sensors

Article
Fast Detection of Missing Thin Propagating Cracks during
Deep-Learning-Based Concrete Crack/Non-Crack Classification
Ganesh Kolappan Geetha 1 , Hyun-Jung Yang 2 and Sung-Han Sim 1, *

1 School of Civil, Architectural Engineering and Landscape Architecture, Sungkyunkwan University,

Suwon 16419, Republic of Korea
2 Smart Convergence Research Department, Power Technology Research Institute, KEPCO E & C,
Gimcheon 39660, Republic of Korea
* Correspondence: [email protected]

Abstract: Existing deep learning (DL) models can detect wider or thicker segments of cracks that
occupy multiple pixels in the width direction, but fail to distinguish the thin tail shallow segment
or propagating crack occupying fewer pixels. Therefore, in this study, we proposed a scheme for
tracking missing thin/propagating crack segments during DL-based crack identification on concrete
surfaces in a computationally efficient manner. The proposed scheme employs image processing
as a preprocessor and a postprocessor for a 1D DL model. Image-processing-assisted DL as a
precursor to DL eliminates labor-intensive labeling and the plane structural background without any
distinguishable features during DL training and testing; the model identifies potential crack candidate
regions. Iterative differential sliding-window-based local image processing as a postprocessor to
DL tracks missing thin cracks on segments classified as cracks. The capability of the proposed
method is demonstrated on low-resolution images with cracks of single-pixel width, captured using
unmanned aerial vehicles on concrete structures with different surface textures, different scenes with
complicated disturbances, and optical variability. Due to the multi-threshold-based image processing,
the overall approach is invariant to the choice of initial sensitivity parameters, hyperparameters, and
the sequence of neuron arrangement. Further, this technique is a computationally efficient alternative
to semantic segmentation that results in pixelated mapping/classification of thin crack regimes,
Citation: Kolappan Geetha, G.; Yang, which requires labor-intensive and skilled labeling.
H.-J.; Sim, S.-H. Fast Detection of
Missing Thin Propagating Cracks Keywords: deep learning; 1D-CNN; concrete crack and non-crack; thin crack classification; structural
during Deep-Learning-Based health monitoring; fast detection; image processing; UAV; image binarization
Concrete Crack/Non-Crack
Classification. Sensors 2023, 23, 1419.
https://doi.org/10.3390/s23031419

Academic Editor: 1. Introduction

Gilbert-Rainer Gillich Rapidly aging civil engineering structures require reliable scheduled maintenance at
Received: 29 December 2022
periodic intervals to increase the safety and longevity of critical load-carrying members.
Revised: 19 January 2023 Concrete cracks on the load transfer path grow continuously owing to cyclic loading. Hence,
Accepted: 22 January 2023 it is critical to identify thin propagating cracks from the perspective of civil engineering
Published: 27 January 2023 structures [1–3]. Conventional scheduled maintenance identifies concrete cracks using
visual inspection, which requires extensive manual observation and the training of inspec-
tors. To a great extent, the quantification parameters of defects depend on the inspectors’
judgmental skills, which are subjective and prone to error [4]. Hence, there is a need for
Copyright: © 2023 by the authors. robust non-destructive inspection (NDI) techniques that result in consistent outcomes for
Licensee MDPI, Basel, Switzerland. different types of test structures.
This article is an open access article NDI inspection methodologies employed by researchers for the inspection of thick
distributed under the terms and
structures/concrete comprise computer-vision-based visual inspection [5–9] and physics-
conditions of the Creative Commons
based NDI techniques, such as ultrasonics [10–13], radiography [14], and thermography [15–18].
Attribution (CC BY) license (https://
Although physics-based NDI techniques are promising for the diagnosis of defects at the
creativecommons.org/licenses/by/
coupon level, the complex necessity of tailor-made auxiliary devices for measurement and
4.0/).

Sensors 2023, 23, 1419. https://doi.org/10.3390/s23031419 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 1419 2 of 21

scalability for in situ field measurements remains questionable. The continuously evolving
semiconductor industry affords consumer-grade inexpensive vision-based sensors to NDI
inspectors. Computer-vision-based inspection enables rapid large-area measurement scala-
bility with image sensors that are easily configurable with robotic systems [17,19–23] or
unmanned aerial vehicles (UAVs) [24–26]. Multiple types of autonomous robotic systems,
such as ring-type climbing robots, camera-equipped mobile robots, and movable fixtures
with multiple degrees of freedom, have recently been employed to explore concrete struc-
tures and assist in maintenance [17,19–23,27]. Although robotic-platform-based inspection
is a promising way forward for automation and scalability, the following limitations exist:
(i) customized systems or auxiliary support systems are required for different structural
components and this is often unsuitable for inaccessible surfaces; (ii) robotic platforms
often have a short working distance from the target structure to achieve sufficient image
resolution for reliable crack detection, i.e., a small field-of-view (FOV), which limits the
range of coverage over a specified inspection duration; (iii) although the dimensions of the
crack can be accurately identified in image coordinates, a lack of reference or calibration
targets within a small FOV results in an erroneous estimation of the true dimensions of the
crack in world coordinates; (iv) regarding bridge components, robotic platforms are diffi-
cult to operate without additional surface preparation due to the continuous exposure to
degrading environmental conditions. UAVs overcome these limitations as they operate at a
long standoff distance from a target structure [24]. Although artificial-intelligence-assisted
UAVs are promising for covering a large area in a short duration, windy weather conditions
in the test field restrict UAVs from flying close to a target structure, thereby reducing the
mm/px resolution. In recent years, UAV-based structural assessments have gained consid-
erable attention owing to their enhanced mobility and superior data acquisition capability.
Postprocessing, segmentation, and classification of acquired images are often performed
using cloud services [26]. Although UAV-enabled edge computing [26] is promising, the
performance of cloud computing services limits the efficiency of UAV bridge inspection,
and often auxiliary costly onboard edge computing units are installed on UAVs, which is
limited to a handful of researchers.
Traditional digital-image-processing-based methods, such as edge detection, mor-
phological operations, digital image correlation, different variants of Otsu’s method, and
pattern matching [28,29], are computationally cost-effective. However, they rely on hand-
crafted selections and predesigned parameters, which hinder automated end-to-end crack
detection. Small, shallow, disconnected segments highlighted by conventional digital
image processing are often misinterpreted as background noise and require careful han-
dling. In the last decade, researchers have worked on computer-vision-based automated
crack inspection for applications in concrete bridge structures [6,30], pavements [31], and
buildings [32]. References [33,34] and references therein exhaustively reported the ad-
vancement of conventional machine-learning- and deep-learning (DL)-assisted computer
vision for automated concrete crack detection. One of the critical challenges in automated
crack detection is identifying missing thin/propagating cracks during DL-based concrete
crack/non-crack classification.
Close observations of concrete cracks, as reported in the literature [35–37], revealed
that the thin faint tail of a crack often follows a thick segment of the crack. The reported
results indicated a drastic reduction in the detection accuracy for thin cracks with widths of
approximately 1–2 px. Concrete cracks propagate over time, and the intrinsic and extrinsic
factors near the crack tip determine the propagation rate [38]. The thin tail of propagating
cracks often occupies only a few pixels or a single pixel in width, whereas broader or thicker
segments of the crack occupy multiple pixels with distinguishable features. The contrast
between the thin crack signature and the surrounding background is negligible, and the
odds of detecting these features require a microscopic lens with a smaller FOV [24]. For
instance, in the case of bridge pier maintenance, as per the standards of ‘Korean bridge
inspection and maintenance’, any crack greater than 0.3 mm should be detected. Although
state-of-the-art image processing and machine learning techniques have shown promising
Sensors 2023, 23, 1419 3 of 21

