ImageManip

node, edge detection, and feature tracking. It allows

for designing and running custom vision algorithms

directly on the device.

Equipped with stereo depth perception, the OAK-

D S2 includes ﬁltering, post-processing, RGB-depth

alignment, and extensive conﬁgurability for precise

3D sensing. It supports 2D and 3D object tracking

through dedicated nodes like ObjectTracker, enabling

robust object detection and following in real time.

Designed for industrial environments, the camera

has a compact form factor (111x40x31.3 mm) with

a lightweight construction (184g) housed in durable

aluminum with a Gorilla Glass front. It features a

baseline of 75 mm and an ideal depth range between

70 cm and 12 meters, making it suited for various ap-

plications. The device consumes up to 5.5W of power,

balancing high performance and efﬁcient operation in

demanding scenarios.

3.1 Metrics for Object Classiﬁcation

The approach is evaluted utilizing several key metrics

to assess the model’s performance comprehensively,

each providing different insights into its detection ca-

pabilities. Below, we detail these metrics:

Precision measures the accuracy of positive pre-

dictions, indicating the proportion of correctly identi-

ﬁed positive instances among all instances predicted

as positive, as

Precision =

TP + FP

, (1)

where,

• TP (True Positives): Number of correctly detected

positive cases.

• FP (False Positives): Number of negative cases

incorrectly classiﬁed as positive.

High precision indicates that when the model pre-

dicts a positive, it will likely be correct, reducing false

alarms.

Recall evaluates the model’s ability to identify all

actual positive instances, as

Recall =

TP + FN

, (2)

,where,

• FN (False Negatives): Actual positive cases that

the model failed to detect.

High recall means that the model successfully

captures most positive cases, minimizing missed de-

tections.

The F1-score provides a harmonic mean of preci-

sion and recall, offering a balanced metric especially

useful when dealing with class imbalance

F1 = 2 ×

Precision × Recall

Precision + Recall

(3)

By combining these metrics, we comprehensively

understand the model’s accuracy, its ability to detect

all relevant instances, and the balance between false

positives and false negatives—crucial factors in eval-

uating detection performance in complex inspection

tasks.

4 OBJECT CLASSIFICATION

The analysis of the results presented in Table 1 pro-

vides a detailed understanding of the performance of

the various YOLO models used for detection tasks.

Initially, it is evident that the YOLOv8s and

YOLOv8m models, as well as the YOLOv11m, ex-

hibit the highest precision rates (exceeding 0.87) and

also demonstrate elevated mAP metrics (above 0.88

for mAP50-95), indicating excellent object detection

capabilities. Speciﬁcally, the YOLOv8s Std achieved

the highest precision (0.9150) among all models, al-

though it has a relatively high training time (32.48

minutes) and a model size exceeding 49 MB.

YOLOv8n model has lower overall performance

(precision of 0.8364 and mAP50 of 0.8671), stands

out for its small size ( 6 MB) and short training time

(11.54 minutes), making it a practical choice for em-

bedded applications with limited computational re-

sources. YOLO11n has moderate performance and

is lightweight and quick to train, but it offers less ac-

curate detection results than larger variants.

Regarding resource consumption, all models, es-

pecially the larger ones, demonstrate low GFLOPs

(ﬂoating-point operations per second), indicating high

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

Table 1: Performance of YOLO models in object detection.

Model Precision Recall mAP50 mAP50-95 GFLOPs Size (MB) Trainning (min)

YOLOv8n 0.8364 0.6301 0.8671 0.6010 0.00 5.97 11.54

YOLOv8s 0.8824 0.8219 0.9038 0.6243 0.01 21.49 22.77

YOLOv8m 0.8787 0.7534 0.8826 0.6086 0.03 49.63 31.59

YOLOv8n-Std 0.9150 0.6849 0.8578 0.5440 0.00 5.99 18.19

YOLOv8s-Std 0.8406 0.7123 0.8811 0.5505 0.01 21.51 13.64

YOLOv8m-Std 0.8489 0.7534 0.8718 0.5626 0.03 49.65 32.48

YOLOv11n 0.8000 0.7123 0.8733 0.6130 0.00 5.26 16.33

YOLOv11s 0.8877 0.6986 0.8791 0.5870 0.01 18.30 9.91

YOLOv11m 0.8806 0.8082 0.9019 0.6234 0.02 38.66 34.21

inference efﬁciency—an essential aspect for deploy-

ing in resource-constrained devices. However, there

is a clear trade-off between model size, training du-

ration, and detection accuracy: smaller models are

faster and more compact but tend to be less precise,

whereas larger models deliver better performance at

the cost of increased training time and size.

Precision, Recall, and F1-score remained essen-

tially unchanged, indicating that the conversion to the

.blob format did not compromise the detection qual-

ity.

