Subsurface Metallic Object Detection Using GPR Data and YOLOv8
Based Image Segmentation
Duarte Branco
1,2 a
, Rui Coutinho
1,2 b
, Armando Sousa
1,2 c
, and Filipe dos Santos
2 d
1
Faculty of Engineering (FEUP), University of Porto, Rua Dr. Roberto Frias, Porto, Portugal
2
INESC Technology and Science (INESC TEC), Rua Dr. Roberto Frias, Porto, Portugal
Keywords:
Ground Penetrating Radar, GPR, Object Detection, Subsurface Characterization, YOLO, Hyperbola Model.
Abstract:
Ground Penetrating Radar (GPR) is a geophysical imaging technique used for the characterization of a sub-
surface’s electromagnetic properties, allowing for the detection of buried objects. The characterization of
an object’s parameters, such as position, depth and radius, is possible by identifying the distinct hyperbolic
signature of objects in GPR B-scans. This paper proposes an automated system to detect and characterize
the presence of buried objects through the analysis of GPR data, using GPR and computer vision data pro-
cessing techniques and YOLO segmentation models. A multi-channel encoding strategy was explored when
training the models. This consisted of training the models with images where complementing data processing
techniques were stored in each image RGB channel, with the aim of maximizing the information. The hy-
perbola segmentation masks predicted by the trained neural network were related to the mathematical model
of the GPR hyperbola, using constrained least squares. The results show that YOLO models trained with
multi-channel encoding provide more accurate models. Parameter estimation proved accurate for the object’s
position and depth, however, radius estimation proved inaccurate for objects with relatively small radii.
1 INTRODUCTION
The characterization of a subsurface is an intrinsic
problem in many fields of science and engineering,
essentially everywhere there is a need to determine
what is beneath a visually opaque surface. Naturally,
non-destructive and non-intrusive techniques to solve
this problem are of great importance, where there is a
need to preserve the probing environment.
In this context, Ground Penetrating Radar (GPR)
has emerged as one of the most useful subsurface
probing methods. In essence, GPR is able to char-
acterize the electromagnetic response of a subsurface,
allowing for the detection of buried objects. Object
detection and characterization are possible by identi-
fying and analysing the distinct hyperbolic signature
of an object in GPR data ((Persico, 2014; Utsi, 2017;
Jol, 2009; Daniels, 2004)).
Traditionally, this process requires human inter-
pretation of the GPR data. However, analysing GPR
a
https://orcid.org/0009-0006-2039-8492
b
https://orcid.org/0000-0001-8821-2224
c
https://orcid.org/0000-0002-0317-4714
d
https://orcid.org/0000-0002-8486-6113
data is a difficult operation that requires relevant
knowledge of the operating principles of GPR, the
propagation of electromagnetic waves, and their inter-
actions with the subsurface and buried objects ((Per-
sico, 2014; Utsi, 2017; Jol, 2009; Daniels, 2004)).
For this reason, research has focused on systems to
model and analyze GPR data.
In general, these systems include a data process-
ing step followed by a hyperbola detection step and
parameter estimation considering the model of a GPR
hyperbola ((Zhou et al., 2018)).
The data processing step has the goal of mitigating
the effect of noise, cluttering signals and attenuation,
such that the hyperbola signature is easier to identify.
Traditional GPR data processing techniques include
dewow, timezero correction, frequency filtering, de-
convolution, background removal and gain ((Persico,
2014; Utsi, 2017; Jol, 2009; Daniels, 2004)). Re-
cently, other approaches based on subspace methods
and deep-learning have been proposed. Subspace-
based methods include singular value decomposition,
principal component analysis and independent com-
ponent analysis (Chen et al., 2019). In essence, these
aim to project the original B-scan data onto clutter
and target subspace. Deep learning methods con-
692
Branco, D., Coutinho, R., Sousa, A. and Santos, F.
Subsurface Metallic Object Detection Using GPR Data and YOLOv8 Based Image Segmentation.
DOI: 10.5220/0013027500003822
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 1, pages 692-699
ISBN: 978-989-758-717-7; ISSN: 2184-2809
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
sist in training neural networks, namely autoencoders
((Bestagini et al., 2020)) or Generative Adversarial
Nets (GANs) ((Ni et al., 2022)), to predict a noise-
free image.