results in detecting cracks, the reported schemes often fail to detect thin segments of
propagating cracks with a single-pixel width. Thus, early diagnosis and maintenance of
such thin propagating cracks will increase the corresponding structural lifespan at the
component and system levels.
Various image processing schemes and conventional machine-learning-based feature
extractions for crack detection are comprehensively discussed in [7,39–42] and [43–46],
respectively. Irrespective of the choice of state-of-the-art algorithms, the derived results
were significantly influenced by field noise in terms of surface textures [47], lighting
conditions [48], and shot noise [49]. These image processing and machine-learning-based
algorithms are not universal; they are contextual and require a priori information on
the potential feature discriminant. Further, although machine learning methods perform
well, they require a large number of structured labels. Extracting handcrafted features at
crack boundaries to differentiate between crack and non-crack images is challenging [46].
In this regard, supervised DL has proved to be a possible way forward for real-world
applications, wherein knowledge of distinct feature extraction is automated. State-of-
the-art DL models account for unforeseen scenarios by considering data for different
environmental, optical, and structural variabilities. Holistic reviews of AI-based automated
surface crack classifications are provided in [34,36,50–53]. Convolutional neural network
(CNN)-based architectures used for crack detection are briefly overviewed in [33]. Every
implemented architecture performed well on the available datasets. However, supervised
DL with a constant sliding window, bounding boxes, and pixel-level segmentation requires
labor-intensive labeling and is computationally costly owing to the deeper layers. Moreover,
labeling the ground truth in semantic segmentation is subjective to human expertise. Hence,
they are unsuitable for real-time applications implemented using robotic systems or UAVs
with limited computational facilities. Although semantic segmentation is widely used for
the pixel-level classification of concrete crack features, one of the major limitations is that
the detected width of a concrete crack is much wider than the ground truth. This attribute
is often associated with a wider width of ground-truth labeling during the training of the
DL model to reduce false positives and false negatives.
Recently, researchers have employed a combination of digital image processing
schemes and DL for automated crack classification [54–57]. In [54,56], image processing was
employed to assist DL, while [55,57] used DL to assist in image processing. Additionally,
Kim et al. [54] used image processing to extract potential crack candidate regions (CCRs)
and filter background features, thereby reducing the volume of data required for the DL
model. Golding et al. [56] transformed RGB images to grayscale before DL implementation
using a pretrained VGG16 architecture, thereby reducing the computation associated with
different input channels. Nomura et al. [55] and Yu et al. [57] first extracted bounding boxes
of cracks using YOLO-based object detection. Subsequently, the former used a morphologi-
cal algorithm to pixelate the defect, and the latter used a mask filter to eliminate statistical
noise and connected component optimization methods with a global threshold to highlight
cracks from the background. Image processing with globally sensitive parameters [55,57] is
ineffective for field testing involving UAVs with long standoff distances, where environ-
mental and optical variability are significant, and when the discriminating crack intensity
is shallow compared with the background. However, the current state of the literature
combining digital image processing schemes and DL still relies on 2D matrix operations
during the forward and backward propagation of the DL model.
Despite the significant success of DL, certain segments of thin shallow propagating
cracks are often not identified using state-of-the-art DL-based algorithms. However, [58]
(i) overcame the limitations of labor-intensive labeling by identifying potential CCRs,
wherein image processing was employed as a precursor to the DL model; (ii) identified
CCRs to solve the problems associated with big data analysis on a robotic platform; and
(iii) achieved computational robustness using a shallow 1D CNN. The potential CCRs
in the Fourier basis identified using image preprocessing are fed into the DL model to
extract crack information inherently in the frequency domain with a smaller number of
Sensors 2023, 23, 1419 4 of 21

neurons and layers. Although state-of-the-art DL shows promising results, it often fails
to detect thin propagating cracks with the width of a single pixel in a computationally
efficient manner.
The contributions of this study are as follows: To overcome the limitations of state-
of-the-art CNNs, pixel-level segmentation, and to utilize the potential of one-dimensional
Fourier-based CNNs [58], the authors proposed a scheme considering multi-threshold-
based image processing combined with a 1D frequency-domain DL model to detect cracks
including thin propagating shallow cracks with a single-pixel width. The proposed scheme
tracked missing thin crack segments during DL-based crack identification on a concrete
surface in a computationally efficient manner. With image processing, as a preprocessor
and postprocessor to the neural network, the overall approach is invariant to the choice of
initial sensitivity parameters, hyperparameters, and the sequence of neuron arrangement.
The remainder of this paper is organized as follows: Section 2 discusses the back-
ground theories on (i) adaptive-threshold-based integral image binarization for image
preprocessing for potential CCR identification and (ii) frequency-domain-based 1D CNN
to extract crack discriminant features. In Section 3, an iterative scheme is detailed for
tracking thin-propagating cracks with different sliding windows as a postprocessor. In
Section 4, we demonstrate the applicability of the developed algorithm for field data from
bridge piers. In addition, we discuss the effects of various image-processing parameters
and their effects on the predicted results. Finally, we summarize and conclude the study
in Sections 5 and 6, respectively.

2. Theory and Method

Here, we concisely discuss the background theory directly used in our proposed
implementation scheme: (i) local-threshold-based image processing is used as a precursor
to DL to eliminate labor-intensive labeling, and as a postprocessor to track missing shallow
propagating crack segments that are often missed by the DL network; (ii) a shallow 1D
CNN-based architecture in the frequency domain is adopted to extract discriminant crack
features and distinguish cracks in CCRs.

2.1. Image Binarization

Although nonlinear neural networks have proven successful in distinguishing crack
features, analyzing high-definition (HD) concrete images is computationally expensive
when using a deep neural network with multiple cascading layers. The area occupied by
the crack or non-crack features in a HD image is generally less than 1%. Local threshold-
based intelligent image preprocessing prior to DL enables the identification of discriminant
features, such as CCRs, which prevents the background without any structural features
from being fed into the DL. Image binarization as a preprocessor eliminates the large
area with a plane structural background, reducing the computational burden during the
forward and backward propagation of the DL model. Supervised DL techniques require
labeled datasets for training, so physically labeling segments of cracks and non-cracks in the
presence of a background is cumbersome and prone to human error. Image processing, as
a preprocessor, eliminates labor-intensive labeling of the plane structural background and
associated human errors. Among various image processing schemes, image binarization
effectively filters out background features from discriminant features; image binarization
pixelates an RGB image into binary pixel space, and an optimal threshold can be deter-
mined that is universally applicable across all sets of training and testing images. Optimal
threshold estimation using global approaches based on Otsu’s method [59] and its improvi-
sation [60–62] fails to detect crack features when the histogram distribution of pixels in the
grayscale (0–255) is unimodal and under non-uniform lighting conditions. However, local
thresholding with a sliding window of constant size overcomes these limitations [58,63,64].
Sensors 2023, 23, x FOR PEER REVIEW 5 of 21

Sensors 2023, 23, 1419 5 of 21

However, local thresholding with a sliding window of constant size overcomes these lim-
itations [58,63,64].
Below, we briefly discuss the theory of local-adaptive-threshold-based integral image
Below, we briefly discuss the theory of local-adaptive-threshold-based integral image
binarization [58,63]. An integral image 𝐼(𝑥, 𝑦) (called the summed-area table) computed
binarization [58,63]. An integral image I ( x, y) (called the summed-area table) computed
for each pixel (𝑥, 𝑦) is given by:
for each pixel ( x, y) is given by:
𝐼(𝑥, 𝑦) = 𝑓(𝑥, 𝑦) + 𝐼(𝑥 − 1, 𝑦) + 𝐼(𝑥, 𝑦 − 1) − 𝐼(𝑥 − 1, 𝑦 − 1) (1)
I ( x, y) = f ( x, y) + I ( x − 1, y) + I ( x, y − 1) − I ( x − 1, y − 1) (1)
where 𝑓(𝑥, 𝑦) denotes pixel intensity at (𝑥, 𝑦). A schematic of the integral image compu-
tation
whereforf ( x,each pixel (𝑥,
y) denotes 𝑦) is
pixel shown at
intensity in (Figure
x, y). A1a. The cumulative
schematic sum over
of the integral the
image ABCD
computa-
region is computed as
tion for each pixel ( x, y) is shown in Figure 1a. The cumulative sum over the ABCD region
is computed as𝑥2 𝑦2
𝑓𝑠 (𝑥, 𝑦) = ∑ ∑x2 𝑓(𝑥, y2 𝑦)
) =1 𝑦=𝑦∑
f s ( x, y𝑥=𝑥 1 ∑ f ( x, y) (2)
x = x1 y = y1 (2)
= 𝐼(𝑥= I (2x) 2−, y𝐼(𝑥
2, 𝑦 2, 𝑦
2) − I1( x−2 ,1)
y1 − 1) 1−−I (1,x1𝑦2−) 1,
−𝐼(𝑥 + y𝐼(𝑥
2 ) 1+−I (
1,x𝑦
1 1−−1,1)
y1 − 1)

( 1, 1) 1, 1) ( , 1)

( , )
( 1, ) ( , )
, )

(a) (b)

(c) (d)
Figure 1. (a) Schematic representation for an integral image, and (b) results for local-threshold-based
Figure 1. (a) Schematic representation for an integral image, and (b) results for local-threshold-based
image binarization. (c) Filtering surface texture and optical noise using aspect ratio of crack, and (d)
image binarization. (c) Filtering surface texture and optical noise using aspect ratio of crack, and
crack candidate regions (CCRs) mapped (marked in red color) to original RGB image.
(d) crack candidate regions (CCRs) mapped (marked in red color) to original RGB image.
For
Foraa square
square window
window of of size 𝑠, the
size s, the local threshold T𝑇(𝑥,
local threshold ( x, y𝑦)
) isis
𝑓 (𝑥,𝑦) 𝑡
( 1 𝑇(𝑥, 𝑦) > f 𝑠( x,y) × (1 − t )

𝑇(𝑥, 𝑦) =1{ T ( x, y) > ss𝑠×𝑠 × 1 − 100 (3)
T ( x, y) = 0 otherwise ×s 100
(3)
0 otherwise
where 𝑇(𝑥, 𝑦) is the binarized value and 𝑡 is the sensitivity factor. Representative results
where
showing T ( x, y) is the
binarization binarized
involving value
cracks, and t is the
non-cracks, andsensitivity
noise arefactor. Representative
presented re-
in Figure 1b.
sults showing binarization involving cracks, non-cracks, and noise are presented in Figure 1b.
Image binarization computes distinct values for each pixel ( x, y), which are governed
by the local window size and sensitivity factor t (Equation (3)). Equations (1)–(3) can be
Sensors 2023, 23, 1419 6 of 21

approximately physically interpreted as the moving average of s’ pixels surrounding every

pixel ( x, y). The adaptive thresholding scheme with a sliding window accounts for spatial
illumination variation and thus the approach is invariant to nonuniform lighting conditions.
However, a high sensitivity factor introduces many statistical features associated with the
surface texture and optical noise [49]; hence, a tradeoff is required to reduce the subsequent
analysis. Post-local image binarization (Figure 1b) and connected component analysis [65]
extracts a single stretch of pixels without any discontinuity. The pixelated residue was
filtered using the aspect ratio of the crack features [54]. A detailed discussion can be
found in Ref. [58]. It is impossible to derive a universal value for t that works well for
every image. The threshold value is often derived from heuristics that largely satisfy the
entire dataset under consideration in concrete structural feature detection. Subsequently,
the bounds of the pixelated residue were extracted with bounding boxes that enclose the
maximum pixel from left to right and from top to bottom, called CCRs, which include
crack and non-crack features. The area of CCRs is not uniform and varies depending on
their features. Although the training dataset for DL requires labeling CCRs into cracks and
non-cracks, the amount of labeling is significantly reduced compared with up-to-date DL
techniques. Figure 1b–d) display the representative results for local image preprocessing
and subsequent postprocessing.

