Thermal Image Super-Resolution Using Real-ESRGAN for Human
Detection
Vin
´
ıcius H. G. Correa
1 a
, Peter Funk
2 b
, Nils Sundelius
2
, Rickard Sohlberg
2
,
Mastura Ab Wahid
3 c
and Alexandre C. B. Ramos
4 d
1
Institute of Mechanical Engineering, Federal University of Itajub
´
a, Itajub
´
a, Brazil
2
School of Innovation, Design and Engineering, M
¨
alardalen University, V
¨
aster
˚
as, Sweden
3
School of Mechanical Engineering, Universiti Teknologi Malaysia, Johor Bahru, Malaysia
4
Institute of Mathematics and Computing, Federal University of Itajub
´
a, Itajub
´
a, Brazil
Keywords:
Digital Image Processing, Generative Adversarial Networks, Target Detection, Search and Rescue.
Abstract:
Unmanned Aerial Vehicles (UAVs) are increasingly crucial in Search and Rescue (SAR) operations due to
their ability to enhance efficiency and reduce costs. Search and Rescue is a vital activity as it directly impacts
the preservation of life and safety in critical situations, such as locating and rescuing individuals in perilous
or remote environments. However, the effectiveness of these operations heavily depends on the quality of
sensor data for accurate target detection. This study investigates the application of the Real Enhanced Super-
Resolution Generative Adversarial Networks (Real-ESRGAN) algorithm to enhance the resolution and detail
of infrared images captured by UAV sensors. By improving image quality through super-resolution, we then
assess the performance of the YOLOv8 target detection algorithm on these enhanced images. Preliminary
results indicate that Real-ESRGAN significantly improves the quality of low-resolution infrared data, even
when using pre-trained models not specifically tailored to our dataset, this highlights a considerable potential
of applying the algorithm in the preprocessing stages of images generated by UAVs for search and rescue
operations.
1 INTRODUCTION
The use of UAVs (Unmanned Aerial Vehicles) like
drones has been extensively studied and applied in
SAR (Search and Rescue) operations, as presented on
(Dousai and Lon
ˇ
cari
´
c, 2022), (Svedin et al., 2021),
(Lygouras et al., 2019), (Kulkarni et al., 2020). This
technology helps save resources and makes it easier
to locate targets in hard-to-reach areas, natural dis-
asters, or accidents. Furthermore, the capability to
detect victims covered in mud or other debris dur-
ing searches makes multispectral sensors important
tools in this field of operation (Pensieri et al., 2020)
(Schoonmaker et al., 2010), and given that the hu-
man and animal body emits electromagnetic waves in
the infrared spectrum, the use of cameras in thermal
a
https://orcid.org/0000-0002-7578-559X
b
https://orcid.org/0000-0002-5562-1424
c
https://orcid.org/0000-0003-0515-0932
d
https://orcid.org/0000-0001-8844-5116
spectrum can greatly aid identification in dark envi-
ronments.
1.1 Related Work
Some studies are being conducted on target detection
in the infrared spectrum using YOLO. (Shen et al.,
2023) introduces an enhanced method called DBD-
YOLOv8 designed specifically to address challenges
associated with low signal-to-noise ratio and lack of
texture detail in infrared images. This method incor-
porates innovative modules like BiRA and Dyheads
to improve object detection accuracy across differ-
ent scales, and handle occluded and small objects.
Moreover, the proposed model significantly enhances
multi-scale feature representation, filters out irrele-
vant regions, and improves feature fusion, resulting
in a notable increase in average detection accuracy.
These improvements enable DBD-YOLOv8 to meet
real-time detection requirements, despite a slight in-
crease in model complexity and minor inference time
Correa, V. H. G., Funk, P., Sundelius, N., Sohlberg, R., Wahid, M. A. and Ramos, A. C. B.
Thermal Image Super-Resolution Using Real-ESRGAN for Human Detection.
DOI: 10.5220/0013078800003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 3: VISAPP, pages
247-254
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
247

overhead. The authors utilize various datasets tailored
for target detection during training.