Computer vision techniques have been used to fil-
ter noise and detect the hyperbola, using noise filter-
ing, thresholding, morphological operations and edge
detection based approaches ((Zhou et al., 2018)).
Several different types of machine learning algo-
rithms and neural networks have been used for the
task of hyperbola detection, including support vector
machines, classification models, different architec-
tures of convolutional neural networks and adversar-
ial neural networks ((Bai et al., 2023)). The You Only
Look Once (YOLO) family of networks is one of the
more advanced one-stage detection models ((Zheng
et al., 2023)), making it a compelling tool for hyper-
bola detection.
The goal of this paper is to implement a system to
detect and characterize objects in GPR data with an
approach based on exploring the training of YOLO
models by exploring a multi-channel encoding strat-
egy. Multi-channel encoding consists in storing dif-
ferent information in each of the three image RGB
colour channels. Thus, it’s possible to take full ad-
vantage of the YOLO networks, which support inputs
with three channels, by providing it with multiple B-
scans images, where different processing techniques
were applied, and combined in an RGB image.
The article starts by introducing background con-
cepts related to GPR operation and data analysis.
Next, the developed system is described in depth and
its results are exposed. Finally, the results are dis-
cussed and the performance of the system is evalu-
ated.
2 BACKGROUND
The working principles of GPR operation are simi-
lar to those of conventional radar but applied to a
different probing medium, the subsurface ((Persico,
2014; Utsi, 2017; Jol, 2009; Daniels, 2004)). The
most common configuration of GPR systems con-
sists of a transmitter and receiver antenna in a fixed
geometry, which is moved over the surface in or-
der to probe it. Essentially, GPR works by trans-
mitting electromagnetic waves into the subsurface,
which are then scattered and/or reflected by changes
in the medium’s electromagnetic properties which are
caused by dielectric discontinuities. The reflected
waves are picked up by the receiver antenna, allowing
for the characterization of the medium and the buried
objects’ depth and geometry. Depth is determined
by the time interval between the transmission and re-
ception of the electromagnetic (EM) waves and their
propagation velocity, which depends on the medium’s
EM properties. This is expressed by equation 1.
d = v
t
2
(1)
In simple terms, since reflection depends on the
object’s electrical and magnetic properties, this means
that GPR can detect any material provided these prop-
erties are sufficiently different from its surrounding
materials.
GPR data can be represented in different ways. An
A-scan is the representation of a single reflected EM
wave on a time-amplitude graph, plotted along an axis
which represents the wave’s travel time, which is pro-
portional to depth. The aggregation of consecutive A-
scans measured along a fixed direction, in this case,
the x-axis, results in a B-scan. The resulting B-scan
can be interpreted as a two-dimensional (2D) image
where the EM waves’ amplitude is represented using
brightness or colour modulation.
The signals transmitted by GPR antennas propa-
gate as ellipses with increasing areas as they prop-
agate deeper, resulting in a beam with an elliptical
cone shape. This phenomenon is described by equa-
tion 2, where E/2 represents the energy radius, H the
propagation distance, ε
r
the relative permittivity and
λ the signal’s wavelength ((Chang et al., 2009)). This
in turn will result in the reflection, and subsequent
recording, of EM waves before and after the GPR is
directly above the object of interest.
E
2
=
λ
4
+
H
ε
r
+ 1
(2)
When the GPR is directly above the object the EM
waves will be reflected along the direct path, corre-
sponding to the object’s depth. Before and after this,
the object will be detected via a longer path, result-
ing in a longer time interval between the transmission
and recording of the EM wave. If the object is cylin-
drical and perpendicular to the scan’s direction, when
plotted in a B-scan, this behaviour will result in a hy-
perbolic shape. The hyperbole’s apex indicates the
object’s depth, since it’s created by the EM wave re-
flected when GPR is directly above the object.
In essence, the GPR hyperbola can be modelled
mathematically by the geometrical parameters of the
scan and the time an EM wave takes to be reflected.
This model assumes some simplifications, namely, a
null offset between the GPR antennas, null distance
between the surface and the GPR. a homogeneous
medium with constant EM wave propagation veloc-
ity and a cylindrical object buried perpendicular to the
scan’s direction ((Chen and Cohn, 2010)).
Subsurface Metallic Object Detection Using GPR Data and YOLOv8 Based Image Segmentation
693
Figure 1: GPR hyperbola model of a cylindrical object, con-
sidering zero antenna offset, and the relationship with the
scan’s geometrical parameters (Chen and Cohn, 2010).