2.2. Frequency Domain 1D CNN

In the case of a 2D CNN, the kernel slides in 2D data, whereas the kernel slides in
1D for the case of a 1D CNN. This reduces the parameters involved in the DL model and
the computational time for training and testing. The advantages of the 1D DL model are
as follows: (i) matrix computations associated with gradient-descent-based optimization
during forward and backward propagation are replaced with vector array operations;
(ii) the 1D DL model learns hidden features with shallow architectures, whereas the 2D
DL models require deeper layers to learn hidden features; (iii) 2D DL models often require
cloud computing or high-performance GPUs; in contrast, the computational efficacy of 1D
CNNs enables their implementation on standard computers, mobile devices, or handheld
devices. Although the superior performance of the 1D CNN has been well reported in [44],
studies on its application in concrete crack detection are limited to that reported in [58].
Reference [66] reported the potential for crack feature extraction in the frequency
domain. Motivated by [66], we input the frequency domain characteristics of CCRs into
the DL model [58]. The authors utilized the advantages of distinct crack signatures in the
frequency domain for feature extraction in a 1D DL network. This advantage effectively
enables a 1D CNN with a shallow architecture and a smaller number of neurons to distin-
guish cracks from those in the region of interest. The 2D discrete Fourier transform (DFT)
(Figure 1a) of an image of size M × N is expressed as:

1 ux vy
F̂ (u, v) =
MN ∑ xM=−01 ∑ yN=−01 f (x, y)e− j2π( M + N ) (4)

ux vy
where f ( x, y) is the pixelated image and e− j2π ( M + N ) is the basis function for each point
F̂ (u, v). The complex-valued Fourier transform includes both phase and magnitude in-
formation. We performed frequency shifting to transform all the higher magnitudes to
the center of the basis system. The transformed 2D pixel space to Fourier basis space
information is mutually orthogonal and independent. Subsequently, we vectorized the 2D
Fourier space into a 1D space and fed it to the 1D CNN. The 2D CCRs of unequal areas were
resized to a uniform size prior to 2D DFT. CCRs greater than or less than 256 px × 256 px
were resized or padded with zeros, thereby normalizing the input size for the 2D DFT. The
authors implemented a frequency-domain-based 1D CNN for fast classification of surface
cracks with a processing rate of approximately 60 images/s [58]. The 1D DFT-CNN was
trained on 1492 crack and 1321 non-crack images. The chosen database includes concrete
images from different bridge piers that account for the effects of optical variability in
terms of lighting conditions, standoff distance between the camera and target structure,
Sensors 2023, 23, x FOR PEER REVIEW 7 of 21

Sensors 2023, 23, 1419 7 of 21

was trained on 1492 crack and 1321 non-crack images. The chosen database includes con-
crete images from different bridge piers that account for the effects of optical variability
in terms of lighting conditions, standoff distance between the camera and target structure,
focal lengths, FOV, and lens. The configuration of the 1D DFT-CNN-based architecture is
focal lengths, FOV, and lens. The configuration of the 1D DFT-CNN-based architecture is
depicted in Figure 2. The architecture includes four convolutional layers followed by three
depicted in Figure 2. The architecture includes four convolutional layers followed by three
fully connected layers. A detailed discussion of the parameters of the architecture can be
fully connected layers. A detailed discussion of the parameters of the architecture can be
found in [58]. The optimal number of epochs for training and validation was determined
found in [58]. The optimal number of epochs for training and validation was determined
from the asymptotic behavior of the binary entropy loss L BE (y, ŷ), given by:
from the asymptotic behavior of the binary entropy loss 𝐿𝐵𝐸 (𝑦, 𝑦̂), given by:
L BE (y,𝐿ŷ) = ylog ŷ + (1 −
̂ +y)(1
log 1 − ŷ)(1 − 𝑦̂)
−(𝑦)log (5)
𝐵𝐸 (𝑦, 𝑦
̂) = 𝑦𝑙𝑜𝑔𝑦 (5)
where y𝑦and
where 𝑦̂ are
andŷ are thethe output
output andand
thethe predicted
predicted output,
output, respectively.
respectively.

Figure 2. Architecture
Figure Architectureof of
1D1D
discrete Fourier
discrete transform
Fourier (DFT)(DFT)
transform convolutional neuralneural
convolutional network (CNN)
network
[58]. [58].
(CNN)

3.
3. Iterative
Iterative Postprocessing
Postprocessing Scheme
Scheme forfor Classification
Classificationof ofMissing
MissingThin Thin Propagating
Propagating Cracks
Cracks
In the current work, we employed a combination of image-processing-assisted 1D
In the current work, we employed a combination of image-processing-assisted 1D
DL and 1D DL-assisted image processing to detect thin propagating shallow cracks. The
DL and 1D DL-assisted image processing to detect thin propagating shallow cracks. The
scheme iteratively uses a differential sliding window, wherein we employ integral image
scheme iteratively uses a differential sliding window, wherein we employ integral image
binarization with an increasing threshold to track missing crack segments. The overall
binarization with an increasing threshold to track missing crack segments. The overall
framework (Figure 3) for crack identification uses image binarization, first as a preprocessor
framework (Figure 3) for crack identification uses image binarization, first as a preproces-
to the recently developed 1D frequency-domain-based DL by the authors, and finally as a
sor to the recently developed 1D frequency-domain-based DL by the authors, and finally
postprocessor to DL to track missing thin shallow cracks with widths less than or equal to
as a postprocessor to DL to track missing thin shallow cracks with widths less than or
a single pixel. This approach combines the advantages of image processing and inherently
equal to a single pixel. This approach combines the advantages of image processing and
extracts distinct crack features from a frequency-domain-based CNN. Image processing, as a
inherently extracts distinct crack features from a frequency-domain-based CNN. Image
preprocessor, identifies potential CCRs and significantly reduces the structural background
processing,
without as astructural
distinct preprocessor, identifies
features potentialDL-based
in subsequent CCRs and significantlyThis
computations. reduces the
enables
structural background without distinct structural features in subsequent
the replacement of computationally intensive matrix calculations involved in forward DL-based com-
putations.
and Thispropagation
backward enables the replacement
in a CNN byofacomputationally
1D array-based intensive matrixnetwork
shallow neural calculations
for
DL-based hidden feature extraction. Post DL, we performed an iterative schemeshallow
involved in forward and backward propagation in a CNN by a 1D array-based with a
neural network
sliding window of forimage
DL-based hiddentofeature
processing extraction.
track thin Post DL,
propagating crackwesegments,
performed an itera-
which are
tive scheme with a sliding window of image processing to track thin propagating
often missed by state-of-the-art DL-based schemes (Figure 4). Multi-threshold-based image crack
segments, which
processing are oftenand
as a precursor missed by state-of-the-art
postprocessor providesDL-based
invariance schemes (Figure 4). Multi-
in the framework to the
threshold-based
choice image processing
of initial sensitivity parameters as during
a precursor
imageand postprocessor
processing, provides invariance
hyperparameters, and the
in the framework
number of layers intothe
theDL
choice of initial sensitivity parameters during image processing,
model.
hyperparameters, and the number of layers in the DL model.
Sensors 2023, 23, 1419 8 of 21
Sensors
Sensors2023,
2023,23,
23,xxFOR
FORPEER
PEERREVIEW
REVIEW 88 ofof 21
21

Figure
Figure3.3.Overall
Overallalgorithmic
algorithmicscheme
schemefor
foridentifying
identifyingthin
thinpropagating
propagatingcracks,
cracks,wherein
whereinimage
imagepro-
pro-
Figure 3. Overall algorithmic scheme for identifying thin propagating cracks, wherein image process-
cessing
cessingisisused
usedas
asaapreprocessor
preprocessorand
andpostprocessor
postprocessortoto1D
1Dfrequency-domain-based
frequency-domain-baseddeep
deeplearning
learning
ing(DL).
is used as a preprocessor and postprocessor to 1D frequency-domain-based deep learning (DL).
(DL).

