Hindawi

Advances in Civil Engineering

Volume 2020, Article ID 8829715, 16 pages
https://doi.org/10.1155/2020/8829715

Research Article
Image Processing-Based Spall Object Detection Using Gabor
Filter, Texture Analysis, and Adaptive Moment Estimation (Adam)
Optimized Logistic Regression Models

1,2
Nhat-Duc Hoang
1
Institute of Research and Development, Duy Tan University, Da Nang 550000, Vietnam
2
Faculty of Civil Engineering, Duy Tan University, Da Nang 550000, Vietnam

Correspondence should be addressed to Nhat-Duc Hoang; [email protected]

Received 16 May 2020; Revised 5 November 2020; Accepted 12 November 2020; Published 30 November 2020

Academic Editor: Moacir Kripka

Copyright © 2020 Nhat-Duc Hoang. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This study aims at proposing a computer vision model for automatic recognition of localized spall objects appearing on surfaces of
reinforced concrete elements. The new model is an integration of image processing techniques and machine learning approaches.
The Gabor filter supported by principal component analysis and k-means clustering is used for identifying the region of interest
within an image sample. The binary gradient contour, gray level co-occurrence matrix, and color channels’ statistical mea-
surements are employed to compute the texture of the extracted region of interest. Based on the computed texture-based features,
the logistic regression model trained by the state-of-the-art adaptive moment estimation (Adam) is utilized to establish a decision
boundary that delivers predictions on the status of “nonlocalized spall” and “localized spall.” Experimental results demonstrate
that the newly developed model is able to achieve good detection accuracy with classification accuracy rate � 85.32%, pre-
cision � 0.86, recall � 0.79, negative predictive value � 0.85, and F1 score � 0.82. Thus, the proposed computer vision model can be
helpful to assist decision makers in the task of the periodic survey of structure heath condition.

1. Introduction materials. Most often, spalling in concrete is due to cor-

rosion of steel reinforcement embedded in the structure. To
Public safety is a major concern of civil engineers who design prevent such accidents and to ensure the safety and ser-
and maintain high-rise buildings. Despite considerable ef- viceability of the built environment, periodic visual surveys
forts in design and advanced knowledge of building of structural heath condition and proper maintenance ac-
structures, accidents can still happen in the built environ- tivities are very crucial [5].
ment due to excessive usage, structural aging, and inclement In developing countries, including Vietnam, manual
weather conditions [1]. Among the hazards occurred in visual inspection is still the principal method for evaluating
high-rise buildings, falling objects from overhead caused by structural heath conditions. This activity is performed at
concrete spalling can be particularly dangerous and have a regular intervals to identify potential damages and guarantee
high potential severity to occupants’ heath [2]. The effect of the service/safety requirements of high-rise buildings.
concrete debris can be devastating for human lives if it gets Provided the well-trained technicians experienced in
broken off from surfaces of exterior wall systems of high-rise structural heath assessment, manual visual inspection is able
buildings [3]. to providing accurate surveying outcomes. Nevertheless, due
A concrete spall (Figure 1) is regarded as flakes of to the increasing numbers of buildings needed to be
concrete/mortar broken off from a concrete element (e.g., inspected periodically and the limited number of experience
beam, wall, and ceiling) [4]. Spalling is typically caused by technicians, timely evaluation of building elements becomes
stresses brought about by differential movement of infeasible and inspection deficiencies become a major
2 Advances in Civil Engineering

(a) (b) (c)

Figure 1: Spall objects on (a) ceiling, (b) beam, and (c) wall structures.

concern of property owners. Therefore, there is a practical including support vector machines, Naı̈ve Bayesian classi-
need to substitute the unproductive manual visual survey fier, and random forest have been employed to identify
with a more effective approach. concrete defects [8, 26]. A model for localization and
Recently, due to the ease of access to low-cost and high- quantification of concrete spalling defects based on terres-
quality visual sensing equipment including digital cameras, trial laser scanning has been proposed in [27]. Dawood et al.
computer vision-based models have been increasingly used [21] presented a computer vision-based model for spalling
for automatic structural heath monitoring [6]. These ad- detection used in environment of subway networks.
vanced approaches have been proved to be viable alterna- Hoang [28] relies on a steerable filter and machine
tives to replace the labor-intensive and subjective methods learning to recognize wall defects such as cracks and spalls. A
relied on manual survey. With the use of advanced image concrete spalling detection model for metro tunnel from
processing techniques operated on image samples collected point cloud that employs a roughness descriptor has been
from digital cameras, the physical condition of civil struc- developed by Wu et al. [24]. Hoang [29] presents an image
tures can be continuously surveyed and reported to main- processing approach for automatic detection of concrete
tenance agencies. This evaluation outcome can be effectively spalling using machine learning algorithms and image
used to support the decision-making process regarding texture analysis. Nevertheless, this model focused on ma-
maintenance prioritization and funding allocation. chine learning-based texture discrimination and was not
Due to such reasons, a large number of computer vision- capable of isolating the entire individual spall object.
based approaches have been proposed to successfully detect Yao et al. [30] establishes a convolutional neural net-
various forms of structural defects such as cracking and work-based model for concrete bughole detection; a large
spalling. Abdel-Qader et al. [7] employs a principle com- number (about 10,000) of image examples have been used as
ponent analysis-based model to recognize cracking defects a training dataset. Li et al. [31] proposed a model for
appeared on bridge surfaces; the principle component detecting exposed aggregate appeared on stilling basin slab
analysis is utilized to support data cluster identification with using the attention U-Net network. Chow et al. [32] employs
a large database of bridge images. O’Byrne et al. [8] utilizes deep learning of a convolutional autoencoder for anomaly
texture analysis for detecting damages appeared on infra- detection of defects existing on concrete structures. A model
structural elements; the texture-based image segmentation for recognizing damaged ceiling areas in large-span struc-
relies on pixel intensity values and gray level co-occurrence tures has been proposed by Wang et al. [33]; this model
matrix. Subsequently, a support vector machines model is employs a convolutional neural network for pattern rec-
then employed for the data classification task. Lattanzi and ognition. Although deep learning-based models are capable
Miller [9] rely on the data clustering approach for image of performing feature extraction phase automatically, these
segmentation based on the Canny and k-means algorithm; supervised learning models generally demand a large-size
the research finds that the combined algorithms can deliver training dataset and a meticulous process of data labeling
good accuracy of crack recognition under different envi- [34–36]. This data labeling process itself can be time-con-
ronmental circumstances. suming as well as error prone [5]. In addition, the deep
As can be seen from the literature, a large number of learning models also require experience and the trial-and-
previous studies have been dedicated to crack detection for error process to adjust a significant amount of model’s free
concrete structures [10–20]. Only recently, there is an in- parameters.
creasing focus on detecting other forms of damage including An effort of combining unsupervised learning and
concrete spalling [21–24]. German et al. [25] constructs a machine learning-based data classification has been recently
combination of segmentation, template matching, and introduced in [37]. The k-means clustering algorithm and
morphological preprocessing for detecting spall appeared on machine learning classifier have been integrated to form an
surfaces of concrete columns. Machine learning models automatic approach for estimating stripping of asphalt
Advances in Civil Engineering 3