Considering figure 1, where Z and Z
0
represent
the distance travelled by the EM wave to an object of
radius R, and equation 1:
(x x
0
)
2
+ (Z
0
+ R)
2
= (Z + R)
2
(x x
0
)
2
+
vt
0
2
+ R
2
=
vt
2
+ R
2
(3)
By re-writing the previous expression, the GPR
hyperbola model can be represented by the general
hyperbola equation, given by equation 4. Thus, equa-
tion 3 is re-written as equation 5.
(y y
0
)
2
a
2
(x x
0
)
2
b
2
= 1 (4)
(t +2R/v)
2
(t
0
+ 2R/v)
2
(x x
0
)
2
(vt
0
/2 + R)
2
= 1 (5)
a = t
0
+
2R
v
t
0
= a
2R
v
(6)
b = v
t
0
2
+ R (7)
y
0
=
2R
v
(8)
Following these equations, an object’s position
corresponds to the horizontal position of the hyper-
bola apex, x
0
, and the depth is from the reflection
time, t
0
, and equation 1. The remaining parameters,
namely, the object’s radius and EM wave velocity are
related to the hyperbola parameters, by the following
equations:
v =
2b
a
(9)
R =
v(a t
0
)
2
R =
b(y
0
)
a
(10)
Consequently, by fitting the parameters of the
mathematical hyperbola model it’s possible, namely
a, b, x
0
and y
0
, it’s possible to determine the object’s
parameters and wave velocity. Fitting has been de-
fined following a constrained least-squares minimiza-
tion problem ((Zhou et al., 2018)).
While the hyperbola fitting process works well for
well-shaped hyperbolas ((Mertens et al., 2015; Gian-
nakis et al., 2022)), in real GPR data hyperbolas are
often miss-shaped. In these conditions, the parame-
ters estimated will be affected by an error, since they
will estimated by noisy points. In particular, the ef-
fect of noise is heavily detrimental to an accurate ra-
dius estimation. The remaining parameters, namely,
EM wave velocity and object depth and position can
be sufficiently estimated even considering noise.
Considering this, the radius can be estimated
through an alternative method, proposed in (Chang
et al., 2009). This method relates the number of A-
scans where a reflection is detected, given by the hy-
perbola width, L, and the energy radius, expressed in
equation 2. Thus, the radius is given by:
S = L E (11)
R =
S
2π
(12)
3 METHODOLOGY
The system’s architecture is characterized by three
main components: Image processing, Hyperbola de-
tection and Hyperbola fitting. Image processing was
done using both GPR and computer vision process-
ing techniques, with the goal of making it easier to
identify hyperbolic patterns. Hyperbola detection is
achieved using a YOLOv8 segmentation neural net-
work, trained with images with different processing
techniques in each image colour channel. Hyperbola
fitting was done using the segmentation masks pro-
vided by the network, with the goal of estimating the
object’s parameters using the GPR hyperbola model.
This step also includes a hyperbola mask processing
component, used to determine the hyperbola fitting
points. The high-level architecture of the system is
illustrated in figure 2.
3.1 Dataset Creation
A dataset comprising 294 GPR B-scans was collected
and annotated, which was further augmented using
image augmentation techniques, in order to train the
neural networks. The original scans were split into
training, testing and validation subsets, following an
80%, 10%, and 10% splits, respectively. The augmen-
tation techniques in CLAHE, flipping, resizing and
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
694
Figure 2: System architecture.
rotation ((Shorten and Khoshgoftaar, 2019)). In or-
der to generate a greater number of images, all com-
binations of these techniques were applied to the orig-
inal image and between the augmented images, with
the exception of resize and rotation. These were only
applied to images with at least one other technique.
After applying these transformations 8 images were
generated from an original scan. Augmentation was
applied to the training and validation datasets. A sec-
ond smaller dataset with the object’s, composed of
30-B-scans, depth and position ground-truth was col-
lected, in order to verify the precision of the devel-
oped system. For both datasets, objects consisted of
solid metallic cylinders with radii of 0.5, 1.4 and 1.5
cm. Scans were taken following straight lines where
objects were buried in soil perpendicular to the di-
rection of the scan. The GPR equipment used was
OERAD’s Concretto GPR system.
Data will be made available upon the acceptation
of the paper.