The ITD-YOLOv8, as discussed in (Zhao et al.,
2024), is tailored for detecting infrared targets in com-
plex scenarios and across different scales, all while
reducing computational complexity. It incorporates
improvements in YOLOv8’s feature extraction back-
bone, including modules like GhostHGNetV2, AK-
Conv, VoVGSCSP, and the CoordAtt attention mech-
anism. These enhancements are aimed at enhancing
multi-scale feature extraction capabilities and effec-
tively detecting hidden infrared targets in challeng-
ing environments. Moreover, ITD-YOLOv8 intro-
duces the XIoU loss function to improve the accu-
racy of target localization, thereby reducing the rates
of missed and false detections. Experimental results
highlighted in the study show that ITD-YOLOv8 out-
performs YOLOv8n significantly by notably decreas-
ing the number of missed and false detections. Addi-
tionally, the model achieves a reduction in both model
parameters and floating-point operations. The aver-
age precision (mAP) achieved is 93.5%, confirming
the model’s efficacy in detecting infrared targets in
UAV applications.
(Luo and Tian, 2024b) presents YOLOv8-EGP,
which proposes improvements and optimizations
based on the original YOLOv8 to overcome chal-
lenges such as low detection accuracy, low robust-
ness, and missed detections in infrared images. The
enhancements include replacing the C2f module with
a more flexible and adaptive convolution module
called SCConv to improve feature diversity in the out-
put. It also incorporates the dyhead detection header
combined with multi-attention to enhance the expres-
sion capability of the detection head for infrared tar-
gets, along with adding a small target detection layer
(min) to reduce missed detections of small targets
and improve overall detection precision. These opti-
mizations and improvements have led to a significant
increase in detection accuracy, precision, and recall
compared to the original YOLOv8, demonstrating the
effectiveness of the enhanced model in detecting tar-
gets in the infrared spectrum.
(Luo and Tian, 2024a) incorporates the CPCA
attention mechanism to enhance the model’s focus
on specific areas of infrared images. The authors
replace the original downsampling layer with the
CGBD module to preserve edge information and ef-
fectively handle local and contextual features. Ad-
ditionally, they adopt the Weighted Intersection over
Union (WIoU) loss function to more accurately as-
sess target box coverage, thereby improving evalua-
tion precision. These improvements result in a 1.4%
increase in average precision (mAP) compared to the
YOLOv8s model, demonstrating significant enhance-
ments in precision and recall for detecting targets in
the infrared spectrum. Moreover, the enhanced model
is better suited for practical applications such as as-
sisted driving and road monitoring platforms.
(Wang et al., 2023) introduces the Small Tar-
get Detection (STC) structure in the network, which
serves as a bridge between shallow and deep features
to enhance semantic information gathering for small
targets and improve detection accuracy. Addition-
ally, the algorithm incorporates the Global Attention
Mechanism (GAM) to capture multi-dimensional fea-
ture information, thereby enhancing detection perfor-
mance by integrating features from different dimen-
sions. These improvements in small target detection
in the infrared spectrum are crucial for addressing
challenges faced by UAVs, such as detecting small
targets and mitigating the blur caused by high flight
speeds. Therefore, the algorithm proposed in this
article has the potential to significantly enhance ob-
ject detection in the infrared spectrum by UAVs, con-
tributing to improved performance in detecting small
targets in real-world scenarios of industrial inspection
by UAVs.
Tables 1 and 2 present the YOLO metrics re-
sults for related studies, this comparison should be
considered only for contextual understanding, as the
datasets in the studies presented in the tables differ.
The ITD-YOLOv8 model on (Zhao et al., 2024) was
trained using 300 epochs on the HIT-UAV dataset,
which consists of 2008 training images, 571 testing
images, and 287 validation images. The dataset in-
cludes three classes: people, bicycles, and vehicles;
The YOLOv8-EGP in (Luo and Tian, 2024b) was
trained over 300 epochs using a dataset of 10,467 in-
frared images. The dataset was divided into train-
ing, testing, and validation sets, with 7,326 images
for training, 2,094 images for testing, and 1,047 im-
ages for validation. The model proposed on (Luo and
Tian, 2024a) was also trained over 300 epochs using
the same dataset and configuration as (Luo and Tian,
2024b) and the classes of objects detected in the train-
ing dataset included person, bike, car, bus, light, and
sign. These classes were carefully selected to improve
the model’s generalization ability for real-world sce-
narios in infrared object detection tasks.