Figure
Figure
Figure 4. 4.Algorithm
4.Algorithm
Algorithmforfor 1D
for1D DL-assisted
DL-assistedimage
1DDL-assisted imageprocessing
image processing
processing for
for identification
foridentification
identificationofofthin
thin propagating
propagating
of thin propagating
cracks
cracks post
post DL-based
DL-based crack and
crackand non-crack
non-crackclassification.
andnon-crack classification.(a)
(a)Identify
Identify endpoints
endpoints ofofpixelated
pixelated residue
residue
cracks post DL-based crack classification. (a) Identify endpoints of pixelated residue
ininthe
thecrack
cracksegments
segmentsthatthatthethe1D
1DDFT-CNN
DFT-CNNmodel modelclassifies
classifies(markers
(markersindicates
indicatesthetheendpoints
endpointsofof
in pixelated
the crackresidue;
segments that the 1D DFT-CNN model classifies (markers indicates the endpoints of
pixelated residue;red redcolor
colorrectangle
rectangleisisthe
thebounds
boundsofofthe thepixelated
pixelatedresidue;
residue;circle
circleindicates
indicatesregion
region
pixelated
ofof zoom
zoom residue;
window).red(b)
window). color rectangle
(b) Define
Define slidingis the
sliding bounds
window
window (redofcolor
(red the
colorpixelated
square) residue;
square) with
with each
eachcircle indicates
endpoint
endpoint (redregion
(red star of
star
marker)
zoom marker) as
window).asthe centroid
the(b) Define
centroid ofofsliding
the
thewindow.
window
window. (c)
(c)Map
(redand
Map andextract
color square)
extract the corresponding
thewith
corresponding sliding
each endpointsliding window
(red star marker)
window inin
as the
theraw
the rawimage.
image.
centroid ofAdaptive
Adaptive integral
the window. (c)binarization
integral binarization on
onsliding
Map and extract thewindow
sliding window with
with(d)
corresponding (d)original
slidingsensitivity
original window factor
sensitivity factor
in the raw
image. Adaptive integral binarization on sliding window with (d) original sensitivity factor (t) and
(e) differential sensitivity factor (t + ∆t). (f) Filtered sliding window using geometric properties of
crack. (g) Stitch sliding window in pixelated image to connect missing segments of crack.
Sensors 2023, 23, 1419 9 of 21

In the overall approach, we first generated CCRs using image processing, wherein
adaptive-threshold-based integral image binarization was employed. A large portion of the
CCRs arising from shot noise, surface texture, and statistical features associated with image
binarization were filtered using the geometric properties of cracks, such as eccentricity and
area-based criteria. Next, we mapped the rectangular bounds of filtered CCR segments
derived in the binary pixel space to the raw image space (red-colored bounding boxes in
the image processing (Figure 3)). Subsequently, the bounding boxes from the raw image
were separated from the background and labeled ‘crack’ and ‘non-crack’ for supervised
learning. These segmented images of cracks and non-cracks were transformed to a Fourier
basis space using 2D DFT and vectorized to 1D to reduce the dimensional complexity. The
vectorized DFT was fed as an array input to the 1D DL model. Using the DL model, we
extracted discriminant features and classified CCRs as either cracks or non-cracks. The
segmented CCRs classified based on the 1D DL model were mapped to the original raw
image using the original bounding box coordinates, which were carried forward from
image preprocessing to classification CCRs in the original raw image. This approach is
sufficient irrespective of the size of the segment. The corresponding classified CCRs were
mapped onto the raw images. Although the implementation scheme is similar to the
bounding-box-based approach, the outcome is equivalent to semantic segmentation, where
we perform pixel-level classification of structural features in a computationally efficient
manner by combining image processing with a 1D DFT-CNN. The nonuniform size of
CCRs enables scalability, irrespective of the size of the image.
Thin-propagating shallow cracks, which often have widths of approximately a single
pixel, are difficult to identify because the distinct features between the crack and surround-
ing background are negligible. In the scheme for identifying thin shallow propagating
cracks, we first identified the endpoints of CCRs that were classified as cracks by the DL
model. Although a concrete structure is homogenous, the propagation of cracks at the
microscale is governed by the local intrinsic microstructure and extrinsic environmental
effects. Hence, we identified all possible endpoints of the pixelated residue (Figure 4a).
Next, we constructed a square sliding window spatially in the pixelated binarized domain,
with each endpoint as the centroid (Figure 4b). Then, we mapped the corresponding
coordinates of the square sliding window from the pixelated binarized domain to the
raw image (Figure 4c). We subsequently performed adaptive-threshold-based integral
image binarization with an increasing sensitivity factor t + ∆t inside the mapped window
(Figure 4d,e); a detailed discussion of adaptive-threshold-based integral image binarization
is provided in Section 2.1. We filtered out the noise in the sliding window using the geo-
metric properties of the crack (Figure 4f). The thin segment of the shallow crack extracted
from the sliding window was connected to the pixelated crack residue obtained from the
DL model (Figure 4g). We repeated the procedure using new endpoints of the tracked crack
segment. Crack segments are typically wide and deep at the center and gradually become
shallow toward their ends. Hence, after each iteration, we increased the threshold used for
image processing.
Figure 4 demonstrates the sequence of steps for the 1D DL-assisted image processing.
The algorithm sequence was automated without manual intervention. Manual predefined
inputs to the algorithm included (i) the size of the sliding window, (ii) incremental sensitiv-
ity factor (∆t) for image processing, and (iii) the eccentricity and area threshold to filter the
noisy background in the binarized image and texture-related features.

4. Results and Discussion of Proposed Scheme on Bridge Images

In this study, we employed a UAV to inspect bridge piers. The UAV prototype is a
highly stable, cost-effective, and off-the-shelf quadcopter that can be dynamically controlled
using a smartphone-based interface. The UAV has a built-in stabilizer with an inertial unit
system and a gyroscope. The prototype is equipped with a sensing and communication
module, a camera that records HD images, an inertial measurement system for acceler-
ation, a distance measurement module, and a Wi-Fi module that provides a connection
 Sensors 2023, 23, x FOR PEER REVIEW 10 of 21

inertial unit system and a gyroscope. The prototype is equipped with a sensing and com-
Sensors 2023, 23, 1419 munication module, a camera that records HD images, an inertial measurement 10 system
of 21
for acceleration, a distance measurement module, and a Wi-Fi module that provides a
connection between the UAV-based system and the remote system of the operator. Re-
mote access
between to a UAV-based
the UAV-based system system
and theallows
remotedynamic
system control,
of the updated
operator.position/distance
Remote access
information, and instantaneous data/image transfer.
to a UAV-based system allows dynamic control, updated position/distance The cameras were mounted
information,on a
gimbal
and platform that
instantaneous enabledtransfer.
data/image the operator to instantaneously
The cameras were mounted adjust and monitor
on a gimbal platform the
video
that and images
enabled acquired
the operator to by the camera. The
instantaneously camera
adjust andwas equipped
monitor with aand
the video 1/2”images
CMOS
image sensor
acquired by thetocamera.
obtain an TheRGB pixel was
camera image of 3648 with
equipped px × 5472
a 1/2”px,CMOS
with the maximum
image sensor op-to
eratinganFOV
obtain RGBofpixel
15°. Because
image ofof3648safetypxconcerns
× 5472 px, regarding
with thethe wind-induced
maximum impact
operating FOVof the
of
15 ◦ . Because
UAV on the target bridge
of safety pier, we
concerns operatedthe
regarding thewind-induced
UAV at a longimpact standoff ofdistance,
the UAVtypically
on the
in thebridge
target range pier,
of 3–4
wem from the
operated thetarget
UAVstructure; hence, the
at a long standoff resolution
distance, per pixel
typically in thewas
rangeap-
ofproximately
3–4 m from0.8–1 the mm/px. Often, thehence,
target structure; width theof a resolution
thin propagating
per pixelcrack is less
was than a pixel,
approximately
and the
0.8–1 crack characteristics
mm/px. Often, the width areof indistinguishable
a thin propagating compared
crack iswithlessthe background,
than a pixel, and which
the
crack
addscharacteristics
to the complexity are indistinguishable
of the problem, where compared
bridge with the background,
maintenance whichthe
necessitates adds to
detec-
the
tioncomplexity of the of
of crack widths problem, wherethan
sizes greater bridge maintenance
or equal to 0.3 mm. necessitates
However,the detection
a distinct of
crack
crack widths
signature is of sizes greater
recorded thanprovided
at a pixel or equal to 0.3the
that mm. However,
crack signature a distinct crack
is greater signature
than half the
iswidth
recorded of theat pixel
a pixelresolution.
provided A that the crack signature
representative image of is greater than halfprototype
the UAV-based the widthem- of
the pixelfor
ployed resolution.
bridge pier A representative image ofin
inspection is presented the UAV-based
Figure 5. Sample prototype
images employed
obtained from for
bridge
the UAV pierfor
inspection is presented
the inspection of theinbridge
Figurepier
5. Sample
are shownimages obtained
in Figure A1.from
Thethe UAV for
database in-
the inspection
cludes of thefeatures
crack-like bridge pier
and are shown in Figure
complicated A1. Thesuch
disturbances, database includes crack-like
as structural edges and
features
markings, andconcrete
complicated disturbances,
peeling, surface stains, such as spider
and structural edges
stains. Theandchosenmarkings,
databaseconcrete
of con-
peeling, surface stains, and spider stains. The chosen database of
crete images of bridge piers accounts for the effects of optical variability in terms ofconcrete images of bridge
light-
piers accounts forstandoff
ing conditions, the effects of optical
distance variability
between in terms
the camera andoftarget
lighting conditions,
structure, focalstandoff
lengths,
distance
FOV, and between
lens. Inthe camera we
addition, andconsidered
target structure,
concretefocal lengths,
surface FOV,variability
texture and lens. In byaddition,
populat-
weingconsidered
a databaseconcrete surface
with images from texture variability
different bridge by populating
piers. The proposed a database
modelwithwasimages
trained
from different bridge piers. The proposed model was trained and
and tested using this broad database, thereby accounting for environmental uncertainty tested using this broad
database, thereby accounting
and noise-induced effects. for environmental uncertainty and noise-induced effects.