3.2 Data Processing and YOLOv8
Training
A sequential pipeline composed of the following GPR
processing techniques was applied to every scan: de-
wow, mean background removal and band-pass fil-
ter. The dewow window was set correspond to the
GPR pulse length and the cutoff frequencies of the
band-pass filter to correspond to the spectrum of the
GPR system. After this, the following processing
techniques were applied. Automatic gain control
(AGC), with a window set to correspond to the pulse
length, was implemented to mitigate the attenuation
effect. Computer vision techniques included tem-
plate matching, implemented through the normalized
cross-correlation (NCC) of A-scans with a pulse ref-
erence signal, and a Sobel edge detection filter. These
processes are illustrated in figure 3.
Different processing techniques were stored in
Figure 3: GPR processing applied to a B-scan (Left to
right): Original, Dewow, Background removal, Bandpass
filter, AGC, NCC, Sobel.
each RGB image channel, chosen with the goal of
providing complementary information to the neural
network. These multi-channel encoding configura-
tions are listed in table 1.
Table 1: Multi-channel encoding configurations (MCEC)
of different B-scan processing techniques, stored in the im-
age’s BGR channels.
MCEC B G R
1 GPR Processing GPR Processing GPR Processing
2 GPR Processing AGC -
3 GPR Processing NCC -
4 GPR Processing Sobel -
5 GPR Processing AGC NCC
6 GPR Processing AGC Sobel
7 AGC AGC AGC
8 AGC NCC -
9 AGC Sobel -
10 AGC NCC Sobel
11 NCC NCC NCC
12 Sobel Sobel Sobel
YOLOv8 nano models were trained for each con-
figuration, in order to determine which provided the
best results. Then a YOLOv8 medium model was
trained in order to determine if better results could
be obtained with a more complex model. These mod-
els were trained for 40 and 20 epochs, respectively,
using the default training hyperparameters, with the
exception of the learning rate, 0.0001, batch-size, 24,
Subsurface Metallic Object Detection Using GPR Data and YOLOv8 Based Image Segmentation
695
dropout, 0.2, optimizer, SGD and early stopping dis-
abled. After training, the best model was imple-
mented in the complete system to determine the hy-
perbola segmentation masks. The YOLO predictions
are illustrated in figure 4.
Figure 4: YOLO hyperbola mask predictions from the
trained YOLO model implemented in the system.
3.3 Model Fitting
The segmentation masks obtained from the YOLO
model were processed by an algorithm, in order
to obtain the points used in the fitting procedure.
Since YOLO predictions could be ill-shaped hy-
perbola masks, the algorithm was used to deter-
mine the correct hyperbola branches by considering
the skeletonized mask. The algorithm worked by
determining the intersection points where multiple
skeleton branches connected, and then determining
the biggest continuously connected branches, which
would match the actual hyperbola branches. Taking
into account the hyperbola’s shape, branches which
would have a horizontal overlap weren’t considered
when verifying branch connections. After determin-
ing the skeleton which effectively describes the hy-
perbola, the fitting points are obtained by considering
the mean point in each branch.
The fitting problem was defined as a constrained
least-squares problem, based on what’s described in
(Zhou et al., 2018). The general hyperbola equation
can be re-written as:
(y y
0
)
2
a
2
(x x
0
)
2
b
2
= 1
a
2
b
2
x
2
+ y
2
+ 2
a
2
b
2
x
0
x 2y
0
y + y
2
0
a
2
a
2
b
2
x
2
0
= 0
(13)
Thus, the problem function can be represented by
the following equations, which relate the fitting coef-
ficients with the hyperbola parameters:
F(i) = a
c
x
2
+ b
c
xy + c
c
y
2
+ d
c
x + e
c
y + f
c
= 0 (14)
a
2
= (
e
c
2
)
2
f
c
+
d
2
c
4a
c
b
2
=
(e
c
/2)
2
f
c
d
c
a
c
x
0
=
d
c
2a
c
y
0
=
e
c
2
(15)
Given the physical restrictions on the parameters
associated with the scan, it’s necessary to constrain
the minimization problem. Consecutively, bearing in
mind equations 9 and 10 and the system of equations
15, the coefficients are constrained such that the min-
imum radius was half of our GPR’s horizontal resolu-
tion, 0.5mm, and velocity had a positive value smaller
than the speed of light in vacuum. The fitting process
was implemented using the Pyomo software package
((Hart et al., 2011)) and the IPOPT solver ((W
¨
achter
and Biegler, 2006)).