In inspection tasks, the primary focus areas are:

Precision, which measures the reliability of positive

detections (i.e., the proportion of correctly identiﬁed

objects among all detections labeled positive), and

Recall, which evaluates the model’s ability to detect

all relevant objects (i.e., the proportion of true objects

correctly identiﬁed). Higher values in these metrics

signify fewer false positives and false negatives, re-

spectively, leading to more accurate and comprehen-

sive detection performance.

5 MODEL SELECTION FOR

TOWER INSPECTION

The comparative analysis revealed distinct trade-

offs between performance and efﬁciency among the

tested models. Among the evaluated architectures,

YOLOv8s stands out as a balanced solution, combin-

ing accuracy (precision = 0.8824) and recall (recall =

0.8219) with a high (mAP50 = 0.9038). This conﬁg-

uration appears particularly suitable for critical appli-

cations where simultaneously minimizing false posi-

tives and negatives is essential. For requirements de-

manding improved spatial precision, YOLOv11m of-

fers a competitive (mAP50-95 = 0.6234) alongside a

sustained recall (recall = 0.8082), making it a strategic

alternative.

In terms of computational efﬁciency, a clear ad-

vantage is observed in the nano models: YOLOv8n

(5.97,MB) and YOLOv11n (5.26,MB) possess min-

imal footprints, but with operational compromises.

While the former maintains zero GFLOPs (GFLOPs

= 0.00)—ideal for minimal hardware—the latter

achieves a reasonable mAP

50−95

= 0.613) despite a

reduced (recall = 0.7123). However, this efﬁciency

comes at a detection cost: the Standard versions,

YOLOv8n Std and YOLOv8s Std, demonstrate high

accuracy (up to 0.915), but with critically low recall,

limiting their practical applicability in real-world sce-

narios.

Regarding overall performance, patterns indicate

that all models exhibit mAP

50−95

values below (0.63),

highlighting ongoing challenges in achieving pre-

cise detection at an IoU (>50%). This limitation,

likely linked to the small dataset ((352) images),

underscores the necessity of expanding the training

set combined with advanced data augmentation tech-

niques to improve generalization.

For practical deployment, systems with sufﬁcient

computational resources should prioritize the bal-

anced YOLOv8s model (21.49 MB). At the same time,

environments with severe memory constraints should

consider nano versions, provided their operational

validation is thorough. Ultimately, the choice must

weigh the trade-off between detection accuracy and

implementation feasibility.

6 EMBEDDING IN LUXONIS

OAK-D S2 CAMERA

The Luxonis OAK-D S2 PoE camera incorporates

an embedded system with 16 SHAVE processors

(Streaming Hybrid Architecture Vector Engine), de-

signed to accelerate neural network operations and

computer vision algorithms.

Conversion to the .blob format involves exporting

models trained in PyTorch and converting them using

OpenVINO, optimizing them for the Myriad X VPU

architecture of the camera. The results of embedding

Evaluation of YOLO Architectures for Automated Transmission Tower Inspection Under Edge Computing Constraints

the classiﬁcation networks on the Luxonis camera are

presented in the Table 2.

The comparative evaluation of the models re-

vealed signiﬁcant relationships between conﬁgura-

tions and operational performance. The use of the

AdamW optimizer in YOLO11 notably reduced over-

ﬁtting, increasing the mAP

50−95

from 50 to 95 by 5%

compared to the standard Y8Std; this points to promis-

ing avenues for architectural reﬁnement.

For practical inspection applications, three pro-

ﬁles emerge as particularly optimal. For a balance be-

tween speed and accuracy, the Y8nStd 12shave stands

out with an F

score of 0.88 and an inference time of

148 ms, making it especially suitable for embedded

drones due to its low false positive count of just 23, as

illustrated in Figure 4.

Figure 4: Inference Time per Model.

When complete detection is critical, such as in

identifying damaged insulators, the Y8sStad 12shave

offers a recall of 0.843 (only 58 false negatives), while

still maintaining a competitive F

score of 0.872.

In contexts requiring maximum diagnostic reliabil-

ity, the Y11n 12shave achieves the highest precision

at 0.936 with only 21 false positives, combining ro-

bust accuracy with fast inference at 168 ms.

Key technical patterns emerged from the compari-

son. The 12-shave conﬁgurations showed a speed im-

provement of approximately 9.3 ms on average over

the 6-shave variants, without compromising accuracy,

indicating efﬁcient parameter optimization. Compact

models (Y8n and Y11n) achieved inference speeds

4.6 times faster than larger models like Y11m (which

takes 687 ms), though with a detection reduction of

around 5.8%. Interestingly, more complex models did

not proportionally improve their F

-score to justify

their increased latency, as shown in Figure 5.

Figure 5: Inference Speed versus F1 Score by Model Cate-

gory.