Figure5.5.Representative
Figure Representativeimage
imageshowing
showingunmanned
unmannedaerial
aerialvehicle
vehicle(UAV)-based
(UAV)-basedprototype
prototypeused
usedfor
for
bridge pier inspection.
bridge pier inspection.

With the proposed scheme, we tracked the missing thin shallow cracks on bridge piers
from low-resolution UAV images, which are often undetected by DL models. Figure 6
displays a representative low-resolution image where the width of the crack is a single
Sensors 2023, 23, x FOR PEER REVIEW 11 of 21

Sensors 2023, 23, 1419 With the proposed scheme, we tracked the missing thin shallow cracks on bridge 11 of 21
piers from low-resolution UAV images, which are often undetected by DL models. Figure
6 displays a representative low-resolution image where the width of the crack is a single
pixel. The
pixel. The iterative
iterativescheme
schemeshowsshowspromising
promisingresults
resultsfor
fortracking
trackingmissing
missing thin segments
thin segments of
cracks (Figure 7). This approach is superior to existing semantic segmentation
of cracks (Figure 7). This approach is superior to existing semantic segmentation for the for the fol-
lowing reasons:
following reasons: (a) (a)
although
althoughsemantic segmentation
semantic segmentation predicts the the
predicts location of aof
location crack, the
a crack,
width of the crack region predicted by semantic segmentation is always
the width of the crack region predicted by semantic segmentation is always greater than greater than the
crack
the width,
crack and and
width, hence, the exact
hence, widthwidth
the exact of theofcrack
the in pixels
crack cannotcannot
in pixels be quantified; (b) the
be quantified;
proposed
(b) schemescheme
the proposed does not
doesrequire ground-truth
not require labeling
ground-truth of thin
labeling cracks,
of thin whereas
cracks, whereasthe the
se-
semantic segmentation requires intensive labeling; (c) a deeper DL architectural layerof-
mantic segmentation requires intensive labeling; (c) a deeper DL architectural layer is is
ten required
often required forfor
metadata
metadata feature extraction
feature extractionwhenwhenthethe
crack signature
crack is confined
signature to a to
is confined sin-a
gle pixel;
single however,
pixel; however, such
sucha ascheme
schemeisisunsuitable
unsuitablefor for real-time
real-time bridge
bridge inspections using
UAVs with
UAVs with limited
limited computational
computational facilities.
facilities.

(a) (b)

(c)
Figure 6. (a) Representative raw image and (b) zoomed region showing single-pixel width crack. (c)
Figure 6. (a) Representative raw image and (b) zoomed region showing single-pixel width
Loss of crack information during noise filtering post adaptive-threshold-based integral image bina-
crack. (c) Loss of crack information during noise filtering post adaptive-threshold-based integral
rization.
image binarization.

Despite the promising results, it is noteworthy to discuss the effects of various parame-
ters on integral image binarization contributing to the final results, including (a) the effect of
the initial threshold percentage (Figure 8) (b) the effect of the initial window size on tracking
thin shallow cracks (Figure 9). Moreover, one limitation is that when the crack segment is
close to the non-crack segment, the sliding window from the crack segment also searches
for thin cracks or shallow features from the non-crack regime, which is undesired and
requires additional filtering. This will be addressed in future studies. Another limitation
of the proposed scheme is the misrecognition of disturbances, in addition to the detection
of tiny cracks. Future work will focus on retaining the tiny cracks and filtering additional
disturbances by combining the outputs of the DL model and the proposed research.
Sensors 2023,
Sensors 23,23,
2023, x FOR
1419PEER REVIEW 12 of 21 12 of 21

(a) (b)
× ×

(c) (d)
×
×

(e) (f)
Figure 7. Results for identification of thin shallow cracks of single-pixel width. (a) Crack feature
Figure 7. Results for identification of thin shallow cracks of single-pixel width. (a) Crack feature
identification using 1D DL model. (b–f) Iteratively tracking missing segments of thin cracks (see (b))
postidentification using 1D DL
DL using a differential model.
sliding (b–f) with
window Iteratively tracking
increasing missing segments of thin cracks (see (b))
threshold.
post DL using a differential sliding window with increasing threshold.
Despite the promising results, it is noteworthy to discuss the effects of various pa-
rameters on integral image binarization contributing to the final results, including (a) the
effect of the initial threshold percentage (Figure 8) (b) the effect of the initial window size
 segment is close to the non-crack segment, the sliding window from the crack segment
also searches for thin cracks or shallow features from the non-crack regime, which is un-
desired and requires additional filtering. This will be addressed in future studies. Another
limitation of the proposed scheme is the misrecognition of disturbances, in addition to the
detection of tiny cracks. Future work will focus on retaining the tiny cracks and filtering
Sensors 2023, 23, 1419 13 of 21
additional disturbances by combining the outputs of the DL model and the proposed re-
search.

(a) (b)

(c) (d)

(e) (f)
Figure 8. Effect of initial threshold percentage on tracking thin shallow cracks (a) 𝑡 = 57 %, (c) 𝑡 =
Figure 8. Effect of initial threshold percentage on tracking thin shallow cracks (a) t = 57 %,
58 %, (e) 𝑡 = 59 %. (b),(d),(f) Tracking missing thin crack segments after 5th iteration for corre-
= 58 %, pixelated
(c) tsponding (e) t = 59residue
%. (b,d,f)
with Tracking missing thin
an initial threshold crack parameters
sensitivity segments after 5th iteration
as shown for corre-
in (a),(c),(e).
sponding pixelated
Note: Δ𝑡 for eachresidue
iterationwith
is 0.5an
%.initial threshold sensitivity parameters as shown in (a,c,e). Note:
∆t for each iteration is 0.5 %.

Table 1. Computational time comparison between the proposed framework and existing DL models.

Method Training Time Testing/Inference Time

Standard 2D-CNN (multi-pixel crack) 2 h 45 min 16 s 38–59 s/image [67]
Semantic segmentation 18 h 350 s/image [68]
Approx 0.02 s/image (60 images/s) (excluding
image preprocessing and postprocessing)
Image processing, preprocessing, and post
Approx 0.1–0.2 s/image (5–10 images/s)
processing to 1D-DFT-CNN (proposed 11 min 55 s (inclusive of preprocessing and DL)
(including image preprocessing and DL testing)
framework)
Approx 120–180 s/image (including image
preprocessing, DL testing, and iterative image
postprocessing)
 Sensors 2023, 23, x FOR PEER REVIEW 14 of 21
Sensors 2023, 23, 1419 14 of 21

×
×

(a) (b)

× ×
× ×

(c) (d)
Figure 9. (a) Crack feature identification using 1D DL model with an initial sensitivity
Figure 9. (a) Crack feature identification using 1D DL model with an initial sensitivity factor t = 58 %
factor 𝑡 = 58 % during image preprocessing. Effect of initial window size (b) 3 × 3,
during image preprocessing. Effect of initial window size (b) 23 × 23, (c) 28 × 28, and (d) 35 × 35
(c) 8 × 8, and (d) 35 × 35 on tracking thin shallow cracks. The software and compu-
on tracking thin shallow cracks. The software and computational details of the desktop system
tational details of the desktop system employed for the crack detection scheme are as fol-
employed for the crack detection scheme are as follows: Windows 10 operating system with Intel
lows: Windows 10 operating system with Intel (R) Core (TM) i9-10900F CPU @2.80 GHz
(R) Core (TM) i9-10900F CPU @2.80 GHz and 16 GB RAM memory. Various hidden processing steps
and 16 GB RAM memory. Various hidden processing steps to track missing shallow prop-
to track missing shallow propagating thin crack segments in post-1D CNN-based DL are discussed
agating thin crack segments in post-1D CNN-based DL are discussed in Section 3 (Figure
in Section 3 (Figure 5). The computational time for the first iteration with the proposed algorithm
5). The computational time for the first iteration with the proposed algorithm (Figure 5)
(Figure 5) for an image as large as 3648 px × 5472 px is approximately 2–3 min. The computation
for an image as large as 3648 px × 5472 px is approximately 2–3 min. The computation
time increases nonlinearly after each iteration. Wider segments of the crack are detected in the initial
time increases nonlinearly after each iteration. Wider segments of the crack are detected
iterations, while thin shallow segments of the crack are detected in the subsequent iterations, which
in the initial iterations, while thin shallow segments of the crack are detected in the sub-
require a higher threshold that leads to additional features or disturbances that need to be filtered;
sequent iterations, which require a higher threshold that leads to additional features or
hence, they are computationally expensive. A computational time comparison between the proposed
disturbances that need to be filtered; hence, they are computationally expensive. A com-
framework and existing state-of-the-art DL models is presented in Table 1.
putational time comparison between the proposed framework and existing state-of-the-
art DL
Onemodels
of the is presentedofinimage
limitations Tableprocessing
1. is the selection of predefined parameters. It
is impossible to find a universal choice of parametric values that are versatile and suitable
in all instances. Although the authors have used image processing with DL in our current
work, we implemented the approach with a conservative choice of parameters, meaning
that for image processing as a preprocessor, we chose a minimum threshold value that
caters to a wide range of databases. We compensated for this minimal thresholding by
incrementing the thresholding parameters in each iteration during image postprocessing
on crack-classified segments from the DL.
Sensors 2023, 23, 1419 15 of 21

Figure 10 shows the results for crack and non-crack feature classification, and the
subsequent detection of missing thin crack segments using the proposed scheme as dis-
cussed before. The proposed scheme of image-processing-assisted 1D-DFT-DL can detect
cracks even when they are within the bounding box of non-crack features (Figure 10a,d).
Pixelated results of crack and non-crack features are shown in Figure 10b,e. We further
detected missing thin shallow propagating cracks of single-pixel width using 1D-DFT-
DL-assisted image processing (Figure 10c,f)). The proposed approach could distinguish
crack features from non-crack features including structural edges, spider stains, peeling
of concrete surface coating, and structural markings. The pixel resolution of Figure 10a,b
are 0.43 and 0.21 mm/px, respectively. We applied the proposed method on various pixel
resolution images (mm/px), i.e., various widths of cracks of single-pixel size and the re-
sults are promising (Figure 10c,f). However, the proposed approach cannot distinguish
high-stress-zone structural cracks from cracks appearing on the coating surface, which
requires additional investigation. Few non-crack features whose major and minor axis are
comparable are classified as background by the precursor image processing and are thus
undetected during the proposed procedure.
For the chosen database with a wide range of optical variabilities, environmental
uncertainties, and surface texture variability, we quantified the quality of each image
using a baseline-free indicator—natural image quality evaluation (NIQE) [69]. NIQE
predicts deviations/artifacts in image statistics arising from optical and surface variability.
Figure 11a presents the optical image quality quantified using the NIQE histogram. The
quantitative range of the NIQE is 3–8; a low NIQE metric implies the best perceptual
quality, with minimal undistortion. Ideally, we anticipate a reasonably high accuracy
within the database uncertainty range or variability and a decrease for any unaccounted
effects. However, the performance metrics deviate from the anticipated trend beyond the
NIQE metric > 6 (Figure 11b) because of the skewed unbalanced dataset. The performance
metrics of the proposed scheme are consistent for the NIQE range ≤ 6.
The proposed scheme is generic, can be adopted in a plug-and-play scheme, and
can be employed regardless of the DL model used. A higher sensitivity/thresholding for
precursor image processing increases the computational cost of DL. Higher sensitivity
increases CCRs, which require additional detailed labeling prior to DL. Often, labeling
with finer CCRs is misinterpreted and subjective to human interpretation. Moreover, for
these types of problems, researchers have employed a deeper layer of architecture to extract
hidden distinct features within a single-pixel width, thereby increasing the computational
cost nonlinearly. However, the proposed scheme is computationally effective, wherein
a cushion is provided to perform DL with a shallow architecture and extract missing
information in an iterative manner post DL. Generally, the choice of the initial parameters
in image processing determines the extent to which shallow cracks are determined. In
addition, there can be a tradeoff between the choice of these parameters, which affects the
computational cost and the sensitivity of crack detection using DL models; hence, the DL
approach is sensitive to initial parameters such as the sensitivity factor in image processing.
Accordingly, the proposed multi-threshold-based image processing, one as a precursor and
the other as a postprocessor to the DL with a differential sliding window, is less sensitive to
the initial handcrafted selections and predesigned parameters.
Sensors
Sensors 23,23,
2023,
2023, 1419
x FOR PEER REVIEW 16 of16
21of 21

(a) (b)

(c) (d)

(e) (f)
Figure 10. (a,d) Classification of crack in presence of non-crack features including structural edge,
Figure 10. (a,d) Classification of crack in presence of non-crack features including structural
spider stains, peeling of concrete surface coating, stains and discoloration, and markings. (b,e) Rep-
edge, spiderpixelated
resentative stains, peeling ofcrack
results for concrete surface coating,
and non-crack featuresstains
shownand discoloration,
in (a,d), respectively.and
(c,f)markings.
Miss-
(b,e)
ing Representative pixelated
thin shallow cracks results for
of single-pixel crack
width (seeand non-crack
(b,e)) detected features (a,d), pro-
shown inimage
using DL-assisted respec-
tively. (c,f) Missing thin shallow cracks of single-pixel width (see (b,e)) detected using DL-assisted
cessing.
image processing.
 11a presents the optical image quality quantified using the NIQE histogram. The quanti-
tative range of the NIQE is 3–8; a low NIQE metric implies the best perceptual quality,
with minimal undistortion. Ideally, we anticipate a reasonably high accuracy within the
database uncertainty range or variability and a decrease for any unaccounted effects.
However, the performance metrics deviate from the anticipated trend beyond the NIQE
Sensors 2023, 23, 1419 17 of 21
metric > 6 (Figure 11b) because of the skewed unbalanced dataset. The performance met-
rics of the proposed scheme are consistent for the NIQE range ≤ 6.

(a) (b)
Figure
Figure11.11.(a)
(a)Optical
Opticalimage
imagequality
qualitythat
thataccounts
accounts for
for environmental
environmental uncertainty
uncertainty and optical variabil-
and optical varia-
bility
ity of the database, quantified using natural image quality evaluation (NIQE) histogram. (b) Cor-
of the database, quantified using natural image quality evaluation (NIQE) histogram. (b) Corre-
responding performance metrics scores for feature classification.
sponding performance metrics scores for feature classification.

The proposed scheme is generic, can be adopted in a plug-and-play scheme, and can
5. Summary
be employed regardless
In the proposed of the DLwe
approach, model
first used.
appliedA higher
image sensitivity/thresholding
binarization and DL-based for pre-
crack
cursor image processing increases the computational cost of
identification. Subsequently, we detected thin propagating cracks that are often missedDL. Higher sensitivity in-in
creases CCRs, which require additional detailed labeling prior to
DL-based classification by applying a differential sliding window of image processing to DL. Often, labeling with
finer
the endCCRs is misinterpreted
segments and subjective
of cracks classified from the to DL
human
model. interpretation.
In general, aMoreover,
high initialfor these
threshold
types
is used of during
problems, researchers
image have as
binarization employed
a precursor a deeper layer of architecture
for identifying CCRs, resultingto extract
in an
hidden distinct features within a single-pixel width, thereby increasing
excessive amount of surface background and statistical features that are not associated the computational
cost
withnonlinearly.
concrete surface However, the proposed
features. The proposed scheme is computationally
approach effective,
provides a cushion or wherein
tradeoff afor
cushion
reducing the sensitivity parameters in the precursor, thereby reducing the potentialinfor-
is provided to perform DL with a shallow architecture and extract missing CCRs
mation
and theinassociated
an iterative manner post DL.
computational costsGenerally,
during thethe choice and
training of the initialofparameters
testing the DL model. in
image processing determines the extent to which shallow cracks are
The loss of sensitivity in the initial stage is compensated for during the image-binarization- determined. In addi-
tion,
based there can be a tradeoff
postprocessing. Hence,between the choice of these
multi-threshold-based image parameters,
processingwhich affects the
as a precursor and
computational
postprocessor cost and the
provides sensitivityinofthe
invariance crack detection to
framework using
the DL models;
choice hence,
of initial the DL
sensitivity
approach
parameters is sensitive
during image to initial parameters
processing, such as the sensitivity
hyperparameters, factor in
and the number ofimage
layers pro-
in the
cessing. Accordingly, the proposed multi-threshold-based
DL model. Meanwhile, semantic segmentation can result in a masked segmented image processing, one as aregion
pre-
cursor and the other as a postprocessor to the DL with a differential
of a crack whose mask width is greater than that of a single-pixel propagating crack. The sliding window, is
less sensitive to the initial handcrafted selections and predesigned
proposed approach enables image scalability with an efficient computational scheme while parameters.
maintaining high accuracy in detecting thin shallow propagating crack segments.
5. Summary
The tracking of missing crack segments using differential sliding-window-based local
image In processing
the proposed is equivalent
approach, to weconnecting
first applied crack segments
image based on
binarization andtheDL-based
slope continuity
crack
information ofSubsequently,
identification. the crack. However, in certain
we detected thininstances,
propagating the crack
cracksslope
that continuity fails when
are often missed in
the direction
DL-based of crack propagation
classification by applyingisa governed
differential bysliding
the irregular
window orientation of the material
of image processing to
microstructure
the end segments and
of grains. Althoughfrom
cracks classified the proposed scheme
the DL model. Inexhibited
general, apromising
high initial results,
thresh- the
overall
old is used approach
during is limited
image to the detection
binarization of thin crack
as a precursor segmentsCCRs,
for identifying or propagating
resulting in cracks
an
on the surface.
excessive amount Alternatively, physics-embedded-based
of surface background NDI approaches
and statistical features that are not combined
associated with
computer-vision-based sensing are required to detect the propagating cracks hidden within
the structure. The proposed computer-vision- and DL-based approach inherently uses the
geometrical information of cracks in concrete structures, wherein a sharp crack signature is
followed by the tail of a thin shallow crack region or a propagating crack. This has potential
applications in identifying real-time thin propagating cracks catering to a wide range of
civil structures, wherein we employ UAVs or robotic setups with limited computational
facilities to obtain low-resolution images owing to operating distance constraints.

6. Conclusions
The present study adopted multi-threshold image processing combined with a 1D
frequency-domain DL model to classify cracks, including thin propagating shallow cracks
with a single-pixel width. This approach effectively combines the advantages of (i) image
Sensors 2023, 23, 1419 18 of 21

binarization, (ii) computational effectiveness of a shallow 1D DL model, and (iii) inherent

frequency-domain feature extraction characteristics of cracks in the Fourier-based DL
model (1D-DFT-CNN). Specifically, the proposed scheme overcomes the limitations of
(i) the fixed window size of sliding-window-based conventional CNNs and (ii) the time and
effort required in pixelated labeling as employed in existing semantic segmentation. We
demonstrated the proposed algorithm using low-resolution images obtained from different
bridge piers using UAVs under different ambient lighting conditions and crack-like features.

Author Contributions: Conceptualization, G.K.G. and S.-H.S.; methodology, G.K.G.; software,

G.K.G.; validation, G.K.G.; formal analysis, G.K.G.; investigation, S.-H.S.; resources, S.-H.S.; data
curation, G.K.G.; writing—original draft preparation, G.K.G.; writing—review and editing, S.-H.S.;
visualization, G.K.G.; supervision, S.-H.S.; project administration, S.-H.S. and H.-J.Y.; funding acquisi-
tion, S.-H.S. and H.-J.Y. All authors have read and agreed to the published version of the manuscript.
Funding: This research was financially supported by the Ministry of Trade, Industry and Energy
(MOTIE) and the Korea Institute for Advancement of Technology (KIAT) through the National
Innovation Cluster R & D program (P0002073_Open LAB Connected support project in future traffic
safety to promote regional innovative enterprises).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Sensors 2023, 23, x FOR PEER REVIEW
Data Availability Statement: Data available on request due to restrictions from19
theoffunding
21 agencies.
Conflicts of Interest: The authors declare no conflict of interest.
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A. Sample Images Obtained from Unmanned Aerial Vehicle (UAV) for the
Inspection
Appendix A.of a Bridge
Sample Pier
Images Obtained from Unmanned Aerial Vehicle (UAV) for the
Inspection of a Bridge Pier

Figure A1. Representative images obtained from UAV for inspection of the bridge pier.
Figure A1. Representative images obtained from UAV for inspection of the bridge pier.
References
1. Brownjohn, J.M.W. Structural health monitoring of civil infrastructure. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 2007, 365,
589–622.
2. Cross, E.J.; Worden, K.; Farrar, C.R. Structural health monitoring for civil infrastructure. In Health Assessment of Engineered
Structures: Bridges, Buildings and Other Infrastructures; World Scientific: Singapore, 2013; pp. 1–31.
3. Sharma, S.; Dangi, S.K.; Bairwa, S.K.; Sen, S. Comparative study on sensitivity of acceleration and strain responses for bridge
health monitoring. J. Struct. Integr. Maint. 2022, 7, 238–251.
Sensors 2023, 23, 1419 19 of 21

References
1. Brownjohn, J.M.W. Structural health monitoring of civil infrastructure. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 2007, 365,
589–622. [CrossRef] [PubMed]
2. Cross, E.J.; Worden, K.; Farrar, C.R. Structural health monitoring for civil infrastructure. In Health Assessment of Engineered
Structures: Bridges, Buildings and Other Infrastructures; World Scientific: Singapore, 2013; pp. 1–31.
3. Sharma, S.; Dangi, S.K.; Bairwa, S.K.; Sen, S. Comparative study on sensitivity of acceleration and strain responses for bridge
health monitoring. J. Struct. Integr. Maint. 2022, 7, 238–251. [CrossRef]
4. Campbell, L.E.; Connor, R.J.; Whitehead, J.M.; Washer, G.A. Human factors affecting visual inspection of fatigue cracking in steel
bridges. Struct. Infrastruct. Eng. 2021, 17, 1447–1458. [CrossRef]
5. Dong, C.-Z.; Catbas, F.N. A review of computer vision-based structural health monitoring at local and global levels. Struct. Health
Monit. 2021, 20, 692–743. [CrossRef]
6. Spencer, B.F., Jr.; Hoskere, V.; Narazaki, Y. Advances in computer vision-based civil infrastructure inspection and monitoring.
Engineering 2019, 5, 199–222. [CrossRef]
7. Koch, C.; Georgieva, K.; Kasireddy, V.; Akinci, B.; Fieguth, P. A review on computer vision based defect detection and condition
assessment of concrete and asphalt civil infrastructure. Adv. Eng. Inform. 2015, 29, 196–210. [CrossRef]
8. Bhattacharjee, S.; Deb, D. Automatic detection and classification of damage zone (s) for incorporating in digital image correlation
technique. Opt. Lasers Eng. 2016, 82, 14–21. [CrossRef]
9. Gehri, N.; Mata-Falcón, J.; Kaufmann, W. Automated crack detection and measurement based on digital image correlation. Constr.
Build. Mater. 2020, 256, 119383. [CrossRef]
10. Dumoulin, C.; Deraemaeker, A. Real-time fast ultrasonic monitoring of concrete cracking using embedded piezoelectric transduc-
ers. Smart Mater. Struct. 2017, 26, 104006. [CrossRef]
11. Kolappan Geetha, G.; Roy Mahapatra, D.; Gopalakrishnan, S.; Hanagud, S. Laser Doppler imaging of delamination in a composite
T-joint with remotely located ultrasonic actuators. Compos. Struct. 2016, 147, 197–210. [CrossRef]
12. Park, S.; Ahmad, S.; Yun, C.-B.; Roh, Y. Multiple crack detection of concrete structures using impedance-based structural health
monitoring techniques. Exp. Mech. 2006, 46, 609–618. [CrossRef]
13. Climent, M.Á.; Miró, M.; Carbajo, J.; Poveda, P.; de Vera, G.; Ramis, J. Use of non-linear ultrasonic techniques to detect cracks due
to steel corrosion in reinforced concrete structures. Materials 2019, 12, 813. [CrossRef]
14. Loeffler, C.M.; Qiu, Y.; Martin, B.; Heard, W.; Williams, B.; Nie, X. Detection and segmentation of mechanical damage in concrete
with X-ray microtomography. Mater. Charact. 2018, 142, 515–522. [CrossRef]
15. Kolappan Geetha, G.; Munian, R.K.; Roy Mahapatra, D.; In, C.-W.; Raulerson, D.A. Ultrasonic horn contact-induced transient
anharmonic resonance effect on vibro-thermography. J. Sound Vib. 2022, 525, 116786. [CrossRef]
16. Kolappan Geetha, G.; Roy Mahapatra, D.; In, C.-W.; Raulerson, D. Transient vibro-thermography and nonlinear resonant modes.
J. Vib. Acoust. 2020, 142, 061007. [CrossRef]
17. Jang, K.; Jung, H.; An, Y.-K. Automated bridge crack evaluation through deep super resolution network-based hybrid image
matching. Autom. Constr. 2022, 137, 104229. [CrossRef]
18. Kolappan Geetha, G.; Mahapatra, D.R. Modeling and simulation of vibro-thermography including nonlinear contact dynamics of
ultrasonic actuator. Ultrasonics 2019, 93, 81–92. [CrossRef] [PubMed]
19. Jiang, S.; Zhang, J. Real-time crack assessment using deep neural networks with wall-climbing unmanned aerial system. Comput.
Civ. Infrastruct. Eng. 2020, 35, 549–564. [CrossRef]
20. Xie, R.; Yao, J.; Liu, K.; Lu, X.; Liu, Y.; Xia, M.; Zeng, Q. Automatic multi-image stitching for concrete bridge inspection by
combining point and line features. Autom. Constr. 2018, 90, 265–280. [CrossRef]
21. Lim, R.S.; La, H.M.; Sheng, W. A robotic crack inspection and mapping system for bridge deck maintenance. IEEE Trans. Autom.
Sci. Eng. 2014, 11, 367–378. [CrossRef]
22. Jang, K.; An, Y.-K.; Kim, B.; Cho, S. Automated crack evaluation of a high-rise bridge pier using a ring-type climbing robot.
Comput. Civ. Infrastruct. Eng. 2021, 36, 14–29. [CrossRef]
23. Yang, L.; Li, B.; Feng, J.; Yang, G.; Chang, Y.; Jiang, B.; Xiao, J. Automated wall-climbing robot for concrete construction inspection.
J. Field Robot. 2022, 40, 110–129. [CrossRef]
24. Kim, H.; Lee, J.; Ahn, E.; Cho, S.; Shin, M.; Sim, S.-H. Concrete crack identification using a UAV incorporating hybrid image
processing. Sensors 2017, 17, 2052. [CrossRef] [PubMed]
25. Li, Y.; Liu, C. Applications of multirotor drone technologies in construction management. Int. J. Constr. Manag. 2019, 19, 401–412.
[CrossRef]
26. Yang, J.; Li, H.; Zou, J.; Jiang, S.; Li, R.; Liu, X. Concrete crack segmentation based on UAV-enabled edge computing. Neurocomput-
ing 2022, 485, 233–241. [CrossRef]
27. Oh, J.-K.; Jang, G.; Oh, S.; Lee, J.H.; Yi, B.-J.; Moon, Y.S.; Lee, J.S.; Choi, Y. Bridge inspection robot system with machine vision.
Autom. Constr. 2009, 18, 929–941. [CrossRef]
28. Zawad, M.; Shahriar, R.; Zawad, M.; Shahriar, F.; Rahman, M.; Priyom, S.N. A comparative review of image processing based
crack detection techniques on civil engineering structures. J. Soft Comput. Civ. Eng. 2021, 5, 58–74.
29. Munawar, H.S.; Hammad, A.W.A.; Haddad, A.; Soares, C.A.P.; Waller, S.T. Image-based crack detection methods: A review.
Infrastructures 2021, 6, 115. [CrossRef]
Sensors 2023, 23, 1419 20 of 21

30. Mohan, A.; Poobal, S. Crack detection using image processing: A critical review and analysis. Alex. Eng. J. 2018, 57, 787–798.
[CrossRef]
31. Yang, X.; Guan, J.; Ding, L.; You, Z.; Lee, V.C.S.; Hasan, M.R.M.; Cheng, X. Research and applications of artificial neural network
in pavement engineering: A state-of-the-art review. J. Traffic Transp. Eng. Engl. Ed. 2021, 8, 1000–1021. [CrossRef]
32. Ekanayake, B.; Wong, J.K.-W.; Fini, A.A.F.; Smith, P. Computer vision-based interior construction progress monitoring: A literature
review and future research directions. Autom. Constr. 2021, 127, 103705. [CrossRef]
33. Ali, R.; Chuah, J.H.; Talip, M.S.A.; Mokhtar, N.; Shoaib, M.A. Structural crack detection using deep convolutional neural networks.
Autom. Constr. 2022, 133, 103989. [CrossRef]
34. Hamishebahar, Y.; Guan, H.; So, S.; Jo, J. A Comprehensive Review of Deep Learning-Based Crack Detection Approaches. Appl.
Sci. 2022, 12, 1374. [CrossRef]
35. Chun, P.; Izumi, S.; Yamane, T. Automatic detection method of cracks from concrete surface imagery using two-step light gradient
boosting machine. Comput. Civ. Infrastruct. Eng. 2021, 36, 61–72. [CrossRef]
36. Song, L.; Sun, H.; Liu, J.; Yu, Z.; Cui, C. Automatic segmentation and quantification of global cracks in concrete structures based
on deep learning. Measurement 2022, 199, 111550. [CrossRef]
37. Wang, W.; Su, C.; Fu, D. Automatic detection of defects in concrete structures based on deep learning. Structures 2022, 43, 192–199.
[CrossRef]
38. Bu, J.; Chen, X.; Hu, L.; Yang, H.; Liu, S. Experimental study on crack propagation of concrete under various loading rates with
digital image correlation method. Int. J. Concr. Struct. Mater. 2020, 14, 25. [CrossRef]
39. Cubero-Fernandez, A.; Rodriguez-Lozano, F.; Villatoro, R.; Olivares, J.; Palomares, J.M. Efficient pavement crack detection and
classification. EURASIP J. Image Video Process. 2017, 2017, 39. [CrossRef]
40. Koch, C.; Paal, S.G.; Rashidi, A.; Zhu, Z.; König, M.; Brilakis, I. Achievements and challenges in machine vision-based inspection
of large concrete structures. Adv. Struct. Eng. 2014, 17, 303–318. [CrossRef]
41. Talab, A.M.A.; Huang, Z.; Xi, F.; HaiMing, L. Detection crack in image using Otsu method and multiple filtering in image
processing techniques. Optik 2016, 127, 1030–1033. [CrossRef]
42. Oliveira, H.; Correia, P.L. CrackIT—An image processing toolbox for crack detection and characterization. In Proceedings of the
2014 IEEE International Conference on Image Processing (ICIP), Paris, France, 27–30 October 2014; pp. 798–802.
43. Hsieh, Y.-A.; Tsai, Y.J. Machine learning for crack detection: Review and model performance comparison. J. Comput. Civ. Eng.
2020, 34, 4020038. [CrossRef]
44. Shi, Y.; Cui, L.; Qi, Z.; Meng, F.; Chen, Z. Automatic road crack detection using random structured forests. IEEE Trans. Intell.
Transp. Syst. 2016, 17, 3434–3445. [CrossRef]
45. Zhang, A.; Wang, K.C.P.; Fei, Y.; Liu, Y.; Chen, C.; Yang, G.; Li, J.Q.; Yang, E.; Qiu, S. Automated pixel-level pavement crack
detection on 3D asphalt surfaces with a recurrent neural network. Comput. Civ. Infrastruct. Eng. 2019, 34, 213–229. [CrossRef]
46. Mir, B.A.; Sasaki, T.; Nakao, K.; Nagae, K.; Nakada, K.; Mitani, M.; Tsukada, T.; Osada, N.; Terabayashi, K.; Jindai, M. Machine
learning-based evaluation of the damage caused by cracks on concrete structures. Precis. Eng. 2022, 76, 314–327. [CrossRef]
47. Srivani, A.; Xavior, M.A. Investigation of surface texture using image processing techniques. Procedia Eng. 2014, 97, 1943–1947.
[CrossRef]
48. Geffray, S.; Klutchnikoff, N.; Vimond, M. Illumination problems in digital images. A statistical point of view. J. Multivar. Anal.
2016, 150, 191–213. [CrossRef]
49. Julliand, T.; Nozick, V.; Talbot, H. Image noise and digital image forensics. Lect. Notes Comput. Sci. 2016, 9569, 3–17.
50. Wu, P.; Liu, A.; Fu, J.; Ye, X.; Zhao, Y. Autonomous surface crack identification of concrete structures based on an improved
one-stage object detection algorithm. Eng. Struct. 2022, 272, 114962. [CrossRef]
51. Ye, W.; Deng, S.; Ren, J.; Xu, X.; Zhang, K.; Du, W. Deep learning-based fast detection of apparent concrete crack in slab tracks
with dilated convolution. Constr. Build. Mater. 2022, 329, 127157. [CrossRef]
52. Xiang, C.; Wang, W.; Deng, L.; Shi, P.; Kong, X. Crack detection algorithm for concrete structures based on super-resolution
reconstruction and segmentation network. Autom. Constr. 2022, 140, 104346. [CrossRef]
53. Gharehbaghi, V.R.; Kalbkhani, H.; Noroozinejad Farsangi, E.; Yang, T.Y.; Nguyen, A.; Mirjalili, S.; Málaga-Chuquitaype, C. A
novel approach for deterioration and damage identification in building structures based on Stockwell-Transform and deep
convolutional neural network. J. Struct. Integr. Maint. 2022, 7, 136–150. [CrossRef]
54. Kim, H.; Ahn, E.; Shin, M.; Sim, S.-H. Crack and noncrack classification from concrete surface images using machine learning.
Struct. Health Monit. 2019, 18, 725–738. [CrossRef]
55. Furuta, H.; Nomura, Y.; Inoue, M. Evaluation of Crack Propagation in Concrete Bridges from Vehicle-Mounted Camera Images
Using Deep Learning and Image Processing. Front. Built Environ. 2022, 8, 972796.
56. Golding, V.P.; Gharineiat, Z.; Munawar, H.S.; Ullah, F. Crack Detection in Concrete Structures Using Deep Learning. Sustainability
2022, 14, 8117. [CrossRef]
57. Yu, L.; He, S.; Liu, X.; Jiang, S.; Xiang, S. Intelligent Crack Detection and Quantification in the Concrete Bridge: A Deep
Learning-Assisted Image Processing Approach. Adv. Civ. Eng. 2022, 2022, 1813821. [CrossRef]
58. Kolappan Geetha, G.; Sim, S.-H. Fast identification of concrete cracks using 1D deep learning and explainable artificial intelligence-
based analysis. Autom. Constr. 2002, 143, 104572. [CrossRef]
59. Otsu, N. A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man. Cybern. 1979, 9, 62–66. [CrossRef]
Sensors 2023, 23, 1419 21 of 21

60. Fan, J.-L.; Lei, B. A modified valley-emphasis method for automatic thresholding. Pattern Recognit. Lett. 2012, 33, 703–708.
[CrossRef]
61. Ng, H.-F.; Jargalsaikhan, D.; Tsai, H.-C.; Lin, C.-Y. An improved method for image thresholding based on the valley-emphasis
method. In Proceedings of the IEEE Asia-Pacific Signal and Information Processing Association Annual Summit and Conference,
Kaohsiung, Taiwan, 29 October–1 November 2013; pp. 1–4.
62. Truong, M.T.N.; Kim, S. Automatic image thresholding using Otsu’s method and entropy weighting scheme for surface defect
detection. Soft Comput. 2018, 22, 4197–4203. [CrossRef]
63. Bradley, D.; Roth, G. Adaptive thresholding using the integral image. J. Graph. Tools 2007, 12, 13–21. [CrossRef]
64. Wellner, P.D. Adaptive thresholding for the Digital Desk. In Xerox, EPC1993-110; Rank Xerox Ltd.: Cambridge, UK, 1993; pp. 1–19.
65. Yamaguchi, T.; Hashimoto, S. Automated crack detection for concrete surface image using percolation model and edge information.
In Proceedings of the IECON 2006—32nd Annual Conference on IEEE Industrial Electronics, Paris, France, 7–10 November 2006;
pp. 3355–3360.
66. Abdel-Qader, I.; Abudayyeh, O.; Kelly, M.E. Analysis of edge-detection techniques for crack identification in bridges. J. Comput.
Civ. Eng. 2003, 17, 255–263. [CrossRef]
67. Zhang, Q.; Barri, K.; Babanajad, S.K.; Alavi, A.H. Real-Time Detection of Cracks on Concrete Bridge Decks Using Deep Learning
in the Frequency Domain. Engineering 2020, 7, 1786–1796. [CrossRef]
68. Zhou, S.; Song, W. Crack segmentation through deep convolutional neural networks and heterogeneous image fusion. Autom.
Constr. 2021, 125, 103605. [CrossRef]
69. Mittal, A.; Soundararajan, R.; Bovik, A.C. Making a “completely blind” image quality analyzer. IEEE Signal Process. Lett. 2012, 20,
209–212. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.