Finally, the scan’s parameters were determined.
Due to the fact that the radius determined from hy-
perbola fitting is heavily affected by miss-shaped hy-
perbolas, the radius was calculated using the energy
radius method. The hyperbola width was obtained
from the width of the segmentation mask or from dou-
ble the distance from the x position of the determined
apex to the furthest x coordinate of a point belong-
ing to the fitted points. The biggest of these distances
would be considered the hyperbola’s width. This was
done with the goal of mitigating the effect of asym-
metric segmentation masks, where the full length of
both of the hyperbola branches wouldn’t be detected,
such that the mask’s width wouldn’t effectively cor-
respond to the actual width of the hyperbola. Ad-
ditionally, it was observed that multiple hyperbolic
reflections were created by an object. This is due
to the ringing effect, caused due to signals repeat-
edly bouncing within an object, which usually oc-
curs in metallic objects ((Persico, 2014; Utsi, 2017;
Jol, 2009; Daniels, 2004)). Consequently, all of the
hyperbolas from an object were grouped given their
apexes similar x position. The first-hyperbola, that is,
the one with smaller depth, correctly indicated the ob-
jects depth, while the repetitions echos, gave an exag-
gerated depth. All of the parameters were calculated
from the first hyperbolic reflection and also from a
mean of the parameters considering the first hyper-
bola and its echos, and the results were compared.
This was done with the goal of trying to take advan-
tage of the multiple hyperbola reflections and obtain-
ing a more accurate result, by mitigating the effect of
a miss-shaped hyperbola in the fitting process.
The fitting results are illustrated in figure 5.
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
696
Figure 5: Constrained least-squares fitting of GPR hyper-
bola model of a B-scan. Fitting points are shown in red,
fitting results are shown in green and the hyperbola apex is
shown in blue. Due to the ringing effect multiple hyperbo-
las are observed.
4 RESULTS
A concise summary of the YOLO model evaluation
metrics for the testing dataset is shown in table 2.
By analysing the evaluation metrics of the segmenta-
tion masks, in particular the precision, recall, mAP50
and mAP50-95, it’s possible to compare the differ-
ent models. For the sake of conciseness, each multi-
channel encoding strategy is referred to by its config-
uration, as listed in table 1.
Multi-channel configuration 6 stands out due to its
good performance both in the validation and testing
data. This configuration achieves the highest results
in many of the metrics and closely matches other con-
figurations when it’s not the best. Notably, its robust-
ness to false positives, given by the recall, stands out
in the testing dataset. The precision value, mAP50
and mAP50-95 are also some of the highest. For these
reasons, multi-channel encoding configuration 6 was
considered the best.
Next, a YOLOv8 medium model was trained with
this configuration. The metrics from training and val-
idation per epoch are illustrated in figure 7. The re-
sults after training was concluded are shown in ta-
ble 3. Considering the improved results in the recall,
mAP50, mAP50-95 and fitness for the testing dataset,
when compared to the YOLOv8 nano segmentation
model trained with the same multi-channel encoding
configuration, this model was used in the final sys-
tem. Considering the F1-score-confidence curve, a
confidence threshold of 0.15, was used to accept a
prediction. This rather low value was chosen since it
was considered more detrimental to have missed de-
tections rather than false positives.
In order to evaluate the success of the whole
system in estimating the scan’s parameters, the sys-
tem verification dataset was used. As previously
mentioned, the parameters were obtained considering
only the first hyperbola, and the first hyperbola and all
its repetitions.
The absolute error for the estimated parameters
is shown next. In order to better represent the esti-
mation error, some considerations were taken, such
that the estimated parameters are valid and the error
properly indicated the accuracy of estimation. For the
position error, all of the scans were considered when
calculating, since at least one true-positive hyperbola
was detected in all scans. For the depth error, only
scans where the first-hyperbola was correctly labelled
were considered. For the radius error, the previous
consideration was also used as well as the radius be-
ing greater than zero, since otherwise it would indi-
cate that the hyperbola’s width wasn’t properly deter-
mined.