Table 1: Comparison of Precision(P) and Recall(R) metrics
from other studies. Here Eps. means the number of epochs
applied to YOLO training.
Study P(%). R(%). Eps.
(Zhao et al., 2024) 90.3 88.6 300
(Luo and Tian, 2024b) 85.6 74.0 300
(Luo and Tian, 2024a) 84.2 70.2 300
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
248

Table 2: Comparison of mAP50 and F1 Score metrics from
other studies.
Study mAP50(%) F1(%)
(Zhao et al., 2024) 93.5 89.4
(Luo and Tian, 2024b) 82.9 79.3
(Luo and Tian, 2024a) 78.2 76.5
1.2 Super-Resolution
Enhancing images is a critical preprocessing step for
raw data because factors like lighting conditions, sen-
sor movement during capture, and camera quality can
introduce artifacts and degradation. Improving these
images is essential to mitigate these issues for their
subsequent use, such as target detection in search and
rescue operations conducted by unmanned aerial ve-
hicles.
Classical algorithms for image enhancement, such
as Nearest Neighbour; Bilinear, Bicubic (Rahim et al.,
2015) and Lanczos (Fadnavis, 2014) increase an im-
age’s scale by interpolating pixel values through dif-
ferent mathematical methods. Nearest Neighbour du-
plicates the nearest pixel values, leading to a blocky
look, while Bilinear averages the four closest pixels
for smoother results. Bicubic interpolation, which
considers a 4x4 grid of pixels, offers even smoother
and sharper images, and Lanczos, using a sinc func-
tion, delivers high-quality results with minimal blur-
ring and aliasing. Importantly, these methods do
not involve artificial intelligence; they rely purely
on mathematical computations rather than learning-
based approaches. Each method has its strengths and
trade-offs, impacting the final image quality.
Regarding the use of Convolutional Neural Net-
works (CNNs) for super-resolution, especially Gener-
ative Adversarial Networks (GANs) applied to UAV
images, (Correa et al., 2024) provides a compre-
hensive systematic literature review of the applica-
tion of GANs to drone images, analyzing trends and
methodologies to improve Search and Rescue opera-
tions. The study highlights the effectiveness of GANs
in enhancing image quality through super-resolution,
which is crucial for improving target detection in SAR
missions. Additionally, it discusses the integration
of GANs with traditional object detection algorithms,
such as YOLO and Faster R-CNN, to enhance the
identification of targets in images captured by UAVs.
The authors also propose areas for further investiga-
tion, including the use of pre-trained models and real-
time applications of GANs in SAR operations, em-
phasizing the need for tailored datasets and method-
ologies. Overall, the findings suggest that GANs can
significantly improve UAV sensor capabilities, en-
abling more effective and timely identification of tar-
gets in various operational conditions.
1.3 Real-ESRGAN
A widely employed algorithm for image enhancement
that utilizes super-resolution techniques with GAN
approach is Real-ESRGAN (Wang et al., 2021b), an
upgraded work of the ESRGAN algorithm (Wang
et al., 2018). Its generator utilizes deep neural
networks with Residual-in-Residual Dense Blocks
(RRDB) to capture fine details and produce high-
quality images from low-resolution inputs. The train-
ing process involves comparing high-resolution im-
ages with their degraded counterparts to teach the
model how to reconstruct high-quality details. Real-
ESRGAN’s training is more complex than its pre-
decessor, ESRGAN, due to a broader range of im-
age degradations and the use of a U-Net architec-
ture combined with Spectral Normalization (SN) for
improved stability and performance. Additionally, a
pixel-unshuffle technique is employed to reduce com-
putational load by decreasing the spatial size of in-
puts and increasing the channel size, which optimizes
GPU memory usage. Real-ESRGAN’s training on
synthetic images allows it to handle diverse degra-
dation scenarios, enhancing its performance on real-
world images.
State-of-the-art (SOA) metrics like Precision, Re-
call, and mAP are critical for evaluating object detec-
tion performance, especially in tasks such as human
detection. These metrics, commonly used in models
like YOLO, provide insights into the accuracy and re-
liability of the detection system. Precision and Re-
call focus on identifying true positives and minimiz-
ing errors, while mAP measures overall performance
across multiple detection thresholds. In applications
like search and rescue, where accurate human detec-
tion is vital, SOA metrics are essential for tracking
and improving detection effectiveness.
In contrast, traditional image quality metrics like
Peak Signal-to-Noise Ratio (PSNR) and Structural
Similarity Index (SSIM) are often used to evaluate
super-resolution algorithms. While these metrics as-
sess pixel-level similarity between original and en-
hanced images, they may not fully reflect improve-
ments in visual quality, which are more relevant
for tasks like object detection. GAN-based super-
resolution methods often result in lower PSNR and
SSIM values (Xue et al., 2020); (Zhang et al., 2021);
(Wang et al., 2021a); (Lucas et al., 2019). However,
this does not indicate a flaw in the algorithm but high-
lights a trade-off: GANs are designed to optimize
perceptual quality, improving visual details that are
more important for object detection, even at the cost
Thermal Image Super-Resolution Using Real-ESRGAN for Human Detection
249

of lower pixel accuracy (Wang et al., 2020).
In this study, we sought to assess the effectiveness
of Real-ESRGAN in enhancing infrared images, as
this spectrum often exhibits higher noise levels and is
crucial for detecting human body heat signatures. We
present an overview of the YOLOv8 model applied to
our dataset, comparing the results between standard
and super-resolved image samples. We then analyze
and compare the super-resolved images produced by
Real-ESRGAN with those generated using traditional
non-AI super-resolution techniques, including Bilin-
ear, Bicubic, Lanczos, and Nearest Neighbor Interpo-
lation methods.
2 METHODOLOGY
To evaluate the performance of Real-ESRGAN us-
ing a pretrained model for enhancing infrared spec-
trum images, we propose the following methodologi-
cal steps:
1. Extraction of frames from our infrared video se-
quences to compile a dataset for analysis.
2. Application of Real-ESRGAN Algorithm on the
Dataset Samples previously obtained.
3. Assessing the effectiveness of image enhance-
ment by comparing person detection results using
YOLOv8 on both the original and enhanced im-
ages.
4. Comparison of the performance of Real-
ESRGAN with classical super-resolution
algorithms by evaluating Peak Signal-to-Noise
Ratio (PSNR) and Structural Similarity Index
(SSIM) metrics.
The initial phase of this research involved the cre-
ation of a dataset, which was constructed from a se-
ries of frames, each measuring 640x512 pixels. These
frames were extracted from three videos captured by
an infrared camera and represent the thermal signa-
tures of individuals in an outdoor setting. Figures 1
and 2 provides examples of the frames used in the
dataset construction.
In total, 10,043 frames ranging from 7 through 14
µm in wavelength were extracted from the 3 videos,
forming our initial dataset. Subsequently, we created
collections to facilitate our tests. Collection I was
composed of 500 randomly selected frames from the
first and second videos. Collection II consists of 100
randomly selected frames from all three videos. Col-
lection III consists of 100 randomly selected frames,
which were subsequently enhanced using the Real-
ESRGAN algorithm with 4x upscaling through the
RealESRGAN x4plus model.
Figure 1: Examples of frames captured from the videos
recorded by an infrared camera. Plot (a) is from the first
video, plot (b) from the second.
Figure 2: Examples of frames captured from the videos.
Both plots are from the third video.
Collection I was used as training data for detect-
ing people with YOLOv8. Except for collection I,
collections II and III, which were used for validating,
consist of frames extracted from all three videos.
After completing the super-resolution and collec-
tion creation phases, we proceeded with target detec-
tion using the YOLOv8 tool. During the annotation
process, we labeled all the targets as ”human” in the
dataset. We then selected an image from our dataset
and applied classical interpolation methods, including
Nearest Neighbor, Bilinear, Bicubic, and Lanczos, to
enhance the image. Each interpolation method was
evaluated for its impact on image quality. Next, the
same image was enhanced using the Real-ESRGAN
for comparison with the classical algoritms.
Following the enhancement process, we compared
the PSNR and SSIM metrics of the images enhanced
by the classical interpolation methods with that en-
hanced by the Real-ESRGAN algorithm.
3 RESULTS
After applying the Real-ESRGAN algorithm to our
dataset, we observed significant qualitative improve-
ments in the images. The enhanced images exhibited
sharper details, improved clarity, and more defined
features. Figure 3 illustrates a qualitative result of the
algorithm on one of the images.
Despite notable improvements in image clarity,
minimal contrast differences were observed in some
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
250

Figure 3: Example of satisfactory contrast enhancement in
image resolution by Real-ESRGAN. (a) Image before en-
hancement. (b) Image after upgrading (Correa, 2024).
noisy frames even with the increased scaling, as il-
lustrated in Figure 4 showing a struggle to improve
contrast in regions dominated by noise, as its focus
is on texture refinement rather than contrast enhance-
ment. Additionally, artifacts were present in specific
areas of the images, as depicted in Figure 5. These
artifacts, also observed in (Wang et al., 2021b), were
attributed to aliasing by the authors.
Figure 4: Example of minimal contrast improvement in an
image after super-resolution enhancement. (a) Image before
enhancement. (b) Same image after enhancement (Correa,
2024).
Figure 5: Artifacts observed in a region of an image after
applying the enhancement algorithm. (a) Original image
depicting a window. (b) Observed artifacts (Correa, 2024).
Tables 3 and 4 shows the performance of SOA
metrics from our tests. Test 1 entailed training with
the 500 frames from collection I and validating with
the 100 frames from collection II. Test 2 involved
training with the same 500 frames from collection I
and validating with the 100 frames from collection III.
We conducted training for 300 epochs in both tests 1
and 2.
Table 3: Comparison of Precision(P) and Recall(R) metrics
from tests 1 and 2. (Correa, 2024).
Study P(%). R(%). Eps.
Test 1 55.6 57.7 300
Test 2 92.9 71.6 300
Table 4: Comparison of mAP50 and F1 Score metrics from
tests 1 and 2 (Correa, 2024).
Study mAP50(%) F1(%)
Test 1 48.7 56.6
Test 2 83.4 80.9
Figures 6 and 7 display the evolution of the per-
formance metrics across epochs for Tests 1 and 2. In
these figures, we observe a significant improvement
in Test 2. Specifically, the images show higher values
for Precision, Recall, and mAP50 in Test 2. This indi-
cates that the model performed better in Test 2, where
the images were enhanced using the Real-ESRGAN
algorithm. The increased precision, recall, mAP50
and mAP50-95 suggest that the super-resolution tech-
nique contributed positively to the detection perfor-
mance, helping the model achieve more accurate and
reliable results in identifying human targets.
Figure 6: Precision and recall evolution for Tests 1 and 2
(Correa, 2024).
Based on the literature review by (Correa et al.,
2024), we found that PSNR and SSIM are commonly
used metrics in many studies that utilize GANs for
super-resolution. These metrics compare values be-
tween an original image and its modified version, val-
idating the degree of similarity between them. Thus,
Thermal Image Super-Resolution Using Real-ESRGAN for Human Detection
251

Figure 7: mAP50 and mAP50-95 evolution for Tests 1 and
2 (Correa, 2024).
we compared PSNR and SSIM across various interpo-
lation methods, including Nearest Neighbor, Bicubic,
Bilinear, and Lanczos. These classical interpolation
methods were implemented using the OpenCV library
and Python. The PSNR and SSIM metrics require
that the compared images be of the same size. Since
the original frames in the dataset have dimensions of
640x512 pixels, all interpolation methods were ap-
plied with a scaling factor of 4x to ensure that the met-
rics could be compared accurately against each other.
Tables 5, 6, 7, 8 and 9 present the results of the com-
parison of these metrics.
Table 5: Comparison of metrics against Bicubic Interpola-
tion (Correa, 2024).