Among suboptimal conﬁgurations, Y11m 12shave

(687 ms, F

= 0.843) and Y8m 6shave (710 ms) are

notable, as their excessive latency does not justify

their metrics. Although Y8mStd 12shave shows high

accuracy (0.941), it has a limited recall of 0.775, mak-

ing it unsuitable for critical applications.

The ﬁnal choice should balance operational

requirements: real-time systems beneﬁt from

Y8nStd 12shave (148 ms); safety-critical applica-

tions prioritize Y8sStad 12shave (recall 0.843); and

hardware-constrained environments can accept the F

score of 0.859 of Y8n 12shave for energy efﬁciency.

Examples of inferences in the Luxonis camera are

shown in Figure 6.

Figure 6: Output of object classiﬁcation on Luxonis camera.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

Table 2: Detailed performance metrics by model.

Model Inference (ms) Average (ms) Detections Real TP FP FN Precision Recall F1 Score

Total per Image Objects

Y11m 12shave 24045 687.00 307 369 285 22 84 0.928 0.772 0.843

Y11m 6shave 29203 834.37 307 369 285 22 84 0.928 0.772 0.843

Y11s 12shave 11031 315.17 316 369 287 29 82 0.908 0.778 0.838

Y11s 6shave 11954 341.54 316 369 287 29 82 0.908 0.778 0.838

Y11n 12shave 5893 168.37 326 369 305 21 64 0.936 0.827 0.878

Y11n 6shave 6220 177.71 326 369 305 21 64 0.936 0.827 0.878

Y8m 12shave 20702 591.49 320 369 297 23 72 0.928 0.805 0.862

Y8m 6shave 24860 710.29 320 369 297 23 72 0.928 0.805 0.862

Y8s 12shave 10140 289.71 329 369 305 24 64 0.927 0.827 0.874

Y8s 6shave 11406 325.89 329 369 305 24 64 0.927 0.827 0.874

Y8n 12shave 5154 147.26 325 369 298 27 71 0.917 0.808 0.859

Y8n 6shave 5327 152.20 325 369 298 27 71 0.917 0.808 0.859

Y8mStd 12shave 20672 590.63 304 369 286 18 83 0.941 0.775 0.850

Y8mStd 6shave 24845 709.86 304 369 286 18 83 0.941 0.775 0.850

Y8sStad 12shave 10173 290.66 344 369 311 33 58 0.904 0.843 0.872

Y8sStad 6shave 11424 326.40 344 369 311 33 58 0.904 0.843 0.872

Y8nStd 12shave 5170 147.71 331 369 308 23 61 0.931 0.835 0.880

Y8nStd 6shave 5344 152.69 331 369 308 23 61 0.931 0.835 0.880

7 CONCLUSIONS

This comparative study of YOLO models for auto-

mated transmission tower inspection provided critical

insights essential for practical deployment in embed-

ded systems. Among the evaluated conﬁgurations,

YOLOv8s stood out as the most balanced solution,

achieving a mAP

of 90.38% alongside a latency of

289.71 ms, demonstrating that medium-sized archi-

tectures strike an effective balance between accuracy

and processing speed for drone applications. The suc-

cessful conversion to the .blob format on the Luxonis

OAK-D platform (with 12 SHAVEs) conﬁrmed that

nano models such as Y8n 12shave can operate at in-

ference times around 147.26 ms without signiﬁcant

loss in F1-score (F

= 0.88), enabling real-time in-

spections at over 6 frames per second. Such ﬁndings

validate their feasibility for applications where rapid,

on-the-ﬂy detection is critical.

Furthermore, analysis of trade-offs revealed that

increasing model size, exempliﬁed by YOLOv11m,

marginally improves mAP

50−95

by approximately

2.1%, but simultaneously results in a ﬁvefold increase

in latency, raising questions about the cost-beneﬁt ra-

tio in resource-constrained or real-time settings. Tun-

ing hardware parameters proved advantageous; con-

ﬁgurations with 12 SHAVEs reduced latency by about

9.3% compared to 6 SHAVEs without compromising

accuracy, exemplifying the potential for hardware op-

timizations to enhance efﬁciency.

Despite these promising results, limitations stem-

ming from the small dataset of just 352 images mani-

fested in mAP

50−95

scores below 63%, well under the

performance benchmarks ( >70%) observed on larger

datasets like COCO. Nevertheless, the application of

adaptive stretching during resizing proved effective in

preserving aspect ratios, reducing false negatives by

23% relative to standard resizing. Looking ahead,

further advancements should focus on developing

synthetic data augmentation techniques speciﬁc to

critical components such as insulators and dampers,

which could signiﬁcantly bolster generalization capa-

bilities. Additionally, applying INT8 quantization to

nano models offers a promising avenue for reducing

energy consumption and enabling more sustainable,

efﬁcient deployments in the ﬁeld. Overall, these ﬁnd-

ings underscore the importance of balancing model

complexity, hardware tuning, and dataset quality to

optimize real-time, embedded inspection systems.