5 CONCLUSION
In this paper, a system for object detection and char-
acterization in GPR data was implemented, which
used YOLOv8 segmentation model for hyperbola de-
tection and constrained least-squares fitting consid-
ering the GPR hyperbola model. The approach ex-
plored the combination of different data processing
techniques via a multi-channel encoding strategy that
had the goal of providing the model with a B-scan
image where complementing processing techniques
were applied. These techniques consisted of both
traditional GPR processing and also computer vision
methods.
In order to determine which multi-channel config-
uration produced the best detection results, multiple
YOLOv8 segmentation nano models were trained and
their metrics were compared. This analysis concluded
that the best multi-channel encoding configuration
consisted of the GPR processing steps, AGC and So-
bel filter. Next, a YOLOv8 segmentation medium
model was trained with this configuration achieving
improved results, thus being implemented in the final
system.
The segmentation masks provided by the trained
model were processed, to obtain the correct hyper-
bola branches from masks with potentially ill-shaped
masks. Next, constrained least squares was used to
obtain the hyperbola’s parameters considering the hy-
perbola model. The scan’s parameters, namely, the
object’s position, depth and radius, and the medium’s
EM wave propagation velocity were then estimated,
considering only the first-hyperbola, and the first-
hyperbola and its repetitions.
Subsurface Metallic Object Detection Using GPR Data and YOLOv8 Based Image Segmentation
697
Table 2: Testing metrics of YOLOv8 segmentation nano models on multi-channel encoding configurations (MCEC).
MCEC Precision Recall mAP50 mAP50-95 Fitness
1 0.8940 0.7063 0.8100 0.3140 0.8730
2 0.9219 0.6650 0.8135 0.3022 0.8722
3 0.7726 0.5714 0.6715 0.2236 0.6664
4 0.8172 0.6798 0.7828 0.2854 0.8226
5 0.8369 0.6650 0.7856 0.2863 0.8143
6 0.9136 0.7340 0.8382 0.3080 0.8776
7 0.9219 0.6650 0.8135 0.3022 0.8722
8 0.7757 0.6643 0.7561 0.2744 0.7798
9 0.8546 0.6897 0.8214 0.3088 0.8640
10 0.8283 0.6700 0.7623 0.2814 0.8043
11 0.7806 0.5783 0.6721 0.2006 0.6401
12 0.5793 0.5468 0.5296 0.1491 0.5416
Figure 6: Training and validation results of YOLOv8 segmentation medium model on multi-channel encoding configuration
6.
Figure 7: F1-score confidence curve considering validation
dataset on YOLOv8 segmentation medium model on multi-
channel encoding configuration 6.
Table 3: Validation and testing metrics of YOLOv8 seg-
mentation medium model on multi-channel encoding con-
figuration 6.
Metric Validation Dataset Testing Dataset
Precision 0.8801 0.8624
Recall 0.7733 0.7931
mAP50 0.8617 0.8770
mAP50-95 0.3461 0.3339
Fitness 0.9372 0.9394
The error values of the estimated parameters show
that an accurate result can be obtained for the position
and depth estimates. The results obtained considering
the hyperbola repetitions indicate that this method can
obtain a more accurate depth estimate. This is essen-
tially due to a better velocity estimate when consid-
ering the mean of multiple estimates of repeated hy-
perbolas. Similar error values occur for the position
estimate, considering both methods, which indicate
that the position when estimated from the average of
multiple hyperbola repetitions is also valid. However,
rather high results for radius are present, noting that
the objects used had radii of 0.5, 1.4 and 1.5 cm. Er-
ror in estimation velocity or hyperbola apex position
will affect the calculation of the energy radius, due to
the effect on estimated depth. Additionally, an error
in the value of the hyperbola’s width will also affect
the radius calculation.
The success of the entire system essentially de-
pends on the good performance of the YOLO model.
If the mask predictions don’t effectively match the hy-
perbola’s shape the estimation of all of the parameters
will be affected. Additionally, the system is also af-
fected by false positive and false negative detections.
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
698
Table 4: Average absolute error, , of valid estimated parameters. Medium’s EM wave propagation velocity, v, and object’s
position, x, depth, y, and radius, r. Calculated considering only first-hyperbola, 1st, and all hyperbolas from an object, group.
x
1st
(cm) x
group
(cm) y
1st
(cm) y
group
(cm) r
1st
(cm) r
group
(cm)
2.7 2.9 2.3 1.4 0.4 0.6
The developed system works as expected when
hyperbola segmentation provided well-fitting masks,
such that the parameters were estimated from fitting
points which correctly described the hyperbola.
ACKNOWLEDGEMENTS
This work is co-financed by Component 5 - Capital-
ization and Business Innovation, integrated in the Re-
silience Dimension of the Recovery and Resilience
Plan within the scope of the Recovery and Resilience
Mechanism (MRR) of the European Union (EU),
framed in the Next Generation EU, for the period
2021 - 2026, within project AgendaTransform, with
reference 34.
REFERENCES
Bai, X., Yang, Y., Wei, S., Chen, G., Li, H., Li, Y., Tian, H.,
Zhang, T., and Cui, H. (2023). A comprehensive re-
view of conventional and deep learning approaches for
ground-penetrating radar detection of raw data. Ap-
plied Sciences, 13(13):7992.
Bestagini, P., Lombardi, F., Lualdi, M., Picetti, F., and
Tubaro, S. (2020). Landmine detection using au-
toencoders on multipolarization GPR volumetric data.
IEEE Transactions on Geoscience and Remote Sens-
ing, 59(1):182–195.
Chang, C. W., Lin, C. H., and Lien, H. S. (2009). Measure-
ment radius of reinforcing steel bar in concrete using
digital image gpr. Construction and Building Materi-
als, 23(2):1057–1063.
Chen, G., Fu, L., Chen, K., Boateng, C. D., and Ge, S.
(2019). Adaptive ground clutter reduction in ground-
penetrating radar data based on principal component
analysis. IEEE Transactions on Geoscience and Re-
mote Sensing, 57(6):3271–3282.
Chen, H. and Cohn, A. (2010). Probabilistic robust hyper-
bola mixture model for interpreting ground penetrat-
ing radar data. Proceedings of the International Joint
Conference on Neural Networks, pages 1–8.
Daniels, D. J. (2004). Ground Penetrating Radar, 2nd
Edition. Institution of Electrical Engineers, London,
United Kingdom.
Giannakis, I., Zhou, F., Warren, C., and Giannopoulos, A.
(2022). On the limitations of hyperbola fitting for esti-
mating the radius of cylindrical targets in nondestruc-
tive testing and utility detection. IEEE Geoscience
and Remote Sensing Letters, 19:1–5.
Hart, W. E., Watson, J.-P., and Woodruff, D. L. (2011). Py-
omo: modeling and solving mathematical programs
in python. Mathematical Programming Computation,
3(3):219–260.
Jol, H. M. (2009). Ground Penetrating Radar Theory and
Applications. Elsevier Science, Amsterdam, Nether-
lands.
Mertens, L., Persico, R., Matera, L., and Lambot, S. (2015).
Automated detection of reflection hyperbolas in com-
plex gpr images with no a priori knowledge on the
medium. IEEE Transactions on Geoscience and Re-
mote Sensing, 54(1):580–596.
Ni, Z.-K., Shi, C., Pan, J., Zheng, Z., Ye, S., and Fang,
G. (2022). Declutter-gan: Gpr b-scan data clutter re-
moval using conditional generative adversarial nets.
IEEE Geoscience and Remote Sensing Letters, 19:1–
5.
Persico, R. (2014). Introduction to Ground Penetrating
Radar: Inverse Scattering and Data Processing. John
Wiley and Sons, Inc., Hoboken, New Jersey.
Shorten, C. and Khoshgoftaar, T. M. (2019). A survey on
image data augmentation for deep learning. Journal
of Big Data, 6:1–48.
Utsi, E. C. (2017). Ground Penetrating Radar: Theory and
Practice. Butterworth-Heinemann, Oxford, United
Kingdom.
W
¨
achter, A. and Biegler, L. T. (2006). On the imple-
mentation of a primal-dual interior point filter line
search algorithm for large-scale nonlinear program-
ming. Mathematical Programming, 106(1):25–57.
Zheng, X., Fang, S., Chen, H., Peng, L., and Ye, Z. (2023).
Internal detection of ground-penetrating radar images
using yolox-s with modified backbone. Electronics,
12(16):3520.
Zhou, X., Chen, H., and Li, J. (2018). An automatic gpr b-
scan image interpreting model. IEEE Transactions on
Geoscience and Remote Sensing, 56(6):3398–3412.
Subsurface Metallic Object Detection Using GPR Data and YOLOv8 Based Image Segmentation
699