coating. The k-means clustering algorithm is utilized to ⎨ 1 x2 y 2 ⎫

⎧ ⎬
separate pixels with similar values on the surface of ag- h(x, y) � exp⎩ − ⎡⎢⎣ 2 + 2 ⎤⎥⎦⎭ cos 2πu0 x􏼁, (1)
2 σx σy
gregates; subsequently, machine learning models are used to
categorize the identified clusters into groups of asphalt-
where u0 denotes the frequency of a sinusoidal plane wave
coated and uncoated areas.
along the x axis. σ x and σ y represent the space constants of
As pointed out by previous studies, the current chal-
the Gaussian envelope along the x and y axes, respectively. It
lenges faced by computer vision-based concrete damage
is noted that the GF with arbitrary orientations can be
detection including spall recognition are complex envi-
obtained via a rigid rotation of the x-y coordinate system.
ronmental conditions (e.g., noisy background image) [5]
The frequency domain representation of the GF is given
and the difficulty of the image labeling process [32]. More
by [45]
efforts should be dedicated to automatically identify the
2
damage’s region of interest (ROI) via unsupervised learning 1 u − u0 􏼁 v2
methods. Capable machine learning methods with few free H(u, v) � A􏼠exp􏼨− 􏼢 􏼣 + 􏼩
2 σ 2u σ 2v
parameters should be investigated as viable alternatives to (2)
sophisticated models used for data classification. It is be- 2
1 u + u0 􏼁 v2
cause simple and manageable models significantly facilitate + exp􏼨− 􏼢 􏼣 + 􏼩􏼡,
the development and application of hybrid computer vision- 2 σ 2u σ 2v
machine learning approaches for concrete spalling
where σ u � 1/2πσ x , σ v � 1/2πσ y , and A � 2πσ x σ y .
detection.
It is worth noticing that it is necessary to specify tuning
Based on such motivation, this study proposes and
parameters of the GF including the orientation angles and
verifies an automated method for recognizing localized spall
the radial frequency. Based on the suggestions of Jain and
objects based on an integration of a Gabor filter, k-means
Farrokhnia [45], four values of orientations are often
clustering, image texture analysis, and logistic regression
employed: 0°, 45°, 90°, and 135°. Given an image with a width
pattern classification models. The Gabor filter coupled with
of Nw pixels, the√following
� √� √ values
� of radial√� frequency u0 can
the principal component analysis and the k-means clustering
be considered: 1 2, 2 2, 4 2, ..., Nw /4 2.
are used synergistically for automatic identification of ROI
on concrete surface. The image texture analysis combines
powerful texture discriminators of binary gradient contours, 2.2. The K-Means Clustering Algorithm. In this study, the
color channels’ properties, and the gray level co-occurrence unsupervised machine learning approach of k-means clus-
matrix. The logistic regression model trained by the state-of- tering [46] is employed to divide an image into different
the-art adaptive moment estimation (Adam) optimizer is regions based on the analysis results obtained from the GF.
employed for data classification. This unsupervised machine learning method is simple yet
The subsequent sections of the study are organized as powerful algorithm for automatic data grouping [47]. Based
follows: the second section reviews the research method- on such method, image pixels that have the similar prop-
ology. The third section presents the image data collection erties can be grouped in one cluster. Accordingly, data
process. The proposed integrated model used for concrete samples belonging to one cluster feature the smallest degree
spall detection is described in the next section, followed by of variation. The iterative algorithm used to compute the
the experimental results and discussion. The final section cluster centers is presented in Algorithm 1.
summarizes the research findings with several concluding
remarks.
2.3. Image Texture Analysis
2. Research Methodology
2.3.1. Binary Gradient Contours (BGC). The BGC, proposed
2.1. Gabor Filter (GF). Image segmentation is the process of by Fernández et al. [48], is a group of computationally simple
separating an image into distinctive regions [38, 39]. The GF texture descriptors. Given a 3 × 3 grayscale image patch,
is a widely applied approach for segmenting image [40, 41]. these texture descriptors employs a set of eight binary
This approach is inspired by the multichannel operation of gradients between pairs of pixels all along a closed path
the human visual system used for visual interpretation in around the central pixel [49]. The BGC includes three
real-world circumstances [42–44]. Based on experimenta- versions which are single-loop, double-loop, and triple-loop
tion, it has been shown that the GF resembles simple cells in descriptors. Via experimentation, the BGC operator has
the Mammalian vision system. Thus, this filter can be a been found to achieve good texture discrimination
reasonable model of how humans actually recognize and outcomes.
discriminate areas characterized by different texture [45]. A matrix S which is the pixel intensity of an image patch
The GF consists of two-dimensional Gabor filters which of the size 3 × 3 is given as follows:
can be described as complex sinusoidal waves modulated by I7 I6 I5
Gaussian envelopes [43]. This filter carries out a localized ⎡⎢⎢⎢ ⎤⎥⎥⎥
and oriented frequency analysis of a two-dimensional signal. S � ⎢⎢⎢⎢⎣ I0 Ic I4 ⎥⎥⎥⎥⎦, (3)
The GF yields a response that can be mathematically given as I1 I2 I3
follows [45]:
4 Advances in Civil Engineering

Determine the number of cluster k

Randomly assign k centers of data samples
(1) Loop
(2) Assign each data points to the cluster with the nearest mean
(3) Recalculate means for data points assigned to each cluster
(4) Until the data assignments are unchanged.

ALGORITHM 1: The k-means clustering.

where Ic denotes the central pixel. I0, I1, . . ., I7 are the I7 I6 I5

neighboring pixels.
I0 Ic I4
The schematic representations of BGC with three ver-
sions of single, double, and triple loops are presented in I1 I2 I3
Figure 2. In addition, to facilitate the mathematical for- Single loop Double loop Triple loop
mulation of these texture descriptors, a square crop Sm,n is
Figure 2: The graphical representation of BGC.
given by