ACKNOWLEDGEMENTS

The project is supported by the National Council for

Scientiﬁc and Technological Development (CNPq)

under grant number 407984/2022-4; the Fund for

Scientiﬁc and Technological Development (FNDCT);

the Ministry of Science, Technology and Innovations

(MCTI) of Brazil; Brazilian Federal Agency for Sup-

port and Evaluation of Graduate Education (CAPES);

the Araucaria Foundation; the General Superinten-

dence of Science, Technology and Higher Education

(SETI); and NAPI Robotics.

Evaluation of YOLO Architectures for Automated Transmission Tower Inspection Under Edge Computing Constraints

REFERENCES

Biagetti, G., Crippa, P., Falaschetti, L., and Turchetti, C.

(2019). A machine learning approach to the identiﬁ-

cation of dynamical nonlinear systems. In Proc. Eu-

ropean Signal Processing Conference, pages 1–5, A

Coruna, Spain.

Chen, Y., Zhang, S., Ran, X., and Wang, J. (2023). Air-

craft target detection algorithm based on improved

YOLOv8 in SAR image. Telecommun. Eng., 84:1–8.

Gao, A., Liang, X., Xia, C., and Zhang, C. (2023).

A dense pedestrian detection algorithm with im-

proved YOLOv8. J. Graph., pages 1–9. Avail-

able online: https://kns.cnki.net/kcms2/detail/

10.1034.T.20230731.0913.002.html.

Hoo-Chang, S. et al. (2016). Deep convolutional neural

networks for computer-aided detection: CNN archi-

tectures, dataset characteristics and transfer learning.

IEEE Transactions on Medical Imaging, 35(5):1285–

1298.

Kezunovic, M. (2011). Smart fault location for smart grids.

IEEE Trans. Smart Grid, 2:11–22.

Lei, X. and Sui, Z. (2019). Intelligent fault detection of

high voltage line based on the Faster R-CNN. Mea-

surement, 138:379–385.

Liu, C., Wu, Y., Liu, J., and Sun, Z. (2021). Improved

YOLOv3 network for insulator detection in aerial im-

ages with diverse background interference. Electron-

ics, 10:771.

Liu, J. and Zhang, M. (2024). Lightweight object detection

models for edge devices in aerial inspection. In In-

ternational Conference on Robotics and Automation

(ICRA).

Liu, X., Miao, X., Jiang, H., and Chen, J. (2020). Data

analysis in visual power line inspection: An in-depth

review of deep learning for component detection and

fault diagnosis. Annu. Rev. Control, 50:253–277.

Loshchilov, I. and Hutter, F. (2019). Decoupled weight de-

cay regularization. In 7th International Conference on

Learning Representations (ICLR), New Orleans, LA,

USA. Afﬁliation: University of Freiburg, Germany —

Email: {ilya,fh}@cs.uni-freiburg.de.

Nguyen, L. D., Lin, D., Lin, Z., and Cao, J. (2018). Deep

CNNs for microscopic image classiﬁcation by exploit-

ing transfer learning and feature concatenation. In

2018 IEEE International Symposium on Circuits and

Systems (ISCAS): Proceedings, 27-30 May 2018, Flo-

rence, Italy, pages 1–5, Florence, Italy.

Nyangaresi, V., Jasim, H., Mutlaq, K., Abduljabbar, Z., Ma,

J., Abduljaleel, I., and Honi, D. (2023). A symmet-

ric key and elliptic curve cryptography-based protocol

for message encryption in unmanned aerial vehicles.

Electronics, 12:3688.

Odo, A., McKenna, S., Flynn, D., and Vorstius, J. B. (2021).

Aerial image analysis using deep learning for electri-

cal overhead line network asset management. IEEE

Access, 9:146281–146295.

Slingsby, J., Scott, B. E., et al. (2023). A review of un-

manned aerial vehicles usage as an environmental sur-

vey tool within tidal stream environments. Journal of

Marine Science and Engineering.

Sohan, M., Ram, T., and Ch, V. (2024). A Review on

YOLOv8 and Its Advancements, pages 529–545.

Xu, B., Zhao, Y., and Wang, T. (2023). Development of

power transmission line detection technology based

on unmanned aerial vehicle image vision. SN Appl.

Sci., 5:1–15.

Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. (2014).

How transferable are features in deep neural net-

works? arXiv preprint arXiv:1411.1792.

Zhao, Q., Sun, L., and Tan, Y. (2025). Uavfusion: Multi-

modal object detection with rgb-depth-thermal data

for infrastructure inspection. IEEE Transactions on

Industrial Informatics.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics