Hindawi

Computational Intelligence and Neuroscience

Volume 2022, Article ID 1822585, 14 pages
https://doi.org/10.1155/2022/1822585

Research Article
A Pavement Crack Detection Method Based on Multiscale
Attention and HFS

Chun Li ,1,2,3 Yu Wen,1 Qingxuan Shi,1,2,3 Fang Yang,1,2,3 Hongyan Ma,4

and Xuedong Tian 1,2,3
1
School of Cyber Security and Computer, Hebei University, Baoding 071002, China
2
Hebei Machine Vision Engineering Research Center, Hebei University, Baoding 071002, China
3
Institute of Intelligent Image and Document Information Processing, Hebei University, Baoding 071002, China
4
School of Mathematics and Information Science, Hebei University, Baoding 071002, China

Correspondence should be addressed to Xuedong Tian; [email protected]

Received 15 October 2021; Revised 19 December 2021; Accepted 28 December 2021; Published 27 January 2022

Academic Editor: Suneet Kumar Gupta

Copyright © 2022 Chun Li et al. 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.

To solve the problem of low detection accuracy due to the loss of detailed information when extracting pavement crack features in
traditional U-shaped networks, a pavement crack detection method based on multiscale attention and hesitant fuzzy set (HFS) is
proposed. First, the encoding-decoding structure is used to construct a pavement crack segmentation network, ResNeXt50 is used
to extract features in the encoding stage, and a multiscale feature fusion module (MFF) is designed to obtain multiscale context
information. Second, in the decoding stage, a high-efficiency dual attention module (EDA) is used to enhance the ability of
capturing details of the cracks while suppressing background noise. Finally, the membership degree of the crack is calculated based
on the advantages of the HFS in multiattribute decision-making to obtain the similarity of the crack, and the binary image after
segmentation is judged by the hesitation fuzzy measure. The experiment was conducted on the public road crack dataset Crack500.
In terms of segmentation performance, the evaluation indexes Intersection over Union (IoU), Precision, and Dice coefficients of
the proposed network reached 55.56%, 74.26%, and 67.43%, respectively; in terms of classification performance, for transversal
and longitudinal cracks, the classification accuracy was 84% ± 0.5%, while the block and the alligator were both 78% ± 0.5%. The
experimental results prove that the crack details detected by the proposed method are more abundant, and the image detection
effect of complex topological structures and small cracks are better.

1. Introduction Based on the traditional image processing method, it is

the initial attempt to automatically detect road cracks.
As the lifeblood of the national economy, the quality of Akagic et al. [3] proposed a crack image detection method
pavement operation plays a vital role in the normal progress of based on the Otsu threshold and histogram. Although this
production and life. Due to the pressure and climatic effects of method is efficient, the crack area can be accurately found
heavy loads throughout the year, most of the roads have suffered only when the crack pixel is darker than the surrounding
from varying degrees of disease. Cracks are one of the common pixels. Medina et al. [4] used the wavelet transform method
pavement diseases. If not discovered and treated in time, it will to detect cracks, which not only was susceptible to the
affect the service life of the road surface [1]. Traditional detection contrast between crack pixels and surrounding pixels, but
methods are inefficient and costly, and cracks show problems also could not detect cracks with poor continuity. To im-
such as complex topology, poor continuity, low contrast, and prove the effect of detecting continuous cracks, the mini-
strong noise [2], which brings challenges to the automatic mum path selection method [5] is proposed to detect cracks
detection of road cracks. Therefore, it is necessary to design an from a global perspective, which effectively enhances the
efficient automatic road crack detection method. continuity of fractured cracks. Although the minimum path
2 Computational Intelligence and Neuroscience

selection method performs crack detection from a global pavement cracks and applied a principal component analysis
perspective, its detection performance is still unsatisfactory to classify the detected pavement cracks. The crack types
when dealing with cracks with disordered shapes or low were divided into transversal, longitudinal, cracked cracks.
contrast with surrounding pixels. It can be seen that automatic The accuracy scored higher than 90%. Nevertheless, patch
detection of road cracks is still a difficult task for researchers. classification is only suitable for rougher classification tasks.
In recent years, deep learning has been applied to road Cubero-Fernandez et al. [16] classified the discontinuous
crack detection tasks due to its outstanding feature ex- cracks in an image as a whole, though they did not consider
traction capabilities. Pauly et al. [6] cropped each crack the spatial distribution relationship between the cracks.
image into a patch, and then the patch was classified as crack Existing road crack detection methods enhance the
or noncrack after neural network training. Although this extraction and classification capabilities of crack features
method was very efficient, it produced false detections. To through global context modules, attention mechanisms,
further improve its detection accuracy, semantic segmen- and principal component analysis methods to improve
tation algorithms based on the encoding-decoding archi- detection and classification accuracy. Because of the crack
tecture are widely used. Lau et al. [7] introduced U-Net to image detection, the foreground pixels are relatively small
road crack detection. The network introduced skip con- and have different lengths and widths. If a natural image
nections into the encoding-decoding architecture, which detection method with a large proportion of foreground
helped to preserve rich image details, thereby improving the pixels is used, the effect is often poor, and eventually the
detection accuracy. Although U-Net performs well in the information of the detected cracks will be lost, thereby
field of image segmentation, the crack area of the crack affecting the detection effect. Therefore, this paper pro-
image is much smaller than the background area. Cao et al. poses a road crack detection method based on multiscale
[8] replaced the U-Net encoder with ResNet34 to deal with attention and HFS. The method is mainly divided into two
the loss of spatial information caused by continuous pooling. tasks: semantic segmentation of the crack image [1] to
Effectively avoiding gradient disappearance or gradient realize the separation of the crack area and the noncrack
explosion, Chen et al. [9] embedded a global context module area and the classification of the segmented binary image.
in the U-Net network structure to give the network the For the first task, the proposed solution uses rectangular
ability to capture global context information, which is soft pooling instead of global average pooling, which
conducive to the detailed segmentation of pavement crack effectively extracts long and narrow fracture feature in-
images. Augustauskas and Lipnickas [10] introduced a kind formation; rectangular pooling is used to fuse multiscale
of attention based on the U-shaped network. The force gate feature information to expand the receptive field of the
model suppresses background noise and strengthens the network, so that the small proportion of crack informa-
ability of the network to capture detailed features of cracks. tion in the image is also noticed, thereby improving the
Fan et al. [11] proposed an end-to-end pixel-level road crack accuracy of segmentation, using channel and spatial at-
detection network. By building multiple expansion convo- tention to assign the importance of two-dimensional
lution modules to help the network obtain the multiscale weights, and based on this importance to improve the
context information of the cracks, a hierarchical feature useful information for crack identification and suppress
learning module is designed to integrate low-level features useless information. Different from existing segmentation
and high-level features. The designed multiscale output methods, our proposed solution is more suitable for
feature map has better performance in fracture information segmenting images with unbalanced aspect ratios such as
inference, thereby improving the robustness and universality cracks. For the second task, the core of the classification
of the network. Ali et al. [12] implemented a deep fully algorithm of the proposed solution is to define the number
convolutional neural network based on residual blocks. For of crack branches, the number of inflection points, and the
the extreme imbalance between target and background centroid distance index and use the multiattribute deci-
pixels in crack images, a local weighting factor was proposed sion-making of HFS to calculate the similarity to classify.
to effectively reduce the trouble caused by pixel imbalance to Compared with the existing classification algorithms, our
the network; a crack image dataset with different crack width method provides a more detailed qualitative classification
directions and a location dataset were developed for re- method, which can derive the crack image category from
searchers to use for training, validation, and testing. Fan the comprehensive analysis of multiple indicators.
et al. [13] proposed a road crack automatic detection and The main contributions of this article are as follows:
measurement network based on probability fusion. Through
(1) Based on the encoding-decoding architecture, a
the designed integrated neural network model, satisfactory
multiscale feature fusion module is designed to
crack detection accuracy is obtained; according to the
obtain more receptive fields, so as to improve the
predicted crack map, the width and length of the crack can
network’s ability to recognize disordered cracks.
be measured effectively. Wang et al. [14] proposed a sem-
isupervised semantic segmentation network for crack de- (2) Design an efficient dual attention module to realize
tection. The model extracts multiscale crack feature the information interaction between spatial features
information through Efficient-UNet; it greatly reduces the and channel features, so as to improve the network’s
workload of labeling while maintaining high labeling ac- anti-interference ability and network feature ex-
curacy. Wang et al. [15] used a neural network to detect traction ability.
Computational Intelligence and Neuroscience 3

(3) The segmented binary image is analyzed by the mechanism [19–21], it does not require dimensionality re-
connected domain algorithm, and the advantages of duction operations to capture rich semantic information.
hesitant fuzzy sets in multiattribute decision-making However, in the high-level features of the cracks, it lacks
are used to calculate the fracture multiattribute sufficient spatial information. Inspired by the cascade of
membership degree to obtain the similarity of the channel and spatial attention in Convolutional Block At-
fractures to determine the fracture category. tention Module (CBAM) [22], this paper introduces the
importance of pooling in the ECA attention mechanism [23]
The rest of the paper is structured as follows: the second
to fuse channel weight information and spatial position
part elaborates the fracture segmentation network based on
information, while also designing an efficient dual attention
multiscale attention and the crack classification method
module EDA. Its operating principle is shown in Figure 2,
based on hesitant fuzzy sets; the third part analyzes and
which can make the network better distinguish the im-
discusses the experimental results; the fourth part sum-
portance of crack features and further improve the accuracy
marizes the paper.
of the segmentation network.
The module is divided into upper and lower branches:
2. Materials and Methods
(1) The upper branch is used to capture channel at-
2.1. Fracture Segmentation Network Based on Multiscale tention features. This branch first obtains the global
Attention. The overall structure of our proposed solution receptive field through a global average pooling
Multiscale Attention Crack Segmentation Network operation and then uses 2D convolution to achieve
(MACSNet) is shown in Figure 1. channel interaction, taking each channel and its K �
The network consists of Encoders (E1, E2, E3, E4), 3 neighbouring channels to generate local cross-
Decoders (D1, D2, D3, D4), an EDA module and a MFF channel attention information, as shown in
module. When designing the network structure, considering
K
that the proportion of crack pixels in the image is small, the ⎝􏽘 F · Y ⎞
Fi � σ ⎛ ⎠ Yi,j ∈ Ωi,k ,
j i,j , (1)
network structure should not be too deep, but a certain j�1
degree of accuracy must be ensured. Therefore, the encoder
uses ResNeXt50 as the basic network to extract the char- where Fj represents the one-dimensional convolu-
acteristics of the input crack image. Its essence is grouped tion of size K , Ωi,k represents the set of K adjacent
convolution, and the algorithm performance is improved by channels of the input feature Yi , σ represents the
increasing the number of branches. The encoder in this sigmoid activation function, and Fi represents the
structure retains the first five feature extraction modules of attention information of K adjacent channels to the
ResNeXt50, which are named pooling and E1-E4, respec- current channel.
tively, as shown in Figure 1.
In addition, to obtain multiscale features, a multiscale (2) The next branch is used to capture spatial attention
feature fusion operation is performed after the E4 encoder to features, which trains a weight function on the
better extract the multiscale context information from the original feature map, similar to the attention func-
crack image and optimise the segmentation effect, as well as tion, and then to perform a weighted average with
to incorporate efficient dual attention into the jump con- the original image. Then, the weight function dis-
nection of encoding and decoding. The module allows the tributes the spatial feature weights through the
network to effectively integrate the low-level spatial reso- sigmoid to obtain the importance space attention, as
lution and the high-level semantic information, while fur- shown in
ther paying attention to the area where the crack is located. 􏽐i∈R ai · exp G ai 􏼁􏼁⎞
The module also combines the advantages of subpixel ⎝conv2d⎛
Fs � σ ⎛ ⎝ ⎠⎞
⎠, (2)
convolution and bilinear interpolation in the decoder 􏽐j�R exp􏼐G􏼐aj 􏼑􏼑
D1–D4 and design a parallel feature fusion structure to
sequentially restore image resolution and detailed infor- where ai represents the original feature map, exp makes the
mation. To further integrate spatial resolution and high-level weight value nonnegative and easy to optimise, and G
semantic information, inspired by dense connection [17], D3 represents the weight function after the network obtains G
and D4 are, respectively, upsampled twice. Then, the results through training to enhance specific features.
of the 3 levels are superimposed in a concatenate manner.
2.1.2. Multiscale Feature Fusion Module. ASPP [24] obtains
2.1.1. Efficient Dual Attention Module. The features multiscale fracture feature information through hole con-
extracted in the fracture segmentation network must not volution with different sampling rates and has achieved good
only contain enough spatial information to locate small- results in classification and segmentation tasks. Because the
scale cracks, but also contain rich semantic information to fracture shape is long and narrow, only the adaptive average
effectively distinguish between cracks and other interference pooling is not enough to obtain the global context infor-
information. The Efficient Channel Attention (ECA) [18] mation of the fracture, while the strip pooling [25] can
strengthens the feature propagation ability of the channel obtain the long-distance dependence. Inspired by this, the
dimension. Compared with the classic channel attention Soft-Pool [26] of strip shape is introduced on the basis of
4 Computational Intelligence and Neuroscience

E1 Output
Input EDA
E2

E3 EDA
E4

MFF
D1
D2
EDA D3
D4
EDA

Figure 1: MACSNet framework.

K=3

GAP
S
1×1×C
Fi:W×H×C

2W×H×C
2D Conv

W×H×C Stri- S
pool
W×H×1 Fs:W×H×C

Figure 2: Efficient dual attention module.

ASPP, which effectively increases the global features. This summing all of the weighted activations in the kernel
research combines the advantages of ASPP and Soft-Pool to neighbourhood R, as shown in the following equation:
design a multiscale feature fusion module MFF, which ef-
fectively combines global information and multiscale con- a � 􏽘 Wi ∗ ai , (4)
i∈R
text information, reducing the discontinuity problem in
fracture segmentation, as shown in Figure 3.
The MFF first undergoes a 1 × 1 convolution to reduce
the dimensionality and then obtains a multiscale parallel 2.1.3. Decoder Module. The decoder restores the image
structure through a variety of sampling rates and pooling resolution through an upsampling method. A common
methods. The first three branches of the parallel structure upsampling method is bilinear interpolation, which restores
fuse hole convolutions with different sampling rates to the resolution through the neighbouring pixel values, but the
obtain multiscale information. The latter of the two branches restored boundary is blurry. The introduction of subpixel
obtains global information that is more suitable for the shape convolution in superresolution [27] can make the details of
of the crack through soft pooling of the strip shape. Strip soft the image clearer. Therefore, the decoding block shown in
pooling mainly uses the maximum approximate R in the Figure 4 is designed by combining bilinear interpolation and
activation area. Each activation ai with index i applies a subpixel convolution. The upper branch undergoes general
weight Wi , and the weight is equal to the natural number of operations, such as a 1 × 1 convolution, batch normalisation,
activation values divided by the sum of the natural expo- and ReLU, and then applies bilinear interpolation for
nents of all activation values, as shown in the following upsampling; the lower branch is subpixel convolution. After
equation: the fusion of the two branch features, the detailed infor-
mation of the detected cracks is more complete, and the
eai amount of calculation added is small. The structure of the
Wi � . (3)
􏽐j�R eaj decoder module is shown in Figure 4.
This weight is multiplied by the corresponding activation
value to make a nonlinear transformation together, and the 2.2. Crack Classification Algorithm Based on Hesitant Fuzzy
higher activation has a more obvious impact on the output. Sets. The type of crack is an important index to evaluate the
Since pooling is performed in a high-dimensional feature quality of the pavement. The evaluation of different types of
space, highlighting the maximum activation effect is more cracks directly affects the decision-making of different
reasonable than directly selecting the maximum value. The maintenance strategies. This section extracts the five features
output value of the Soft-Pool operation is obtained by of the number of cracks in the image, the number of
Computational Intelligence and Neuroscience 5

3×3Conv +
rate=6 W×H× (C/4)

3×3Conv +
rate=12 W×H× (C/4)

3×3Conv +
W×H× (C/4)
W×H×C rate=18 W×H× (5×C/4) W×H×C
SoftPool up
W/4×H× (C/4) W×H× (C/4)

SoftPool
up

W× (H/4)× (C/4) W×H× (C/4)

Figure 3: Multiscale feature fusion module.

BatchNorm
Conv 1×1 Interpolate +
Relu

Sub-pixel
Convolution

Figure 4: Decoder module.

inflection points, the average centroid distance, the angle E � 􏼈 < x, hE (x) > |x ∈ X􏼉, (5)
between centroids, and the area of the cracks while using the
decision-making advantages of hesitation fuzzy set theory to where hE (x) represents the set of membership degrees of the
realize crack classification. element x in the set X to the set E. It is a possible mem-
On the basis of Zadeh fuzzy set theory [28], Torra bership set in [0,1]; then, h � hE (x) is called the hesitant
proposed a hesitating fuzzy set [29], which allows the degree fuzzy element. P and Q are two hesitant fuzzy sets about
to which the element belongs to the set to be given in the X � 􏼈x1 , x2 , x3 , . . . , xn 􏼉; then the generalized hesitant fuzzy
form of a set of multiple possible values, in order to ef- distance measure between them is shown in
fectively characterize the uncertainty in decision-making. In
the formula below, X is a nonempty set and call E the
hesitant fuzzy set:

lx 1/λ
1 n ⎡ 1 i 􏼌􏼌􏼌 σ(j) 􏼌􏼌λ
dghn (P, Q) � 􏽘 ⎣ 􏽘 􏼌􏼌hP xi 􏼁 − hQ xi 􏼁􏼌􏼌􏼌 ⎤⎥⎥⎦ , s(P, Q) � 1 − dghn (P, Q),
⎢
⎢ σ(j) (6)
n i lxi j�1

σ(j)
where n is the number of elements in the set X, hP (xi ) between them. This helps to realize the measurement of
σ(j)
and hQ (xi ) are the j-th largest values in hP (xi ) and hQ (xi ), multiattribute similarity.
respectively, s(P, Q) is the corresponding similarity, and λ is After detecting the pixel area containing the crack, the
the control parameter. The generalized hesitant fuzzy dis- connected area algorithm is used to divide the crack into
tance measure gives the distance calculation formula of two independent crack branch targets. After analysis, the
hesitant fuzzy sets under multiple attributes and multiple number, area, centre position coordinates, and approximate
indicators. The smaller the distance, the greater the similarity length and the width of the crack branches are obtained. To
6 Computational Intelligence and Neuroscience

further analyze the spatial distribution relationship between

adjacent cracks, a Minimum Enclosing Rectangle (MER)
[30] is generated for each crack branch target. As shown in
Figure 5, each MER record includes the position coordinates
of the crack branch target, height, and width. Then by
calculating the MER information of the two branch targets, it
is judged that C1 and C2 can be merged into a new branch
target. When two types of crack branches appear in the crack
image, the types are divided according to the pixel weight.

Definition 1. Cq represents the input road crack binary

image and Cqj represents any crack binary image in the
dataset (j � 1, 2, . . . , n, n is the total number of road crack
binary images in the dataset).

Definition 2. NP is the hesitant fuzzy evaluation attribute of Figure 5: Fracture image analysis.
the crack image.
to 0.5, it is M type. Figure 5 shows the average distance
Definition 3. The cracks are divided into four types:
between the judgement and the calculated centroids. When
transversal, longitudinal, block, and alligator, which are
the average distance between the centroids of Cqj and Cq is
divided into T, V, M, and C.
closer, the similarity is higher.

Definition 6. The membership function of the centroid

2.2.1. Hesitating Fuzzy Attribute Index. (1) The Number of distance index is
Branches Index Tnum . By analyzing the number of branch
1
targets in the crack image, the degree of similarity between μTdis � , (9)
the two images Cqj and Cq is evaluated. When the number of Tdis 􏼐Cqj 􏼑+β
1+e
branches is greater than or equal to 3, the probability of
being M or C type is greater. When the number of branches where Tdis (Cqj ) is expressed as Tdis (Cqj ) � (SC1 /
is less than 3, the probability of being T or V type is greater. (SC1 + SC2 + SC3 )) ∗ dis1 + (SC2 /(SC1 + SC2 + SC3 )) ∗ dis2 ;
the centroid distance weight is larger when the crack area
Definition 4. The membership function of the number of accounts for a larger proportion; β is the balance parameter.
branches is (4) Centroid Included Angle Index Tflg . The angle be-
tween the diagonal of MER and the transversal direction
􏼐−len Cq 􏼁/􏼐len􏼐Cqj 􏼑+β􏼑
(7) judges whether it is T type or V type. The diagonal direction
μTnum 􏼒Cq , Cqj 􏼓 � e ,
always faces the centre of the mass and the direction of the
crack. When the included angle is 0°–45° or 135°–180°, the
where len(Cq ) is the number of crack branches in the input crack type is T; when the included angle is 45°–135°, the crack
crack image, len(Cqj ) represents the number of crack type is V, as shown in Figure 6.
branches in any image in the dataset, and β is the balance
parameter.
(2) Inflection Point Index Tang . To further divide the Definition 7. The membership function of the centroid
cracks into M, C type or T, V type, the Freeman code is included angle index is
introduced to reencode the crack image to obtain the
μTflg 􏼒Cq , Cqj 􏼓 � 􏼚􏼒Levflg j , hSflg j 􏼓j � T, V􏼛, (10)
boundary sequence of the crack and the number of sub-
stantial inflection points. When the number of inflection
where Levflg j represents the type of angle between the
points is greater than or equal to 2, the probability of being
diagonal of the MER and the transversal direction and hSflg j
M and C types is high.
represents the degree of membership corresponding to the
angle.
Definition 5. The membership function of the inflection (5) Crack Area Index Ts . According to the ratio of the
point index is crack area in the image to the original area, the importance
1 of crack branches in the image is divided into three types:
μTang � . (8)
Tang 􏼐Cqj 􏼑+β light (s), medium (m), and heavy (l). When there are
1+e multiple types of cracks in an image of a crack, the type of
(3) Centroid Distance Index Tdis . The centroid average cracks in the image is judged according to the importance.
distance (dis) of the crack branches in the image judges
whether it is M type or C type. When dis is less than the Definition 8. The membership function of the fracture area
threshold 0.5, it is C type, and when it is greater than or equal index is
Computational Intelligence and Neuroscience 7

3.1. Crack Segmentation Experiment Analysis

3.1.1. Dataset and Data Augmentation. To evaluate the
performance of this algorithm and other algorithms more
objectively, the public road crack dataset Crack500 [31] is
selected to evaluate the algorithm. The Crack500 dataset is
composed of 500 pavement crack images with an image
resolution of 2000 × 1500 or more. It was obtained on the
main campus of Temple University using smart devices.
Each image is manually marked by an expert to generate
corresponding Ground Truth data (GT).
The Crack500 dataset has 250 images for the training set, 50
images for the validation set, and 200 images for the test set. In
order to avoid overfitting in the training process, the data need
to be augmented. The training set is randomly increased to
2500 images by rotating at 90°, 180°, or 240°. Because of the
Figure 6: Centroid angle. different sizes of the images, random cropping is used to crop
the original image into a size of 256 × 256 pixels.

μTS 􏼒Cqj 􏼓 � 􏼚􏼒TypeS j , hSStype j 􏼓|j � s, m, l􏼛, (11)

3.1.2. Training Strategy. The experiment in this article is
where TypeS j represents the type of fracture importance implemented under the Windows operating system. The
and hSStype j represents the degree of membership corre- deep learning framework is PyTorch, the processor is I5-
sponding to different types. 8500 NVIDIA, and the graphics card is GTX 1660ti. In the
The abovementioned indexes Tnum , Tdis , and Tflg define experiment, Adam is used to optimise the convergence
the classification of pavement diseases according to the process, the initial learning rate is 0.0001, the batch size is 12,
“China Highway Technical Condition Evaluation Standard,” and the epoch is 50. The learning rate strategy uses an
and the indexes Tang and Ts supplement the above- exponential decay method to solve the problem with slow
mentioned indexes. Tnum represents the number of branches convergence in the network’s pretraining through collabo-
in an image, used to determine whether it is a unidirectional rative optimisation of optimisation algorithms and dynamic
crack or a network crack; Tdis represents the centroid dis- learning rates.
tance between multiple branches in an image, which is used
to further judge whether the network crack is an alligator or
3.1.3. Data Preprocessing. Due to the low contrast between
a block crack; Tflg represents the angle between the crack
the cracks in the original image and the surrounding pixels,
and the transversal line, which is used to further judge
and the susceptibility to uneven illumination, there is a lot of
whether a unidirectional crack is a transversal crack or a
noise in the image. If the segmentation is performed directly,
longitudinal crack; Tang is a supplement to Tflg . When a
the discrimination of the same feature will be low, and the
unidirectional crack has multiple inflection points, the
existing noise will affect the subsequent segmentation effect.
weight of the network crack is judged to increase; Ts is a
Zhu et al. [32] proposed a defogging algorithm based on
supplement to the above four indicators. When there are
image fusion. The underexposed image is obtained on the
multiple types of cracks in the image, it is divided into three
basis of gamma correction, and then the weight is refused by
important levels according to the proportion of pixels, and
analyzing the global and local exposure to improve the
finally the categories are defined according to the high
performance and robustness of image dehazing; Zheng et al.
degree of importance.
[33] proposed an image defogging algorithm based on
adaptive structure decomposition and multiexposure image
fusion. On the basis of gamma correction, images with
2.2.2. Crack Image Classification Based on HFS. After
different exposure levels were subjected to adaptive structure
multiple membership evaluations of fracture images Cq and
decomposition. This algorithm effectively eliminates the
Cqj , hesitant fuzzy sets hq and hqj are formed. hq and hqj are
noise caused by haze on the image. Inspired by the above, in
used to calculate the hesitating fuzzy similarity sim(Cq , Cqj ).
order to improve the contrast between the crack and the
When the similarity is greater, the probability that they is a
surrounding pixels, improve the uneven illumination, and
category is greater. The crack image correlation classification
reduce the image noise, the following preprocessing is
Algorithm 1 based on the hesitant fuzzy set is as follows:
performed: (1) grayscale conversion; (2) image standardi-
zation; (3) gamma correction, as shown in Figure 7.
3. Results and Discussion
In this section, we mainly discussed the implementation 3.1.4. Performance Evaluation Index. To better evaluate the
details of the proposed solution and presented experimental performance of the proposed algorithm, this paper uses
results. four indicators of Accuracy, Precision, Dice, and IoU to
8 Computational Intelligence and Neuroscience

Input: crack binarized image Cq

Output: the crack binarized image Cqj that is the same as Cq or within a certain similarity threshold range
(1) The feature database Crackst(id, sim) is initialised.
(2) hq hqj //Store the hesitant fuzzy set of images Cq and Cqj
(3) for j � 1: n // n is the number of crack images in Crackst
(4) Nq � 􏼚μTqnum , μTqang , μTq , μTq , μTqS 􏼛//Hesitating fuzzy attribute index
dis flg

(5) hTq � Nq � 􏼚μTqnum , μTqang , μTq , μTq , μTq 􏼛

dis S flg

(6) hTq � Nqj � 􏼨μTqj , μTqj , μTqj , μTqj , μTqj 􏼩

j num ang dis flg S
l
(7) dghn (hTq , hTq ) � 1/2 􏽐2m�1[1/lPm 􏽐n�1NPm
|hσ(n) σ(n) λ 1/λ
Tq (NPm ) − hTqj (NPm )| ] //Calculate the generalized hesitation fuzzy distance
j
(8) sim(hTq , hTq ) � 1 − dghn (hTq , hTq ) //Similarity of hesitating fuzzy sets
j j
(9) sim(Cq , Cqj ) � sim(hTq , hTq ) //Convert the calculation of the similarity between the crack images to the calculation of the
j
similarity between the set of hesitation modes
(10) Add to table Crackst(id, sim(Cq , Cqj ))
(11) if sim(Cq , Cqj ) > Q then// Q is the similarity threshold set by the user
(12) return image (Cqj ) //Return the crack image that meets the conditions
(13) end if
(14) end for

ALGORITHM 1.

correctly dividing the cracked pixels. Dice denotes the har-

monic mean of the Precision and Recall.

3.1.5. Loss Function. Since the ratio of cracked pixels to

noncracked pixels is highly imbalanced, the effect of cracked
areas on the loss is relatively small, resulting in low splitting
accuracy of cracked pixels. In order to make the network
balance this problem, the focus loss function [34] is introduced,
focusing on the crack area and difficult-to-separate samples.
c
focal loss pt 􏼁 � −α 1 − pt 􏼁 log pt , (13)
Figure 7: Original image and preprocessed image.
where α is used to balance the imbalance of positive and
negative samples between cracks and background areas. If
objectively evaluate the effect of pavement crack seg- there are few samples of the crack category, the weight of
mentation. The calculation formula of the evaluation its loss increases. c reduces the influence of easy-to-
index is classify samples on the network model and pays attention
TP + TN to the training of difficult samples to further improve the
Accuracy � , segmentation accuracy.
TP + TN + FP + FN
Table 1 shows the effect of weight changes on the results
TP on the Crack500 data set. Since α and c have a mutual
Precision � , influence, when selecting their values, when the value of c
TP + FP
(12) increases, the value of α is correspondingly reduced. It can be
2 × Precision × Recall seen from the experimental results that when the value of c is
Dice � ,
Precision + Recall 1, with the increase of α, the four indicators have different
degrees of growth. When the value of c is 2, the effect is
Precision ∩ GT optimal as α increases to 0.25. The above table is the best part
IoU � ,
Precision ∪ GT of the experimental results. With comprehensive analysis of
where true positive examples (TP) and true negative the values of multiple indicators, the final value of c is 2, and
examples (TN) indicate correctly classified cracks and the value of α is 0.25.
noncrack pixels, respectively; false-positive examples (FP)
and false-negative examples (FN) indicate incorrectly
classified cracks and noncrack pixels, respectively; GT is a 3.1.6. Result Analysis
marked image pixel; Accuracy reflects the performance of the
algorithm for correctly dividing cracked and noncracked pixels; (1) Ablation Experiment. To verify the effectiveness of the
and Precision represents the performance of the algorithm for modules in the proposed method, Table 1 shows the effects
Computational Intelligence and Neuroscience 9

of different modules on the segmentation results. The fol- Table 1: The effect of weight value changes on the results.
lowing modules are added to verify the effectiveness of the c α Accuracy Precision Dice IoU
modules based on the U-shaped network with the ResNeXt
- — 97.12 69.81 65.20 54.12
encoder. The training parameters of each network con- 1 0.70 96.21 70.43 63.75 52.69
taining different modules are consistent with the proposed 1 0.75 96.90 70.49 64.12 53.24
network. 1 0.80 97.12 72.06 66.18 53.73
EDA: the high-efficiency dual attention module is added. 1 0.85 97.33 73.89 66.93 54.56
As seen in Table 1, Precision, Dice, and IoU have increased 1 0.90 97.29 73.60 66.87 54.51
by 0.47%, 0.12%, and 0.51%, respectively. Therefore, we can 2 0.10 98.35 71.85 65.50 53.22
conclude that the attention module is effective for pavement 2 0.15 98.48 74.30 66.24 53.90
crack detection tasks. 2 0.20 98.57 74.27 66.91 55.23
MFF: the multiscale feature fusion module is added. As 2 0.25 98.62 74.26 67.43 55.56
seen in Table 1, Precision, Dice, and IoU have increased by 2 0.30 98.61 74.09 67.38 55.51
2.25%, 1.11%, and 1.23%, respectively, which proves the Bold values are the best performing values.
effectiveness of adding the multiscale feature module.
Focal loss: after replacing the cross-entropy loss function information more abundantly, and qualitatively analyze the
with the focal loss function, the experimental results show effectiveness of MACSNet.
that Precision, Dice, and IoU increase by 1.66%, 0.87%, and Quantitative analysis: According to the evaluation in-
0.60%, respectively. dicators in Section 3.1.4, the test results are obtained on the
The results of the ablation experiment are shown in Table 2. public dataset Crack500, as shown in Table 3. Accuracy and
The focus loss function and the multiscale feature fusion Precision alone are not enough to judge the performance of
module improve the network performance most substantially. each algorithm for splitting cracks. At the same time,
The high-efficiency dual attention module contributes to the comprehensive evaluation indices Dice and IoU are used to
improvement of Precision and IoU. Due to the small and evaluate the performance of each algorithm. It can be
complex topological structure of the pavement cracks, the focus seen from Table 2 that MACSNet’s Dice is 2.10%, 3.58%,
loss function improves the segmentation quality of the small 2.17%, and 1.25% higher than U-Net, CE-Net, Deep-
cracks. The multiscale feature fusion module obtains multiscale Labv3, and DeepLabv3+, respectively. IoU is higher
context information to solve the complex topological structure than U-Net, CE-Net, and DeepLabv3, respectively. And
presented by the crack image. The high-efficiency dual at- DeepLabv3+ is 2.34%, 3.97%, 2.22%, and 1.13% higher.
tention mechanism suppresses noise information, such as Therefore, it is verified that the effect of MACSNet is
shadows and scratches through the importance of the channel substantial, which is consistent with the results of the
and space features, effectively enhancing the characteristic qualitative analysis.
ability of the network. The time complexity of the MACSNet algorithm pro-
posed in this paper is shown in Table 3. Frames Per Second
(2) Compared with Existing Algorithms. Qualitative analysis: (FPS) represents how many frames of images the algorithm
To verify the performance of MACSNet in road crack de- can process in one second, and we use FPS to represent the
tection, the algorithm in this paper is compared with other time complexity. Although our method is slower than the
algorithms on the public dataset Crack500, including U-Net general segmentation methods U-Net and CE-Net, it is
[35], CE-Net [21], DeepLabv3 [24], and DeepLabv3+ [36]. faster than the advanced segmentation method Deep-
Their data enhancement and training methods use the Labv3+. To analyze the reason, we have made a compro-
methods described in 3.1 and 3.2. Figures 8(a)–8(g) show mise in time complexity in order to improve the
some of the output results. segmentation accuracy of the algorithm, but our method
The segmentation results of each algorithm can be seen can still reach the real-time standard and has obvious
directly in Figure 8. When the crack topology in the image is advantages in time complexity.
simple, the above five algorithms can segment the cracks
well, as shown in the first row of the above figure. When (3) Compared with Other Advanced Algorithms. To further
there are shadows, scratches, and other noises in the illustrate the effectiveness of MACSNet, MACSNet is com-
background of the image, as in the second and third rows, pared with other advanced methods under the same dataset,
U-Net, DeepLabv3, DeepLabv3+, and CE-Net all have dif- and the results are shown in Table 4. In this paper, the
ferent degrees of crack segmentation discontinuity prob- MACSNet algorithm’s Accuracy, Precision, and IoU values
lems, and MACSNet can segment continuous cracks. The are better than other road crack segmentation algorithms.
reason may be considering the global context information.
When the topological structure of the crack in the image is
complex, such as in Lines 4–6, the missed detections of 3.2. Crack Classification Experiment Analysis
U-Net, DeepLabv3, and CE-Net are more serious. Although
DeepLabv3+ missed a few cracks, as seen in the fifth line, 3.2.1. Dataset Introduction. In order to verify the effec-
it lacks the integrity of the cracks. Contours and tiveness of the crack classification method based on the
MACSNet also add a local importance attention mechanism, hesitant fuzzy set, the dataset in this section selects 948
which can accurately segment small cracks, extract feature images from the dataset in Section 3.1. It is divided into four
10 Computational Intelligence and Neuroscience

Table 2: Test results of each algorithm’s segmentation index on the Crack500 dataset.
EDA MFF Focal loss Precision % Dice % IoU %
✓ 70.35 65.45 53.73
✓ ✓ 72.60 66.56 54.96
✓ ✓ 73.01 66.90 54.89
✓ 71.43 65.36 53.82
✓ ✓ 73.86 67.21 55.01
✓ ✓ ✓ 74.26 67.43 55.56

(a) (b) (c) (d) (e) (f) (g)

Figure 8: Segmentation results of each algorithm on the Crack500 dataset. (a) Image. (b) GT. (c) MACSNet. (d) U-Net. (e) DeepLabv3.
(f ) DeepLabv3+. (g) CE-Net.

Table 3: Test results of each algorithm’s segmentation index on the Crack500 dataset.
Algorithm Accuracy% Precision% Dice% IoU% FPS
U-Net [35] 96.12 69.88 65.33 53.22 47.49
CE-Net [21] 96.89 70.47 63.85 51.59 43.90
DeepLabv3 [24] 96.91 70.41 65.26 53.34 23.12
DeepLabv3+ [36] 96.94 71.06 66.18 54.43 23.80
MACSNet (our) 98.62 74.26 67.43 55.56 30.91
Computational Intelligence and Neuroscience 11

categories: 404 transversal cracks; 276 longitudinal cracks; Table 4: Comparison with other advanced methods on the
57 block cracks; 40 alligator cracks. Crack500 dataset.
Algorithm Accuracy % Precision % IoU %
3.2.2. Evaluation Index. In order to analyze the effectiveness Chen et al. [9] — — 51.40
of the multiattribute fuzzy classification method on crack Augustauskas et al. [10] 98.32 64.47 53.34
images, the Recall (R) and Precision (P) are selected to Cao et al. [8] N.A 68.05 54.92
evaluate the image classification results. MACSNet (our) 98.62 74.26 55.56

Definition 9. Recall (R) indicates the percentage of the

number of images FS that are similar to the input image in Table 5: Experimental results under different thresholds.
the classification results and the number of all similar images T (%) FS NS AS P (%) R (%)
in the dataset AS.
90 215 251 276 85.66 77.90
FS 85 218 259 276 84.17 78.99
R� ∗ 100%. (14) 80 222 264 276 84.09 80.43
AS
75 225 269 276 83.64 81.52

Definition 10. Precision (P), which represents the per-

centage of the number of images FS that are similar to the
alligator cracks reached 77.98% and 78.38%, respectively.
input image in the classification results to the number of
After the analysis, the massive and cracked cracks are af-
images NS in the total classification results.
fected by the complex geometric structure, as well as the
FS threshold of the average distance between the particles. This
P� ∗ 100%. (15)
NS paper proposes a crack type classification algorithm based on
hesitant fuzzy sets, which can handle the situation where
there are multiple crack targets in an image.
3.2.3. Result Analysis
(2) Compared with Other Advanced Algorithms. In this
(1) Classification Result Analysis. Definitions 9 and 10 prove section, the method in this paper is compared with existing
that FS represents the number of images retrieved when the methods using the Crack500 data set. Safaei et al. [37]
similarity to the image to be retrieved is higher than the proposed a tile-based cracking method, applying local
threshold T. AS represents the number of all similar images threshold technology to each tile. According to the spatial
in the dataset, and NS represents the total number of re- distribution characteristics of crack pixels, the tiles con-
trieved images. For example, there are 276 longitudinal taining cracks are detected, and after fitting a curve, classify
cracks in the 1124 dataset, and the corresponding P and R longitudinal and transverse cracks by setting the slope
values are calculated according to the Precision and Recall threshold, hereinafter referred to as method 1; Song et al.
calculation rules. [30] proposed a crack classification method based on a
It can be concluded from Table 5 that, under different characterization algorithm. Through the connected area
thresholds, the Precision and Recall of the method in this labeling algorithm and the spatial distribution of the cracks,
paper are maintained at approximately 84% and 80%, the cracks are divided into four types: transversal, longi-
respectively. With the decrease of the threshold, the tudinal, block, and alligator. This method effectively detects
number of images similar to the input image in the search the crack information in a complex environment, herein-
result and the number of all search result images are in- after referred to as method 2.
creasing, but the former increases more slowly than the The classification results of this research method and
latter, which leads to a decrease in the P-value and an other methods are shown in Table 7. The P and R values of
increase in the R-value. After the analysis, when the method 1 are similar to the method in this paper, but un-
threshold is lowered, the massive cracks may be classified fortunately, due to the imperfect algorithm, it is impossible
as longitudinal cracks, which causes the P-value of the to distinguish between massive cracks and cracked cracks.
algorithm in this paper to decrease. Compared with method 2, the method in this paper only has
To further verify the effectiveness of the algorithm in this a slightly lower R-value for transverse fractures, and other
paper, the difference between the Recall and the Precision of indicators are higher than those of method 2. A closer look
the four categories is obtained through the experiments. will reveal that the difference between method 2 and the text
When there are different types of cracks in a picture, a larger method in P-value is less than 1%–3%, but the R-value is less
proportion of the cracks prevail. The experimental results are than 1%. Analyze the reason: our method adds hesitation on
shown in Table 6. the basis of the connected domain labeling algorithm. Fuzzy
Judging from the crack classification results in Table 6, feature attributes have a large increase in Precision to a
the classification algorithm proposed in this paper can ba- certain extent, but at this time, some fuzzy categories may be
sically realize the classification of four types of cracks. The introduced to cause a low increase in Recall. The above
Precision rates of transverse cracks and longitudinal cracks experimental results show that the method in this paper has
reached 83.85% and 84.09%, respectively, and block and a better classification effect than the existing methods.
12 Computational Intelligence and Neuroscience

Table 6: Experimental results of fracture classification based on hesitant fuzzy sets.

Crack category FS NS AS P (%) R (%)
T 322 384 404 83.85 79.70
V 222 264 276 84.09 80.43
M 170 218 228 77.98 74.56
C 29 37 40 78.38 72.50
Mean 81.08 76.80

Table 7: Performance comparison with the other two methods.

T V M C
Category FPS
P% R% P% R% P% R% P% R%
Method 1 83.40 80.02 83.40 80.02 — — — — 0.50
Method 2 82.91 80.77 83.94 79.96 75.61 74.23 74.95 72.19 2.13
Our 83.85 79.70 84.09 80.43 77.98 74.56 78.38 72.50 2.04

The time complexity of the hesitant fuzzy set classifi- The follow-up work is mainly carried out from the
cation algorithm proposed in this paper is shown in Table 6. following two aspects: The network training time cost is
It can be seen from the table that our method is better than considered and the lightweight semantic segmentation
method 1 and is slightly lower than method 2. To analyze the network is introduced into the multiscale attention seg-
reason, we introduced the hesitant fuzzy attribute on the mentation network to achieve faster and accurate binary
basis of the connected domain labeling algorithm, which images. According to the complex topology of the cracks, it
increased the calculation degree of the algorithm and caused is necessary to improve the attribute index and optimise the
our time complexity to be slightly lower than that of method classification method of crack images, especially for classi-
2. But our method has a better balance between accuracy and fying massive cracks and cracked cracks.
time complexity and has obvious advantages compared to
methods 1 and 2. Data Availability
4. Conclusion The data that support the findings of this study are available
upon request.
In the current crack detection methods, most of them only
segment the crack images and do not involve classification, Conflicts of Interest
but the type of cracks is very important to the evaluation of
the road health status, so we propose multiscale attention The authors declare that they have no conflicts of interest.
and HFS crack detection and classification method. This
method distributes the weights of the two dimensions of
Acknowledgments
channel and space through cross-channel attention and local
importance pooling, so that the network automatically pays This work is supported by the Natural Science Foundation of
more attention to the characteristic information of the crack Hebei Province, China (Grant Nos. F2019201329 and
area and further improves the detection accuracy of the F2019201451), and the Science and Technology Project of
crack. A multiscale feature fusion module is designed to fuse Hebei Education Department, China (Grant Nos. ZD2019131
multiscale context information, and the rectangular pooling and QN2018214).
method is used instead of average pooling to retain im-
portant fracture information. This detection method is more
suitable for the detection of crack images with unbalanced
References
aspect ratios than existing methods. On this basis, a road [1] Z. Liu, Y. Cao, Y. Wang, and W. Wang, “Computer vision-
crack image classification method based on HFS is designed. based concrete crack detection using U-net fully convolu-
On the basis of the connected domain algorithm, using the tional networks,” Automation in Construction, vol. 104,
advantages of HFS in multiattribute decision-making, the pp. 129–139, 2019.
membership degree of the cracks is calculated, and the [2] W. Song, G. Jia, H. Zhu, D. Jia, and L. Gao, “Automated
similarity of the cracks is obtained for classification pavement crack damage detection using deep multi-scale
judgement. This classification method uses fuzzy multi- convolutional features,” Journal of Advanced Transportation,
vol. 2020, Article ID 6412562, 11 pages, 2020.
attribute special features to further improve classification
[3] A. Akagic, E. Buza, S. Omanovic, and A. Karabegovic,
accuracy. Through comparative experiments, the effective- “Pavement crack detection using Otsu thresholding for image
ness of the above methods is verified. Experimental results segmentation,” in Proceedings of the 41st International Con-
show that this method has good crack detection and clas- vention on Information and Communication Technology,
sification effects, and it has a certain auxiliary effect on the Electronics and Microelectronics (MIPRO), pp. 1092–1097,
evaluation of road health. IEEE, Opatija, Croatia, May 2018.
Computational Intelligence and Neuroscience 13

[4] R. Medina, J. Llamas, E. Zalama, and J. Gómez-Garcı́a-Bermejo, Vision and Pattern Recognition, pp. 7132–7141, Salt Lake City,
“Enhanced automatic detection of road surface cracks by UT, USA, June 2018.
combining 2D/3D image processing techniques,” in Proceedings [20] X. Li, W. Wang, X. Hu, and J. Yang, “Selective kernel net-
of the 2014 IEEE International Conference on Image Processing works,” in Proceedings of the IEEE/CVF Conference on
(ICIP), pp. 778–782, IEEE, Paris, France, October 2014. Computer Vision and Pattern Recognition (CVPR), pp. 510–
[5] R. Amhaz, S. Chambon, J. Idier, and V. Baltazart, “A new 519, California, USA, June 2019.
minimal path selection algorithm for automatic crack de- [21] Z. Gu, J. Cheng, H. Fu et al., “Ce-net: context encoder network
tection on pavement images,” in Proceedings of the 2014 IEEE for 2d medical image segmentation,” IEEE Transactions on
International Conference on Image Processing (ICIP), Medical Imaging, vol. 38, no. 10, pp. 2281–2292, 2019.
pp. 788–792, IEEE, Paris, France, October2014. [22] S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, “Cbam: con-
[6] L. Pauly, D. Hogg, R. Fuentes, H. David, and R. Fuentes, volutional block attention module,” in Proceedings of the
“Deeper networks for pavement crack detection,” in Pro- European Conference on Computer Vision (ECCV), pp. 3–19,
ceedings of the 34th ISARC, pp. 479–485, IAARC, Taipei, Munich, Germany, July 2018.
Taiwan, June 2017. [23] Z. Gao, L. Wang, and G. Wu, “Lip: local importance-based
[7] S. L. H. Lau, E. K. P. Chong, X. Yang, and X. Wang, “Au- pooling,” in Proceedings of the IEEE/CVF International
tomated pavement crack segmentation using u-net-based Conference on Computer Vision (ICCV), pp. 3355–3364,
convolutional neural network,” IEEE Access, vol. 8, Article ID Seoul, Korea, November 2019.
114899, 2020. [24] L. C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Re-
[8] Y. H. Cao, G. T. Yang, and X. Y. Yang, “Deep learning thinking atrous convolution for semantic image segmenta-
pavement crack detection based on attention mechanism,” tion,” 2017, https://arxiv.org/abs/1706.05587.
Journal of Computer-Aided Design & Computer Graphics, [25] Q. Hou, L. Zhang, M. M. Cheng, and J. Feng, “Strip pooling:
vol. 32, no. 8, pp. 1324–1333, 2020. rethinking spatial pooling for scene parsing,” in Proceedings of
[9] J. Chen, G. Liu, and X. Chen, “Road crack image segmentation the IEEE/CVF Conference on Computer Vision and Pattern
using global context U-net,” in Proceedings of the 2019 3rd Recognition (CVPR), pp. 4003–4012, Seattle, Washington,
International Conference on Computer Science and Artificial USA, June 2020.
Intelligence, pp. 181–185, Association for Computing Ma- [26] A. Stergiou, R. Poppe, and G. Kalliatakis, “Refining Activation
chinery, New York, USA, December2019. Downsampling with Softpool,” 2021, https://arxiv.org/abs/
[10] R. Augustauskas and A. Lipnickas, “Improved pixel-level
2101.00440.
pavement-defect segmentation using a deep autoencoder,” [27] W. Shi, J. Caballero, F. Huszár et al., “Real-time single image
Sensors, vol. 20, no. 9, 2020.
and video super-resolution using an efficient sub-pixel con-
[11] Z. Fan, C. Li, Y. Chen et al., “Automatic crack detection on
volutional neural network,” in Proceedings of the IEEE
road pavements using encoder-decoder architecture,” Mate-
Conference on Computer Vision and Pattern Recognition,
rials, vol. 13, no. 13, 2020.
pp. 1874–1883, New York, USA, September2016.
[12] R. Ali, J. H. Chuah, M. S. A. Talip, N. Mokhtar, and
[28] L. A. Zadeh, “Fuzzy sets,” Information and Control, vol. 8,
M. A. Shoaib, “Automatic pixel-level crack segmentation in
no. 1, pp. 338–353, 1965.
images using fully convolutional neural network based on
[29] V. Torra, “Hesitant fuzzy sets,” International Journal of In-
residual blocks and pixel local weights,” Engineering Appli-
telligent Systems, vol. 25, no. 6, pp. 529–539, 2010.
cations of Artificial Intelligence, vol. 104, Article ID 104391,
[30] W. Song, G. Jia, D. Jia, and H. Zhu, “Automatic pavement
2021.
crack detection and classification using multiscale feature
[13] Z. Fan, C. Li, Y. Chen et al., “Ensemble of deep convolutional
neural networks for automatic pavement crack detection and attention network,” IEEE Access, vol. 7, Article ID 171012,
measurement,” Coatings, vol. 10, no. 2, p. 152, 2020. 2019.
[14] W. Wang and C. Su, “Semi-supervised semantic segmentation [31] L. Zhang, F. Yang, Y. D. Zhang, and Y. J. Zhu, “Road crack
network for surface crack detection,” Automation in Con- detection using deep convolutional neural network,” in
struction, vol. 128, Article ID 103786, 2021. Proceedings of the 2016 IEEE International Conference on
[15] X. Wang and Z. Hu, “Grid-based Pavement Crack Analysis Image Processing (ICIP), pp. 3708–3712, IEEE, Phoenix, AZ,
Using Deep Learning,” in Proceedings of the 2017 4th Inter- USA, September 2016.
national Conference on Transportation Information and Safety [32] Z. Zhu, H. Wei, G. Hu, Y. Li, G. Qi, and N. Mazur, “A novel
(ICTIS), pp. 917–924, IEEE, Alberta, Canada, August 2017. fast single image dehazing algorithm based on artificial
[16] A. Cubero-Fernandez, F. J. Rodriguez-Lozano, R. Villatoro, multiexposure image fusion,” IEEE Transactions on Instru-
J. Olivares, and J. M. Palomares, “Efficient pavement crack mentation and Measurement, vol. 70, pp. 1–23, 2020.
detection and classification,” EURASIP Journal on Image and [33] M. Zheng, G. Qi, Z. Zhu, Y. Li, H. Wei, and Y. Liu, “Image
Video Processing, vol. 2017, no. 1, pp. 1–11, 2017. dehazing by an artificial image fusion method based on
[17] H. Huang, L. Lin, R. Tong et al., “Unet 3+: a full-scale con- adaptive structure decomposition,” IEEE Sensors Journal,
nected unet for medical image segmentation,” in Proceedings vol. 20, no. 14, pp. 8062–8072, 2020.
of the IEEE International Conference on Acoustics, Speech and [34] T. Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal
Signal Processing (ICASSP), May 2020. loss for dense object detection,” in Proceedings of the IEEE
[18] Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca-net: international conference on computer vision, pp. 2980–2988,
efficient channel attention for deep convolutional neural Venice, Italy, October 2017.
networks,” in Proceedings of the IEEE/CVF Conference on [35] O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolu-
Computer Vision and Pattern Recognition (CVPR), IEEE, tional networks for biomedical image segmentation,” in
Seattle, WA, USA, June 2020. Proceedings of the International Conference on Medical Image
[19] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation net- Computing and Computer-Assisted Intervention, pp. 234–241,
works,” in Proceedings of the IEEE Conference on Computer Springer, Munich, Germany, October 2015.
14 Computational Intelligence and Neuroscience

[36] L. C. Chen, Y. Zhu, G. Papandreou, F. Schrof, and H. Adam,

“Encoder-decoder with atrous separable convolution for se-
mantic image segmentation,” in Proceedings of the European
Conference on Computer Vision (ECCV), pp. 801–818,
Munich, Germany, February 2018.
[37] N. Safaei, O. Smadi, A. Masoud, and B. Safaei, “An automatic
image processing algorithm based on crack pixel density for
pavement crack detection and classification,” International
Journal of Pavement Research and Technology, vol. 15, pp. 1–
14, 2021.