Im−1,n−1 Im−1,n Im−1,n−1 2.3.3. Gray Level Co-Occurrence Matrix (GLCM). The
⎢
⎡
⎢
⎢ ⎤⎥⎥⎥
⎢
S �⎢ ⎥⎥ GLCM [51, 52] is also an extensively employed approach for
⎣ Im,n−1 Im,n Im,n+1 ⎥⎥⎦,
⎢
⎢ (4)
characterizing image texture. This approach focuses on
Im+1,n−1 Im+1,n Im+1,n+1
capturing the repeated occurrence of certain gray-level
where Im,n represents the pixel at mth row and nth column. patterns [53]. Therefore, indices extracted from a GLCM can
Accordingly, the formulations of the single, double, and be effectively utilized to evaluate the coarseness/fineness of
triple-loop versions are given by [48] an image region. Let r and θ represent a distance and a
rotation relationship between two individual pixels. The
(i) Single-loop version: GLCM, denoted as Pδ , denotes a probability of the two gray
levels of i and j having the relationship specifying by r and θ
7
[54]. Based on the recommendations of Haralick et al. [51],
BGC1 � 􏽘 λ( In − I( n+1 )mod 8 ) × 2n − 1. (5)
the GLCM can be constructed with r � 1 and θ � 0°, 45°, 90°,
n�0
and 135°. Accordingly, for each matrix, four indices of
angular second moment (AM), contrast (CO), correlation
(ii) Double-loop version:
(CR), and entropy (ET) can be computed as follows [29, 55]:
3
Ng Ng
BGC2 � 15 × 􏽘 λ( I2n − I2( n+1 )mod 8 ) × 2n 2
n�0 AM � 􏽘 􏽘 PN
δ (i, j) ,
(6) i�1 j�1
3
n
+ 􏽘 λ( I2n+1 − I( 2n+3 )mod 8 ) × 2 − 16. Ng −1 Ng Ng
n�0
CO � 􏽘 k2 􏽘 􏽘 PN
δ (i, j),
k�0 i�1 j�1
(iii) Triple-loop version:
7 |i−j|�k (8)
BGC3 � 􏽘 λ( I3nmod 8 − I3( n+1 )mod 8 ) × 2n − 1,
N N
n�0
(7) 􏽐i�1g 􏽐j�1g i × j × PN
δ (i, j) − μX μY
1, if x ≥ 0, CR � ,
where λ( x ) � 􏼨 σXσY
0, if x < 0.
Ng Ng
ET � − 􏽘 􏽘 PN N
δ (i, j)log􏼐Pδ (i, j)􏼑,
i�1 j�1
2.3.2. RGB Channels’ Properties. Since the color properties
of spall and nonspall objects are expected to be dissimilar, where Ng denotes the number of gray level values;
this study employs the statistical measurements of three μX , μY , σ X , and σ Y are the means and standard deviations of
color channels: red (R), green (G), and blue (B) as a means of the marginal distribution with respect to PN δ (i, j).
texture description. Given an image sample I, the first-order
histogram P(I) can be computed. Accordingly, the mean
(μc ), standard deviation (σ c ), skewness (Sc ), kurtosis (Kc), 2.4. Logistic Regression Model (LRM). The LRM is a capable
entropy (Ec ), and range (Rc) of the three color channels (R, method for solving binary classification problems
G, and B) can be calculated [29, 50]. [29, 56–58]. The task at hand is to construct a decision
Advances in Civil Engineering 5

boundary that categorizes the input data into two distinctive model training phase by adaptively fine-tuning the Lr
regions. Therefore, given a vector of input data parameter.
xi � xi1 , xi2 , ..., xi D , where D is the number of the features Adam harnesses information obtained from the average
used for classification, the model is able to derive the class of the second moments of the gradients. In detail, this
output y with either y � 0 (for the negative class of nonspall) optimization algorithm computes an exponential moving
and y � 1 (for the positive class of spall). average of the gradient and the square gradient. Moreover, a
The probability of the positive class hθ (xi ) derived by a set of parameters (β1 and β2) is used to dictate the decay rates
LRM is given by [59] of these moving averages [64]. Via experimentation, it can be
1 shown that the advantages of this optimizer include efficient
hθ xi 􏼁 � hθ xi1 , xi2 , ..., xi D 􏼁 � computation, straight forward implementation, no memory
1 + exp −ηi 􏼁
requirements, and the capability of dealing with a large
(9) number of optimized parameters [63].
1
� , In order to implement Adam to optimize a LRM, it is
1 + exp􏼐−θT xi 􏼑 necessary to compute the gradient (gt ). The gradient gt in
the case of using the least square loss function is given by
where ηi � θ0 + θ1 xi1 + θ2 xi2 + · · · + θD xi D � θT xi . The
[60]
vector θ � θ0 , θ1 , θ2 , ..., θD is the model parameter.
As a supervised learning approach, a set of training z Loss(θ)
examples needs to be prepared so that the vector θ can be gt � � 2 y − hθ xi 􏼁􏼁 × hθ xi 􏼁 × 1 − hθ xi 􏼁􏼁 × xi,k .
zθk
adapted during the model training phase. A LRM can be
(14)
trained by either minimizing the least square loss function or
maximizing the log likelihood function. If the log likelihood function is employed, the gradient gt
The least square loss function is given by [60] is given by [22, 61, 62]
M
Loss (θ) � 􏽘 yi − hθ xi 􏼁􏼁 ,
2
(10) z Loss(θ)
gt � � yi − hθ xi 􏼁􏼁xi,k . (15)
i�1 z θk 􏼁
where M is the number of training data. Accordingly, the Adam procedure (illustrated in Algo-
The log likelihood function is described as follows rithm 2) used for training a LRM can be performed itera-
[61, 62]: tively with the following steps:
M (i) Compute gradient gt
Loss ( θ ) � log( L( θ ) ) � 􏽘 yi log( hθ ( xi )
i�1 (11) (ii) Update the biased first and second raw moment
estimates
+( 1 − yi )( 1 − log( hθ ( xi ) ).
(iii) Compute the bias-corrected moment estimates
A LRM can be trained via the stochastic gradient descent (iv) Adapt the optimized parameters
framework [29]. If the least square loss function is used, the
update rule for adapting the model parameters is given by
[60] 3. Collection of Image Samples
θk � θk + Lr × y − hθ xi 􏼁􏼁 × hθ xi 􏼁 1 − hθ xi 􏼁􏼁 × xi,k , The LRM used in this study belongs to the category of
(12) supervised machine learning methods. To train this LRM
with the use of the aforementioned Adam optimizer, it is a
where Lr denotes the learning rate parameter. requisite to prepare a set of training image samples as well as
Meanwhile, if the log likelihood function is selected, the a set of testing image samples to verify the model con-
update rule used that compute θ is given by [61, 62] struction phase. Therefore, this study has carried out field
surveys at several high-rise buildings in Danang city
θk � θk + Lr × yi − hθ xi 􏼁􏼁xi,k . (13)
(Vietnam) to collect a set of 600 image samples. Among
them, 300 samples contain localized spall objects and 300
samples consist of nonlocalized spall objects. Notably, image
2.5. Adaptive Moment Estimation (Adam) Optimizer. samples of the two class of nonspall (class label � 0) and spall
Adam, proposed by [63], is designed as an algorithm for (class label � 1) have been assigned by a human inspector for
first-order gradient-based optimization of stochastic ob- the purposes of model training and testing. The Cannon EOS
jective functions. This algorithm is relied on adaptive esti- M10 (CMOS 18.0 MP) and Nikon D5100 (CMOS 16.2 MP)
mates of lower-order moments. Adam can be considered as have been employed to collect image samples. In addition,
an extension of the stochastic gradient descent employed to the image size has been standardized to be 64 × 64 to fa-
train machine learning models via an iterative weight cilitate the computation process. The image set has been
updating process [64]. It is noted that the conventional collected so that a diverse background (e.g., cracks and
stochastic gradient descent employs a constant learning rate stains) can be included. The collected image set is demon-
(Lr) for all weight updates. Adam seeks for improving the strated in Figure 3.
6 Advances in Civil Engineering

Define step size a � 0.001

Define exponential decay rates β1 � 0.9 and β2 � 0.9999
Define the objective function f(θ)
Randomly initialize the searched variable θ
Assign m0 � 0, v0 � 0, and t � 0
(1) While (θt not converged)
(2) t�t+1
(3) Compute gradient: gt � ∇θ ft (θt−1 )
(4) Update biased 1st moment estimate
(5) mt � β1 mt−1 + (1 − β1 )gt
(6) Update biased 2nd raw moment estimate
(7) vt � β2 vt−1 + (1 − β2 )g2t
(8) Calculate bias-corrected first moment estimate
(9) 􏽢 t � mt /(1 − βt1 )
m
(10) Calculate bias-corrected 2nd raw moment estimate
(11) 􏽢vt � vt /(1 − βt2 )
(12) Update the searched parameter 􏽰��
(13) θt � θt−1 − α × (m 􏽢 t / 􏽢vt + ε)
(14) End While
(15) Return θt

ALGORITHM 2: The Adam optimization procedure.

(a)

(b)

Figure 3: Demonstration of the collected image samples: (a) images containing localized spall objects and (b) images containing non-
localized spall objects.

4. The Proposed Hybrid Approach of Image (ii) Image texture computation

Processing and Machine Learning (iii) Machine learning-based pattern classification
Approach for Automatic Detection of
Concrete Spall
4.1. Automatic Region of Interest (ROI) Identification. To deal
This section of the study aims at describing the structure of with the diverse shapes of localized spall objects, this study
the proposed hybrid approach of image processing and relies on the techniques of GF to automatically identify ROIs
machine learning used for recognizing localized spall object. that contains the potential defects of interest. It is noted that
The overall structure of the proposed approach is presented an image sample has been denoised by a median filter with a
in Figure 4. It is noted that the hybrid model used for window size of 4 pixels and converted to a grayscale one.
automatic concrete spall detection has been developed in After the GFs with different orientations and radial fre-
Visual C#.NET environment (Framework 4.6.2) and quency are computed, the principal component analysis
implemented with the ASUS FX705GE–EW165T (Core i7 (PCA) is performed to transform the set of GFs and reduce
8750H, 8 GB Ram, 256 GB solid-state drive). the data dimensionality (Figure 5). The number of the PCA
The model can be divided into several operational steps: transformed data is selected corresponding to 99% of cu-
(i) Automatic ROI identification mulative variance explained. It is noted that the GF and the
Advances in Civil Engineering 7

Gabor filter Training images

K-Means Region of interest

Training samples Binary gradient
clustering (ROI) extraction
contours
Connected component Image texture Gray level co-
labeling operation computation occurrence matrix
RGB channel’s
statistical properties
Adam-LL
Training method Logistic regression
selection model training
Adam-LS

Testing samples Logistic regression

model prediction
Testing images
Image texture
computation Detected spall
objects

Figure 4: The proposed model structure.

PCA operations have been implemented via built-in func- RGB image Grayscale image
tions provided by the Accord.NET Framework [65].
Based on the PCA result, the k-means clustering algo-
rithm is used to segment the image sample. Via experi-
mentation, the suitable number of clusters for the collected
dataset is found to be 3. Subsequently, the morphological
operation of filling and removing small objects are utilized to
process the segmented image. Moreover, the operation of Gabor filter
background removal is performed to remove redundant
objects. In this study, an object within an image sample is Principal component
considered to be background if its width or height is equal to analysis
that of the image sample.
Accordingly, each image cluster or segment is presented
as a binary image. The connected component labeling al-
gorithm [66] is then used to analyze the position of the
binary-1 pixels and separate them into distinctive compo-
nent regions. Essentially, all pixels having value binary 1 and
are connected to each other are grouped into one object [38].
The 1st principal The 2nd principal The 3rd principal
To remove crack objects, for each grouped pixels obtained component component component
from the connected component labeling analysis, an object
slenderness index (OSI) is computed as follows: Figure 5: Image processed by the Gabor filter.

LOX LOY 4.2. Image Texture Computation. Based on ROIs obtained

OSI � Max􏼠 , 􏼡, (16)
μOX μOY from the previous step, image texture analysis consisting of
statistical measurements of BGC, RGB channels, and GLCM
where LOX and LOY are the object lengths along the X axis is carried out. The BGC texture descriptor includes all of the
and Y axis, respectively. μOX and μOY denote the mean three variants of single, double, and triple loops. Each of the
thicknesses of the object along the X axis and Y axis, variants yields a histogram which describes the texture
respectively. property of an image sample. Accordingly, statistical indices
If the OSI of an object is greater than a certain threshold including mean, standard deviation, skewness, kurtosis, and
(TOSI), this object is classified as a crack. Via several trial-and- entropy can be computed for each histogram. Therefore, the
error experiments with the collected image sample, a suitable BGC results in 15 numerical features. As mentioned earlier,
value for the threshold TOSI is found to be 5. After the ROIs the mean (μc ), standard deviation (σ c ), skewness (Sc ),
have been identified, operations of image convolution and kurtosis (Kc), entropy (Ec ), and range (Rc) of the three color
cropping are employed to isolate the areas of interest within channels (R, G, and B) are used to represent the color
the image sample. The processes of ROI identification and features of image samples. Thus, there are 6 × 3 � 18 addi-
isolation are demonstrated in Figures 6 and 7. tional numerical features. Moreover, properties of the
8 Advances in Civil Engineering

Binarized segments Segmented objects

Image sample Gabor filter Segmented image Extracted ROI

(a)
Binarized segments Segmented objects

Image sample Gabor filter Segmented image Extracted ROI

(b)

Figure 6: ROI extraction for images containing localized spall objects: (a) one object and (b) multiple objects.