Algorithm PSNR
dB
SSIM
Bicubic ∞ 1.00
Lanczos 47.97 0.99
Bilinear 41.60 0.98
NN 36.49 0.92
Real-ESRGAN 32.73 0.79
Table 6: Comparison of metrics against Bilinear Interpola-
tion (Correa, 2024).
Algorithm PSNR
dB
SSIM
Bilinear ∞ 1.00
Bicubic 41.60 0.98
Lanczos 40.50 0.97
NN 36.55 0.92
Real-ESRGAN 32.91 0.80
Table 7: Comparison of metrics against Lanczos Interpola-
tion (Correa, 2024).
Algorithm PSNR
dB
SSIM
Lanczos ∞ 1.00
Bicubic 47.97 0.99
Bilinear 40.50 0.97
NN 36.37 0.91
Real-ESRGAN 32.70 0.79
Table 8: Comparison of metrics against Nearest Neighbour
Interpolation (Correa, 2024).
Algorithm PSNR
dB
SSIM
NN ∞ 1.00
Bilinear 36.55 0.92
Bicubic 36.49 0.92
Lanczos 36.37 0.91
Real-ESRGAN 32.43 0.75
4 DISCUSSION
Infrared images often contain more noise, and the per-
ceptual improvements offered by Real-ESRGAN can
be particularly beneficial in enhancing features such
as human shapes or body heat, which are critical for
human detection. Despite struggling with some noisy
images and the presence of artifacts in certain frames,
as shown in our results, the use of Real-ESRGAN sig-
nificantly improved sharpness and body delineation,
demonstrating its effectiveness in enhancing key fea-
tures for detection tasks.
As shown in Figures 6 and 7, the metric curves for
Test 2 demonstrated better performance. This can be
attributed to the differing image qualities in Collec-
tion II and Collection III. In Test 1, where the model
was validated on low-resolution images, the model
encountered difficulties in extracting detailed features
due to the lower image quality. As a result, perfor-
mance during training was lower, as reflected in the
corresponding curves.
In contrast, Test 2, which used the super-
resolution images from Collection III for validation,
showed a significantly improved performance. The
enhanced images, with more detail and clarity, al-
lowed the detection model to more effectively extract
relevant features, leading to faster convergence and
better performance, as indicated by the higher pre-
cision, recall, and mAP50 scores from Tables 3 and
4 and curves from Figures 6 and 7. These findings
emphasize the impact of image resolution on object
detection performance.
Overall, the results demonstrate the importance of
high-quality images, especially in the case of infrared
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
252

Table 9: Comparison of metrics against Real-ESRGAN
super-resolution (Correa, 2024).
Algorithm PSNR
dB
SSIM
Real-ESRGAN ∞ 1.00
Bilinear 32.91 0.80
Bicubic 32.73 0.79
Lanczos 32.70 0.79
NN 32.43 0.79
imagery, where enhanced details are crucial for ac-
curate detection. This supports the effectiveness of
Real-ESRGAN as a preprocessing step for object de-
tection tasks, particularly when dealing with images
of varying resolutions.
While classical filters are effective for enhancing
image quality, they are limited to manipulating exist-
ing pixels through mathematical transformations. In
contrast, GANs reconstruct the image from the la-
tent space, making more substantial pixel-level mod-
ifications to enhance details. However, these algo-
rithms are constrained by the dataset used for training.
In our study, applying the Real-ESRGAN algorithm
with pre-trained models for infrared image enhance-
ment proved effective for some images, despite the
pre-trained model not being specifically related to our
dataset. This demonstrates a useful capability, partic-
ularly in situations where obtaining data for training
generative models is challenging.
YOLOv8 showed consistent performance and
metrics throughout the training process. The SOA
metrics show that, despite using a smaller dataset than
other studies, the super-resolution method achieved
high precision, recall, and mAP50 scores, underscor-
ing the effectiveness of Real-ESRGAN in image pre-
processing. Our key contribution lies in proposing
image enhancement as a preprocessing step prior to
object detection, especially for YOLO, which requires
intermediate steps such as image annotation and class
creation. We recommend enhancing the images first,
followed by annotation and YOLO training for human
detection.
By comparing Real-ESRGAN with classical inter-
polation algorithms, it can be observed in Tables 5,
6, 7, 8, and 9 that Real-ESRGAN resulted in lower
PSNR and SSIM values. The lower values for these
metrics have also been noted in other studies that uti-
lize GANs for super-resolution. While this might ini-
tially suggest that the algorithm is less effective, it is
important to note that classical interpolation methods
typically make only minor adjustments to the image,
preserving much of the original structure. In contrast,
Real-ESRGAN enhances the image by reconstructing
it from learned feature representations in the latent
space, which can result in significant modifications at
the pixel level. These modifications, while improving
perceptual quality, may introduce dissimilarities from
the original image, leading to lower PSNR and SSIM
values.
Compared to traditional interpolation methods,
Real-ESRGAN is significantly more computationally
intensive. Enhancing 1,000 frames requires approxi-
mately 90 minutes on standard hardware, such as an
RX570 graphics card. This extended processing time
presents a substantial limitation for real-time or auto-
mated applications, particularly in critical fields like
search and rescue operations. Therefore, further re-
search is necessary to optimize super-resolution algo-
rithms for real-time deployment.
5 CONCLUSIONS
This study enabled us to assess the use of a super-
resolution algorithm on a low-resolution dataset, with
a particular focus on improving thermal images for
human detection. By comparing the results before
and after data enhancement using YOLOv8 as the
benchmark, we observed a significant improvement
in both image quality and SOA metrics. The appli-
cation of the Real-ESRGAN algorithm demonstrated
considerable potential for enhancing thermal images,
which is especially beneficial for human detection in
UAV-based applications.
Our study also revealed that the image enhance-
ment process via Real-ESRGAN, which entails pixel
reconstruction and modification, can result in de-
creased PSNR and SSIM values. However, human
detection performance improved compared to using
the original images. This highlights the trade-off:
while the similarity metrics decreased, detection per-
formance increased, suggesting that the enhancement
process benefits detection tasks, even if traditional
quality metrics are lower.
The images in this study were captured at low al-
titudes using an infrared camera designed for short-
range detection, while drones typically operate at
much higher elevations. We also observed challenges
in enhancing some noisy images and the presence
of artifacts in certain regions, highlighting potential
flaws and areas for improvement in the algorithm.
Additionally, the dataset used was relatively small,
comprising 500 training images and 100 validation
images per test. As a result, future research could
focus on improving the quality of high-altitude im-
ages and expanding the dataset size to enable more
robust training, leading to more accurate results. Fur-
thermore, exploring other GAN-based algorithms for
super-resolution is recommended to address the noise
Thermal Image Super-Resolution Using Real-ESRGAN for Human Detection
253

and artifact issues, as different models may offer vary-
ing performance trade-offs. Additionally, training a
GAN from scratch, rather than relying on pre-trained
models, could offer alternative approaches that may
better suit the specific dataset and improve perfor-
mance.
ACKNOWLEDGEMENTS
The authors thank the Fundac¸
˜
ao de Amparo
`
a
Pesquisa do Estado de Minas Gerais (FAPEMIG),
project APQ00234-18, for the financial support that
made this research possible.
ChatGPT was carefully used to refine and clarify
the English expressions in the text, ensuring the ideas
were communicated accurately. Additionally, Chat-
PDF was thoughtfully employed to identify key sec-
tions of the references, which helped the authors ad-
dress the findings effectively and make well-informed
comparisons of the results.
REFERENCES
Correa, V., Funk, P., Sundelius, N., Sohlberg, R., and
Ramos, A. (2024). Applications of gans to aid tar-
get detection in sar operations: A systematic literature
review. Drones, 8(9):448.
Correa, V. H. G. (2024). Application of real-esrgan in im-
proving ir sensor images for use in sar operations.
Master’s thesis, Federal University of Itajub
´
a, Brazil.
Dousai, N. M. K. and Lon
ˇ
cari
´
c, S. (2022). Detecting hu-
mans in search and rescue operations based on ensem-
ble learning. IEEE access, 10:26481–26492.
Fadnavis, S. (2014). Image interpolation techniques in dig-
ital image processing: an overview. International
Journal of Engineering Research and Applications,
4(10):70–73.
Kulkarni, S., Chaphekar, V., Chowdhury, M. M. U., Erden,
F., and Guvenc, I. (2020). Uav aided search and res-
cue operation using reinforcement learning. In 2020
SoutheastCon, volume 2, pages 1–8. IEEE.
Lucas, A., Lopez-Tapia, S., Molina, R., and Katsaggelos,
A. K. (2019). Generative adversarial networks and
perceptual losses for video super-resolution. IEEE
Transactions on Image Processing, 28(7):3312–3327.
Luo, Z. and Tian, Y. (2024a). Improved infrared road object
detection algorithm based on attention mechanism in
yolov8. IAENG International Journal of Computer
Science, 51(6).
Luo, Z. and Tian, Y. (2024b). Infrared road object detec-
tion based on improved yolov8. IAENG International
Journal of Computer Science, 51(3).
Lygouras, E., Santavas, N., Taitzoglou, A., Tarchanidis, K.,
Mitropoulos, A., and Gasteratos, A. (2019). Unsuper-
vised human detection with an embedded vision sys-
tem on a fully autonomous uav for search and rescue
operations. Sensors, 19(16):3542.
Pensieri, M. G., Garau, M., and Barone, P. M. (2020).
Drones as an integral part of remote sensing technolo-
gies to help missing people. Drones, 4(2):15.
Rahim, A. N. A., Yaakob, S. N., Ngadiran, R., and Nas-
ruddin, M. W. (2015). An analysis of interpolation
methods for super resolution images. In 2015 IEEE
Student Conference on Research and Development
(SCOReD), pages 72–77. IEEE.
Schoonmaker, J., Reed, S., Podobna, Y., Vazquez, J., and
Boucher, C. (2010). A multispectral automatic tar-
get recognition application for maritime surveillance,
search, and rescue. In Sensors, and Command, Con-
trol, Communications, and Intelligence (C3I) Tech-
nologies for Homeland Security and Homeland De-
fense IX, volume 7666, pages 374–384. SPIE.
Shen, L., Lang, B., and Song, Z. (2023). Infrared object
detection method based on dbd-yolov8. IEEE Access.
Svedin, J., Bernland, A., Gustafsson, A., Claar, E., and
Luong, J. (2021). Small uav-based sar system using
low-cost radar, position, and attitude sensors with on-
board imaging capability. International Journal of Mi-
crowave and Wireless Technologies, 13(6):602–613.
Wang, F., Wang, H., Qin, Z., and Tang, J. (2023). Uav target
detection algorithm based on improved yolov8. IEEE
Access.
Wang, J., Gao, K., Zhang, Z., Ni, C., Hu, Z., Chen, D., and
Wu, Q. (2021a). Multisensor remote sensing imagery
super-resolution with conditional gan. Journal of Re-
mote Sensing.
Wang, X., Xie, L., Dong, C., and Shan, Y. (2021b). Real-
esrgan: Training real-world blind super-resolution
with pure synthetic data. In Proceedings of the
IEEE/CVF international conference on computer vi-
sion, pages 1905–1914.
Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao,
Y., and Change Loy, C. (2018). Esrgan: Enhanced
super-resolution generative adversarial networks. In
Proceedings of the European conference on computer
vision (ECCV) workshops, pages 0–0.
Wang, Z., Jiang, K., Yi, P., Han, Z., and He, Z. (2020).
Ultra-dense gan for satellite imagery super-resolution.
Neurocomputing, 398:328–337.
Xue, X., Zhang, X., Li, H., and Wang, W. (2020). Research
on gan-based image super-resolution method. In 2020
IEEE International Conference on Artificial Intelli-
gence and Computer Applications (ICAICA), pages
602–605. IEEE.
Zhang, X., Feng, C., Wang, A., Yang, L., and Hao, Y.
(2021). Ct super-resolution using multiple dense
residual block based gan. Signal, Image and Video
Processing, 15:725–733.
Zhao, X., Zhang, W., Zhang, H., Zheng, C., Ma, J., and
Zhang, Z. (2024). Itd-yolov8: An infrared target de-
tection model based on yolov8 for unmanned aerial
vehicles. Drones, 8(4):161.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
254