GLCM including the four indices of angular second moment Adam-LS and Adam-LL. It is noted that 90% of the collected
(AM), contrast (CO), correlation (CR), and entropy (ET) are dataset has been employed to construct the LRM model.
used. It is noted that for each image sample, four GLCMs are Meanwhile, the rest of the dataset is reserved to verify the
established. Thus, the GLCM texture descriptor yields generalization capability of the model.
4 × 4 � 16 features. In total, there are 15 + 18 + 16 � 49 nu-
merical features that can be extracted from the image texture
computation process.
5. Research Findings and Discussion
As mentioned earlier, the whole collected dataset is divided
into two subsets: a training set (90%) and a testing set (10%).
4.3. Pattern Classification Using LRM Trained by the Adam Moreover, to diminish the effect of randomness brought
Optimizer. Using the extracted ROIs and the aforemen- about by data sampling and to assess the generalization
tioned texture descriptors, a dataset with 790 samples and 49 capability of the integrated method reliably, the data sam-
features can be constructed. This dataset contains 465 pling process has been repeated 20 times. A partitioned
nonlocalized spall samples and 325 localized spall samples. datasets used for model training and testing are demon-
As stated earlier, the output class is either 0 for the negative strated in Table 1. In addition, the LRM trained with the
class and 1 for the positive class. Moreover, in order to stochastic gradient descent algorithm with the least square
standardize the input features’ range, the numerical texture and log likelihood loss function are employed as benchmark
descriptors have been normalized by the Z-score equation as models. The stochastic gradient descent models coupled
follows: with the former and later loss function are denoted as LS-LR
Xo − m X and LL-LR, respectively. Furthermore, the two LRMs trained
XZN � , (17) with the Adam optimizer are denoted as Adam-LS-LR and
sX
Adam-LL-LR. All of the LRMs have been trained with 1000
where Xo and XZN represent the original and normalized iterations.
input data, respectively. mX and sX are the mean and the In addition, the classification accuracy rate (CAR),
standard deviation of the original input data, respectively. precision, recall, negative predictive value (NPV), and F1
Based on the aforementioned dataset, the LRM is trained score are computed to quantify the model predictive ac-
with the Adam optimizer using the least square and log curacy. These performance measurement indices are pro-
likelihood loss functions. These two LRM is denoted as vided as follows [67]:
Advances in Civil Engineering 9

Binarized segments Segmented objects

Image sample Gabor filter Segmented image

Extracted ROI

(a)
Binarized segments Segmented objects

Extracted ROI
Image sample Gabor filter Segmented image

(b)

Figure 7: ROI extraction for images containing nonlocalized spall objects: (a) one object and (b) multiple objects.

Table 1: Demonstration of the collected dataset.

Set Case IF1 IF2 IF3 IF4 ... IF46 IF47 IF48 IF49 Class label
1 39.64 37.27 34.53 34.33 ... 0.10 745.72 0.00 6.47 0
2 12.92 11.63 9.60 18.34 ... 0.17 497.69 0.00 3.82 0
3 68.21 65.76 59.38 40.67 ... 0.03 770.71 0.00 6.82 0
4 44.78 45.60 43.86 56.59 ... 0.29 1835.96 0.00 4.39 0
5 65.40 64.84 66.24 60.02 ... 0.12 2458.40 0.00 4.12 0
Training ... ... ... ... ... ... ... ... ... ... ...
707 68.85 58.73 45.76 34.57 ... 0.01 824.92 0.00 6.77 1
708 55.85 51.50 46.05 26.86 ... 0.01 762.79 0.00 7.48 1
709 51.27 47.45 42.09 35.80 ... 0.05 928.80 0.00 6.93 1
710 60.32 55.98 49.76 31.67 ... 0.01 1340.14 0.00 6.67 1
711 73.54 65.91 62.20 44.11 ... 0.02 1180.24 0.00 7.82 1
1 56.26 58.54 61.95 25.44 ... 0.01 498.38 0.00 7.27 0
2 93.11 95.22 103.83 41.81 ... 0.10 1353.60 0.00 3.32 0
3 85.93 83.79 93.25 49.92 ... 0.07 3975.50 0.00 3.95 0
4 58.64 59.44 64.39 55.01 ... 0.18 968.76 0.00 5.57 0
5 83.84 84.93 88.35 68.18 ... 0.14 493.65 0.00 5.54 0
Testing ... ... ... ... ... ... ... ... ... ... ...
75 62.98 47.68 28.07 32.93 ... 0.03 244.76 0.00 7.79 1
76 46.17 44.28 39.69 33.57 ... 0.08 467.10 0.00 6.35 1
77 51.47 46.08 37.80 20.61 ... 0.01 370.11 0.00 8.62 1
78 34.17 30.02 24.89 29.42 ... 0.10 413.32 0.00 6.74 1
79 44.89 37.54 27.15 56.16 ... 0.32 528.69 0.00 5.10 1
10 Advances in Civil Engineering

Table 2: Prediction result comparison.

Adam-LL-LR LL-LR Adam-LS-LR LS-LR
Phase Indices
Mean Std. Mean Std. Mean Std. Mean Std.
CAR (%) 85.27 1.76 84.52 3.82 72.52 1.69 71.81 2.52
TP 246.95 12.48 239.70 26.87 271.85 8.71 265.65 14.59
TN 359.30 13.68 361.25 14.25 243.75 13.76 244.90 11.57
FP 46.85 12.77 53.15 26.97 22.50 7.68 28.25 15.63
Training FN 57.90 14.13 56.90 14.71 172.90 12.92 172.20 10.72
Precision 0.84 0.04 0.82 0.09 0.92 0.03 0.90 0.05
Recall 0.81 0.04 0.81 0.04 0.61 0.02 0.61 0.02
NPV 0.86 0.03 0.86 0.03 0.58 0.03 0.59 0.03
F1 score 0.82 0.02 0.81 0.06 0.74 0.01 0.73 0.03
CAR (%) 85.32 4.64 81.90 4.06 72.03 5.37 70.82 5.92
TP 26.90 5.19 25.25 3.94 27.60 3.29 28.10 6.03
TN 40.50 5.68 39.45 4.78 29.30 3.45 27.85 5.17
FP 4.30 2.26 6.90 2.62 3.05 2.42 3.00 2.02
Testing FN 7.30 2.76 7.40 2.91 19.05 4.34 20.05 3.73
Precision 0.86 0.07 0.79 0.08 0.90 0.07 0.90 0.09
Recall 0.79 0.08 0.78 0.06 0.59 0.07 0.58 0.09
NPV 0.85 0.06 0.84 0.07 0.61 0.08 0.58 0.07
F1 score 0.82 0.06 0.78 0.05 0.71 0.06 0.70 0.08

TP + TN 1.00
CAR � × 100%,
TP + TN + FP + FN 0.90
0.80
TP
Precision � , 0.70
TP + FP
0.60
TP 0.50
Recall � , (18)
TP + FN 0.40

TN 0.30
NPV � , 0.20
TN + FN
0.10
2TP 0.00
F1 score � ,
2TP + FP + FN CAR Precision Recall NPV F1 score
where FN, FP, TP, and TN denote the number of false- Adam-LL-LR Adam-LS-LR
negative, false-positive, true-positive, and true-negative LL-LR LS-LR
samples, respectively.
Figure 8: Performance measurement indices for the training
The experimental results obtained from the repetitive
phase.
data sampling with 20 runs are reported in Table 2. As can be
seen from this table, the Adam-LL-LR has achieved the best
predictive accuracy in both of the training (CAR � 85.25%, performances of LRMs. In addition, to confirm the statistical
precision � 0.84, recall � 0.81, NPV � 0.86, and F1 score- difference of each pair of the localized spall detection
� 0.82) and testing phases (CAR � 85.32%, precision � 0.86, models, the Wilcoxon signed-rank test with a significance
recall � 0.79, NPV � 0.85, and F1 score � 0.82). Since the level (p value) � 0.05 is employed in this section of the study.
prediction performances obtained from the training and The test outcomes are reported in Table 3. Observably,
testing phases of the Adam-LL-LR are relatively similar, it experimental results show that all of the p values are lower
can be shown that this model has not suffered from over- than the significance level. Thus, the null hypothesis shows
fitting. In addition, the LL-LR model is the second best that the performances of the two models under testing are
approach (with CAR � 81.90% and F1 score � 0.78), followed statistically indifferent and can be confidently rejected. This
by the Adam-LS-LL (with CAR � 72.03% and F1 score- hypothesis test asserts the superiority of the Adam-LL-LR
� 0.71) and the LS-LR (with CAR � 70.82% and F1 score- model over other benchmark approaches.
� 0.70). Herein, the index of the F1 score is emphasized Based on the experimental result, the Adam-LL-LR
because it presents the harmonic mean of the precision and model is best suited for the collected dataset at hand. The
recall. performance of this model is further studied in this section.
The training and testing performances of the employed Illustrations of correctly recognized spall objects yielded by
models are graphically presented in Figures 8 and 9. The Adam-LL-LR are presented in Figure 11. As can be observed,
boxplot shown in Figure 10 demonstrates the testing the model can deliver accurate detection results in the
Advances in Civil Engineering 11

1.00
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
CAR Precision Recall NPV F1 score

Adam-LL-LR Adam-LS-LR
LL-LR LS-LR

Figure 9: Performance measurement indices for the testing phase.

95
0.9
90
0.85
85 0.8
80 0.75
0.7
F1 score

75
CAR

0.65
70
0.6
65 0.55
60 0.5
0.45
55
0.4
50
Adam-LL-LR LL-LR Adam-LS-LR LS-LR Adam-LL-LR LL-LR Adam-LS-LR LS-LR
Spall detection models Spall detection models

(a) (b)

Figure 10: Model performances obtained from the repetitive data subsampling process: (a) CAR (%) and (b) F1 score.

Table 3: Wilcoxon signed-rank test results.

Indices Model comparison Test outcome p value
Adam-LL-LR vs. LL-LR Significant 0.0184
CAR (%) Adam-LL-LR vs. Adam-LS-LR Significant 0.0001
Adam-LL-LR vs. LS-LR Significant 0.0001
Adam-LL-LR vs. LL-LR Significant 0.0400
F1 score Adam-LL-LR vs. Adam-LS-LR Significant 0.0004
Adam-LL-LR vs. LS-LR Significant 0.0001

presence of a window (Figure 11(a)) and a minor defect on Adam-LL-LR model can be used to quantify the percentage
the mortar surface (Figure 11(b)). Notably, the localized of damaged areas found in image samples; the computation
spall objects can still be located well in the cases that there are results are demonstrated in Figure 12.
crack objects in the captured scenes (Figures 11(c)–11(e)). Nevertheless, as shown in Figure 13, the newly developed
Furthermore, Adam-LL-LR has also performed well in the model has made incorrect detection results in the cases of
cases that there are multiple spall objects in the image complex background. As observed in Figure 13(a), an area in
samples (Figures 11(f ) and 11(g)). In addition, the proposed the background has the texture property similar to that of the
12 Advances in Civil Engineering

Binary-detected Detected object with

Original image Segmented image positive probability (Prob.)
object

(a)

(b)

(c)

(d)

(e)

(f)

(g)
Figure 11: Illustrations of correctly classified cases.
Advances in Civil Engineering 13

Original image Segmented image Detected objects Computation results

Percentage of damaged
area: 8.70%

Percentage of damaged
area of the 1st object: 4.88%

Percentage of damaged
area of the 2nd object: 4.84%

Figure 12: Demonstration of the calculation of the percentage of damaged areas.

Falsely
Original Segmented Binary-detected
detected/undetected
image image object
object

(a)

Undetected object

(b)

Undetected object

(c)
Figure 13: Illustrations of incorrectly classified cases.

spall object. This can lead to a false-positive detection. localized spall recognition. The newly developed model is a
Complex background (Figure 13(b)) and irregular lighting hybridization of image processing and machine learning
conditions (Figure 13(c)) also tend to reduce the model approaches. Image processing methods of the GF coupled
accuracy. These phenomena can lead to false-negative cases. with k-means clustering and morphological analyses are
used to automatically identify the ROIs that potentially
6. Concluding Remarks contain the defect. The BGC, GLCM, and color channels’
properties are employed as texture descriptors. Based on the
Localized spall is a common defect observed on surfaces of computed image texture, the LRM optimized by the state-of-
reinforced concrete elements. Accurate detection of this the-art Adam is used to construct a decision boundary that
damage is crucial during the phase of the periodic structural separate the data samples into two regions of nonlocalized
heath survey. This study has developed and verified a spall and localized spall. Experimental results show that the
computer vision-based approach for automating the task of LRM trained by the Adam optimizer can deliver the most
14 Advances in Civil Engineering

desired prediction accuracy. Therefore, the proposed inte- [11] Z. Chen and T. C. Hutchinson, “Image-based framework for
grated model can be a useful tool to assist building main- concrete surface crack monitoring and quantification,” Ad-
tenance agencies in the task of evaluating structure heath vances in Civil Engineering, vol. 2010, Article ID 215295,
condition. 18 pages, 2010.
[12] G. K. Choudhary and S. Dey, “Crack detection in concrete
surfaces using image processing, fuzzy logic, and neural
Data Availability networks,” in Proceedings of the 2012 IEEE Fifth International
The image dataset used to support the findings of this study has Conference on Advanced Computational Intelligence (ICACI),
pp. 404–411, Nanjing, China, October 2012.
been deposited in the repository of GitHub: https://github.
[13] S. Dorafshan, R. J. Thomas, and M. Maguire, “Benchmarking
com/NhatDucHoang/LocalizedSpallDetection_AdamLRM. image processing algorithms for unmanned aerial system-
assisted crack detection in concrete structures,” Infrastruc-
Conflicts of Interest tures, vol. 4, no. 2, p. 19, 2019.
[14] N.-D. Hoang, “Detection of surface crack in building struc-
The author declares that there are no conflicts of interest. tures using image processing technique with an improved
otsu method for image thresholding,” Advances in Civil
Acknowledgments Engineering, vol. 2018, Article ID 3924120, 10 pages, 2018.
[15] N.-D. Hoang and Q.-L. Nguyen, “Metaheuristic optimized
This research was financially supported by Duy Tan edge detection for recognition of concrete wall cracks: a
University. comparative study on the performances of roberts, prewitt,
Canny, and sobel algorithms,” Advances in Civil Engineering,
References vol. 2018, Article ID 7163580, 16 pages, 2018.
[16] H. Kim, J. Lee, E. Ahn, S. Cho, M. Shin, and S.-H. Sim,
[1] M. R. Jahanshahi and S. F. Masri, “Adaptive vision-based “Concrete crack identification using a UAV incorporating
crack detection using 3D scene reconstruction for condition hybrid image processing,” Sensors, vol. 17, no. 9, p. 2052, 2017.
assessment of structures,” Automation in Construction, [17] S. Li and X. Zhao, “Image-based concrete crack detection
vol. 22, pp. 567–576, 2012. using convolutional neural network and exhaustive search
[2] B. A. Soeleman, M. V. A. D. Silva, M. Akkari, and C. Santili, technique,” Advances in Civil Engineering, vol. 2019, Article
“Accidents due to falls from roof slabs,” Sao Paulo Medical ID 6520620, 12 pages, 2019.
Journal, vol. 131, no. 3, pp. 153–157, 2013. [18] M.-M. Naddaf-Sh, S. Hosseini, J. Zhang, N. A. Brake, and
[3] S. M. Verhulst, Spalling Is Bad, Falling Is Worse, Pieresearch, H. Zargarzadeh, “Real-time road crack mapping using an
Arlington, TX, USA, 2020, https://pieresearchcom/spalling- optimized convolutional neural network,” Complexity,
is-bad-falling-is-worse/. vol. 2019, Article ID 2470735, 17 pages, 2019.
[4] M. Duckett and G. Pestine, Concrete Spalling Robson Forensic, [19] W. Song, G. Jia, H. Zhu, D. Jia, and L. Gao, “Automated
https://wwwrobsonforensiccom/articles/concrete-spalling- pavement crack damage detection using deep multiscale
expert-article/, 2020. convolutional features,” Journal of Advanced Transportation,
[5] C. Koch, K. Georgieva, V. Kasireddy, B. Akinci, and vol. 2020, Article ID 6412562, 11 pages, 2020.
P. Fieguth, “A review on computer vision based defect de- [20] B. Wang, Y. Li, W. Zhao, Z. Zhang, Y. Zhang, and Z. Wang,
tection and condition assessment of concrete and asphalt civil “Effective crack damage detection using multilayer sparse
infrastructure,” Advanced Engineering Informatics, vol. 29, feature representation and incremental extreme learning
no. 2, pp. 196–210, 2015.
machine,” Applied Sciences, vol. 9, no. 3, p. 614, 2019.
[6] R. Zaurin, “Structural health monitoring with emphasis on
[21] T. Dawood, Z. Zhu, and T. Zayed, “Machine vision-based
computer vision, damage indices, and statistical analysis,”
model for spalling detection and quantification in subway
https://starslibraryucfedu/cgi/viewcontentcgi?
networks,” Automation in Construction, vol. 81, pp. 149–160,
article=4853&context=etdElectronic theses and Dissertations,
2017.
University of Central Florida, Orlando, FL, USA, 2009,
[22] N.-D. Hoang, Q.-L. Nguyen, and X.-L. Tran, “Automatic
https://starslibraryucfedu/cgi/viewcontentcgi?
detection of concrete spalling using piecewise linear stochastic
article=4853&context=etdElectronic theses and Dissertations.
[7] I. Abdel-Qader, S. Pashaie-Rad, O. Abudayyeh, and S. Yehia, gradient descent logistic regression and image texture anal-
“PCA-Based algorithm for unsupervised bridge crack de- ysis,” Complexity, vol. 2019, Article ID 5910625, 14 pages,
tection,” Advances in Engineering Software, vol. 37, no. 12, 2019.
pp. 771–778, 2006. [23] C. Koch, S. G. Paal, A. Rashidi, Z. Zhu, M. König, and
[8] M. O’Byrne, F. Schoefs, B. Ghosh, and V. Pakrashi, “Texture I. Brilakis, “Achievements and challenges in machine vision-
analysis based damage detection of ageing,” Infrastructural based inspection of large concrete structures,” Advances in
Elements Computer-Aided Civil and Infrastructure Engineer- Structural Engineering, vol. 17, no. 3, pp. 303–318, 2014.
ing, vol. 28, pp. 162–177, 2013. [24] H. Wu, X. Ao, Z. Chen, C. Liu, Z. Xu, and P. Yu, “Concrete
[9] D. Lattanzi and G. R. Miller, “Robust automated concrete spalling detection for metro tunnel from point cloud based on
damage detection algorithms for field applications,” Journal of roughness,” Descriptor Journal of Sensors, vol. 2019, Article ID
Computing in Civil Engineering, vol. 28, no. 2, pp. 253–262, 8574750, 12 pages, 2019.
2014. [25] S. German, I. Brilakis, and R. DesRoches, “Rapid entropy-
[10] D. B. Agyemang and M. Bader, “Surface crack detection using based detection and properties measurement of concrete
hierarchal convolutional neural network,” in Advances in spalling with machine vision for post-earthquake safety as-
Computational Intelligence Systems, pp. 173–186, Springer sessments,” Advanced Engineering Informatics, vol. 26, no. 4,
International Publishing, Cham, Switzerland, 2020. pp. 846–858, 2012.
Advances in Civil Engineering 15

[26] P. Prasanna, K. J. Dana, N. Gucunski et al., “Automated crack filters,” Computer-Aided Civil and Infrastructure Engineering,
detection on concrete bridges,” IEEE Transactions on Auto- vol. 29, pp. 342–358, 2014.
mation Science and Engineering, vol. 13, no. 2, pp. 591–599, [43] K. Hammouda, Texture Segmentation Using Gabor Filters
2016. Technical Report, University of Waterloo, Waterloo, Canada,
[27] M.-K. Kim, H. Sohn, and C.-C. Chang, “Localization and 2000, http://wwwpamiuwaterlooca/pub/hammouda/sd775-
quantification of concrete spalling defects using terrestrial paperpdf.
laser scanning,” Journal of Computing in Civil Engineering, [44] N. C. Kim and H. J. So, “Directional statistical Gabor features
vol. 29, Article ID 04014086, 2015. for texture classification,” Pattern Recognition Letters, vol. 112,
[28] N.-D. Hoang, “Image processing-based recognition of wall pp. 18–26, 2018.
defects using machine learning approaches and steerable [45] A. K. Jain and F. Farrokhnia, “Unsupervised texture seg-
filters,” Computational Intelligence and Neuroscience, mentation using Gabor filters,” Pattern Recognition, vol. 24,
vol. 2018, Article ID 7913952, 18 pages, 2018. no. 12, pp. 1167–1186, 1991.
[29] N.-D. Hoang, “Automatic detection of asphalt pavement [46] J. MacQueen, “Some methods for classification and analysis of
raveling using image texture based feature extraction and multivariate observations,” in Proceedings of the Fifth Berkeley
stochastic gradient descent logistic regression,” Automation in Sympos Math Statist and Probability (Berkeley, Calif, 1965/66)
Construction, vol. 105, Article ID 102843, 2019. Vol I: Statistics, pp. 281–297, Univ California Press, Berkeley,
[30] G. Yao, F. Wei, Y. Yang, and Y. Sun, “Deep-learning-based CA, USA, 1967.
bughole detection for concrete surface image,” Advances in [47] R. Duda, P. Hart, and D. Stock, Pattern Classification, John
Civil Engineering, vol. 2019, Article ID 8582963, 12 pages, Wiley & Sons, Hoboken, NJ, USA, 2001, ISBN-13: 978-
2019. 0471056690.
[31] Y. Li, X. Li, H. Wang, S. Wang, S. Gu, and H. Zhang, “Exposed [48] A. Fernández, M. X. Álvarez, and F. Bianconi, “Image clas-
aggregate detection of stilling basin slabs using attention sification with binary gradient contours,” Optics and Lasers in
U-net network,” KSCE Journal of Civil Engineering, vol. 24, Engineering, vol. 49, no. 9-10, pp. 1177–1184, 2011.
no. 6, p. 1740, 2020. [49] F. Alaei, A. Alaei, U. Pal, and M. Blumenstein, “A comparative
[32] J. K. Dey, Z. Su, J. Wu, P. S. Tan, X. Mao, and Y. H. Wang, study of different texture features for document image re-
“Anomaly detection of defects on concrete structures with the trieval,” Expert Systems with Applications, vol. 121, pp. 97–114,
convolutional autoencoder,” Advanced Engineering Infor- 2019.
matics, vol. 45, Article ID 101105, 2020. [50] S. Theodoridis and K. Koutroumbas, Pattern Recognition,
[33] L. Wang, K. i. Kawaguchi, and P. Wang, “Damaged ceiling Academic Press, Printed in the United States of America,
detection and localization in large-span structures using Cambridge, MA, USA, 2009, ISBN 978-1-59749-272-0.
convolutional neural networks,” Automation in Construction, [51] R. M. Haralick, K. Shanmugam, and I. H. Dinstein, “Textural
vol. 116, Article ID 103230, 2020. features for image classification,” IEEE Transactions on Sys-
[34] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning tems, Man, and Cybernetics, vol. SMC-3, no. 6, pp. 610–621,
(Adaptive Computation and Machine Learning Series), The 1973.
MIT Press, Cambridge, MA, USA, 2016, ISBN-10: 026203561. [52] R. M. Haralick and L. G. Shapiro, Computer and Robot Vision,
[35] P. Kim, MatLab Deep Learning with Machine Learning, Addison-Wesley Longman Publishing Co., Inc., Boston, MA,
Neural Networks and Artificial Intelligence, Apress, New York, USA, 1992, ISBN 0201569434.
NY, USA, 2017. [53] G. M. Hadjidemetriou, P. A. Vela, and S. E. Christodoulou,
[36] S. Skansi, Introduction to Deep Learning—from Logical Cal- “Automated pavement patch detection and quantification
culus to Artificial Intelligence, Springer International Pub- using support vector machines,” Journal of Computing in Civil
lishing, New York, NY, USA, 2018. Engineering, vol. 32, Article ID 04017073, 2018.
[37] A. S. Moghaddam, E. R. Azar, Y. Mejias, and H. Bell, “Es- [54] N.-D. Hoang, “Image processing-based pitting corrosion
timating stripping of asphalt coating using k-means clustering detection using metaheuristic optimized multilevel image
and machine learning–based classification,” Journal of thresholding and machine-learning approaches,” Mathe-
Computing in Civil Engineering, vol. 34, Article ID 04019044, matical Problems in Engineering, vol. 2020, Article ID
2020. 6765274, 19 pages, 2020.
[38] L. G. Shapiro and G. C. Stockman, Computer Vision, Prentice- [55] F. Tomita and S. Tsuji, Computer Analysis of Visual Textures,
Hall, Upper Saddle River, NJ, USA, 2001. Springer Science + Business Media, New York, NY, USA,
[39] M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis, 1990, ISBN 978-1-4612-8832-9.
and Machine Vision, Cengage Learning, Printed in the United [56] M. Andrejiova, A. Grincova, and D. Marasova, “Failure
States of America, Boston, MA, USA, 2013, ISBN-10: 1-133- analysis of rubber composites under dynamic impact loading
59360-7. by logistic regression,” Engineering Failure Analysis, vol. 84,
[40] S. Molaei and M. E. Shiri Ahmad Abadi, “Maintaining filter pp. 311–319, 2018.
structure: a Gabor-based convolutional neural network for [57] H. C. Chan, C. C. Chang, P. A. Chen, and J. T. Lee, “Using
image analysis,” Applied Soft Computing, vol. 88, Article ID multinomial logistic regression for prediction of soil depth in
105960, 2020. an area of complex topography in Taiwan,” Catena, vol. 176,
[41] A. Premana, A. P. Wijaya, and M. A. Soeleman, “Image pp. 419–429, 2019.
segmentation using Gabor filter and K-means clustering [58] W. W. Piegorsch, Statistical Data Analytics: Foundations for
method,” in Proceedings of the 2017 International Seminar on Data Mining, Informatics, and Knowledge Discovery, John
Application for Technology of Information and Communica- Wiley & Sons, Hoboken, NJ, USA, 2015, ISBN 978-1-118-
tion (iSemantic), pp. 95–99, Semarang, Indonesia, October 61965-0.
2017. [59] A. Agresti, An Introduction to Categorical Data Analysis, John
[42] Z. Eduardo, G. G. B. Jaime, M. Roberto, and L. José, “Road Wiley & Sons, Hoboken, NJ, USA, 2019, ISBN
crack detection using visual features extracted by gabor 9781119405283.
16 Advances in Civil Engineering

[60] S. J. Russell and P. Norvig, Artificial Intelligence—A Modern

Approach, Pearson Education, Upper Saddle River, NJ, USA,
2010.
[61] M. Gormley, Logistic Regression, https://www.cs.cmu.edu/%
7Emgormley/courses/10701-f16/slides/lecture5.pdf, 2016.
[62] A. Ng, Lecture Notes, Stanford University, Kunnamangalam,
Kerala, 2018, http://cs229.stanford.edu/notes/cs229-notes1.
pdf.
[63] D. P. Kingma and J. Ba, “Adam: a method for stochastic
optimization,” in Proceedings of the 3rd International Con-
ference on Learning Representations (ICLR), San Diego, CA,
USA, May 2015.
[64] J. Brownlee, Gentle Introduction to the Adam Optimization
Algorithm for Deep Learning Machine Learning Mastery,
https://machinelearningmasterycom/adam-optimization-
algorithm-for-deep-learning/, 2017.
[65] Accord, “Accord. NET Framework,” 2019, http://accord-
frameworknet/.
[66] R. M. Haralick, “Some neighborhood operators,” in Real-Time
Parallel Computing: Imaging Analysis, M. Onoe, K. Preston,
and A. Rosenfeld, Eds., Springer, Boston, MA, USA, pp. 11–35,
1981.
[67] D. Tien Bui, N.-D. Hoang, H. Nguyen, and X.-L. Tran, “Spatial
prediction of shallow landslide using Bat algorithm optimized
machine learning approach: a case study in Lang Son Prov-
ince, Vietnam,” Advanced Engineering Informatics, vol. 42,
Article ID 100978, 2019.