Proceedings of the 40th Chinese Control Conference

July 26-28, 2021, Shanghai, China

A Review of Image-Based Pavement Crack Detection Algorithms

Junxian Wu1,2 , Yujia Zhang1 , Xiaoguang Zhao1
1. Institute of Automation, Chinese Academy of Sciences, Beijing 100190, P. R. China
E-mail: [email protected]
2. School of Aritificial Intelligence, University of Chinese Academy of Sciences, Beijing 100190, P. R. China

Abstract: Regular road surface inspection and maintenance on highways, bridges, tunnels and other transportation infrastructure
are of great significance. As one of the common pavement diseases, cracks can pose a potential threat to the safety of bridges,
tunnels, airports, etc., and may cause severe damage to pavements if they are not properly repaired in time. Therefore, crack
detection and segmentation as well as repairing are essential to maintain the overall condition of different pavements. This paper
aims to discuss the development of image-based automatic pavement crack detection algorithms as well the remaining challenges.
Firstly, traditional methods that are based on spatial-domain, frequency-domain, etc. are described. Then the methods using deep
learning techniques are introduced. Datasets and metrics for crack detection are introduced next. Finally, some remaining
problems and the future research directions are presented.

Key Words: Crack detection, Semantic segmentation, Pavement inspection

1 Introduction also presents the characteristics of the narrow operation time

limit and wide operation range. Therefore, a more efficient,
Transportation infrastructure construction is in a period
stable and intelligent image-based automatic road crack de-
of rapid development, which provides convenience for peo-
tection technology is needed for achieving.
ple’s daily life. With the sharply increase of traffic flow, road
Since the 1970s, domestic and foreign researchers have
maintenance becomes more and more significant. In [1] and
conducted in-depth studies on image-based automatic de-
[2], both the authors pointed out that in order to improve the
tection algorithms for pavement cracks. Traditional detec-
maintenance level of the highway pavement and ensure that
tion algorithms mostly utilize the morphological analysis in
the road is in good technical condition, it is crucial to carry
the spatial domain [4], edge detection methods [5], and the
out regular disease investigation and maintenance. In addi-
wavelet transform in the frequency domain [6] in the fre-
tion, in terms of airports, the Capital Airport in Shunyi has
quency domain[6]. Recently, they have been applied into
an average flight taking off and landing in less than 50 sec-
an automated crack detection system for airport pavement
onds. Its total annual passenger throughput has been the sec-
where mobile robots are designed to automatically collect
ond in the world for many consecutive years, second only to
airport pavement crack images [7]. In recent years, with the
Atlanta Airport in the United States [3]. China is in the crit-
wide application of deep learning networks (convolutional
ical stage of transforming from a “large civil aviation coun-
networks [8], fully convolutional networks [9], etc.) in dif-
try” to a “strong civil aviation country”. With the rapid in-
ferent fields, deep learning methods also have good applica-
crease in aircraft takeoffs and landings, the utilization rate of
tions for pavement crack detection [10–12]. However, due to
many airport runways is close to saturation, and road surface
the diverse orientations and sizes of cracks, and other envi-
safety inspection has become the top priority of civil avia-
ronmental factors such as the uneven illumination and stains,
tion operation safety. Bridges, tunnels, etc. are important
it is still a great challenge to achieve accurate pavement crack
components of transportation infrastructure, and the regular
segmentation.
inspection and maintenance of their surfaces are also of great
In this paper, the traditional methods and deep learning-
significance.
based methods for pavement crack detection are reviewed,
As a common pavement disease, cracks on the surface of
datasets and metrics for alogorithms are introduced, and the
concrete and asphalt pose a potential threat to the safety of
prospect of the research is put forward in the end.
bridges, tunnels, and airport pavements. If they are not be-
ing repaired in time, the large-scale damage caused will de- 2 Traditional pavement crack detection algorithm
stroy the pavement structure. Therefore, it is of great sig- The traditional pavement crack detection algorithms
nificance to maintain good road conditions by the regular based on image processing can be divided into the follow-
inspection and repair of cracks. Initial cracks tend to be long ing three categories: 1) spatial domain analysis (including
and thin, which require lots of manpower and material re- morphological methods, edge detection methods, regional
sources for traditional manual or semi-automatic detection. growth methods, etc.), 2) frequency domain analysis (such
And the manual crack detection method completely depends as using wavelet transformation, etc.), and 3) some other de-
on the knowledge and experience of experts which may be tection methods which will be illustrated respectively.
easily affected by different environmental factors and be lack
of objectivity in quantitative analysis. In addition, in order 2.1 Crack detection based on spatial domain analysis
not to affect the normal operation of traffic, the detection Methods based on threshold and morphological analysis
operation has high safety and precision requirements, and usually achieve crack detection by modeling the connectiv-
ity of crack pixels. For example, Amhaz et al. [14] calcu-
This work was supported by the National Key Research and Develop-
ment Project of China (2019YFB1310601), the National Key R&D Program lated the shortest path based on the Dijkstra algorithm [15] to
of China (2017YFC0820203). extract candidate crack areas, and then used morphological

7300
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1: Image processing flow for airport pavement crack detection [13].

post-processing methods to set thresholds to remove small strong consistency, and then follow certain rules to merge
isolated areas and correct false positive pixels. Through into larger regions to achieve segmentation results.
detecting the width of cracks, the black background pixels Zhang et al. [19] extracted the region for rough locations
in the neighborhood of cracks are absorbed as crack pix- of cracks by fusing spatial distribution, intensity and geo-
els for optimizing the result. Peng et al. [13] employed metric features, and then performed post-processing opera-
the secondary threshold segmentation algorithm for the pre- tions. They extracted regions with sufficient confidence as
processed crack image of the airport pavements containing seeds by defining a confidence factor, and finally the grow-
runway markings, which is used to remove the runway mark- ing and merging operations of the region growth algorithm
ings and detect the crack areas. Morphological methods such were used to merge highly similar regions to ensure the in-
as expansion and corrosion are further used for denoising, tegrity of the detected cracks while removing regions with
so as to effectively detect cracks in the image containing low similarity. Li et al. [22] proposed a seed growth algo-
the runway markings. The specific image processing flow rithm to detect cracks at multiple scales. Their method can
is shown in Fig. 1. The disadvantage of crack detection al- obtain a good segmentation result of cracks by matching and
gorithms based on threshold segmentation and morphology fusing the crack regions at different scales. The crack de-
is that the results depend heavily on the selection of differ- tection process of each scale is shown in Fig. 3. Given the
ent parameters [16]. Therefore, additional manual parameter original image (a), it first extracts crack seeds based on in-
adjustments are required with the change of data. formation such as pixel intensity (b), and then performs clus-
Cracks have strong edge characteristics, that is, the gray tering and filtering operations to remove outliers (c). After
level of the crack part has an obvious step response (as in that, the neighboring seeds are connected (d) and further op-
Fig. 2), but the gray level of the background is usually with timization is achieved based on the threshold segmentation
a slow change and a small gradient [17]. Therefore, edge method (e). Finally, the region growth method (f) is used to
detection algorithms are also widely used to address the task detect crack areas at the current scale.
of crack detection. For example, Lim et al. [18] first used the
2.2 Crack detection based on frequency domain analy-
Laplacian of Gaussian operator to highlight the areas where
sis
the pixel intensity value changes rapidly. They then filtered
out noise to achieve edge detection of cracks. However, the Crack detection algorithms based on frequency domain
main drawback of these algorithms is that they can only be analysis are also widely applied in crack detection. These
applied into a set of disjoint crack segments of the images methods usually use the image transformations based on fre-
with high image contrast. These crack detection methods quency domain analysis to process images, and then trans-
may fail when the difference between the crack area and form the extracted information to the space domain. For
the non-crack area is not obvious in the images that contain frequency domain filtering, it mainly includes high-low-
much noise. pass filtering and wavelet processing. The low-pass filtering
Because the cracks are continuous and have obvious di- can smooth high-frequency gray changes and remove high-
rectivity, many researchers proposed to use the method based frequency random noises, however, this may lead to crack
on regional growth for crack segmentation [19–21]. Gener- areas being more blurred. Therefore, in order to enhance the
ally, there are two ways to achieve regional growth: one is contrast between cracks and backgrounds, high-pass filter-
to first give a small block of the target to be segmented in ing methods can be utilized for pavement crack detection to
the image as the seed region, and then to continuously add preserve the areas where the gray level of the image changes
the surrounding pixels according to a certain rule; the other significantly.
is to first divide the image into multiple small areas with Compared with the high-low-pass filtering, wavelet trans-
form is more effective in image processing and can better
suppress noise. Wu and Liu [24] proposed a crack seg-
mentation model based on wavelet transform, which can
achieve robust and efficient crack segmentation by removing
noise and enhancing the edge information of cracks through
wavelet decomposition coefficients. However, wavelet trans-
form cannot well handle the linear singular features in two-
dimensional signals. To address this, Shu et al. [25] pro-
posed an algorithm to apply Contourlet transform for crack
detection, and Wei et al. [26] utilized Beamlet transform for
higher dimensional singularity analysis to realize a better
performance on crack detection.
Fig. 2: The edge feature map of pavement cracks [17].

7301
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3: Li et al. [22] used the area growth algorithm for the crack detection process of each scale.

3 Pavement crack detection algorithms based on curately detect cracks, semantic segmentation algorithm is
deep learning widely used in the pavement crack detection. Semantic seg-
mentation refers to assigning category labels to each pixel
Deep learning has been successfully applied into the fields in the image to achieve pixel-wise classification. The full
of pattern recognition, signal processing, etc., and great convolutional network (FCN) with end-to-end structure [9]
progress has been made in the application of crack segmen- is currently one of the most widely used basic networks in
tation [27–29]. The advantages of deep learning methods the field of semantic segmentation. The FCN uses a con-
include: 1) there is no need to manually design feature rep- volutional layer to replace the last fully connected layer in
resentations to distinguish between cracks and backgrounds; the traditional convolutional network, which can effectively
2) the deep network can implicitly integrate complex illumi- convert image pixels into pixel categories and achieve pixel-
nation changes and irregular changes in the shape of cracks level classification of cracks [28, 32]. For example, Yang et
from the training data; 3) the network can obtain better data al. [28] applied the FCN into crack segmentation, and trained
representations through multi-linear layers. Thus deep learn- the network to segment and predict cracks at of different
ing methods usually have better performances than tradi- scales by taking images of multiple types of cracks as input.
tional image processing technology. Dung and Anh [33] designed a crack segmentation network
based on the fully convolutional network structure, and used
3.1 Crack detection based on CNN different pre-trained networks as the skeleton network in the
Some patch-based crack detection algorithms have been encoder to further compare and analyze the performance of
proposed in the early period. The input of these algorithms different segmentation models.
is the image patches extracted from pictures, and the con- While FCN can obtain pixel-wise segmentation results, it
volutional neural network(CNN) is used to classify patches has the disadvantage that the model tends to ignore the rela-
and identify whether they contain cracks. Zhang et al. [30] tionship between pixels during training, which may lead to
first employed the CNN to achieve image patch-wise crack inferior segmentation results. To solve this problem, Pan et
detection. Doulamis et al. [31] utilized the CNN to classify al. [34] further introduced an attention mechanism to bet-
the tunnel image patches, and then adopted fuzzy spectral ter distinguish pixel-level cracks and non-crack images by
clustering to process the binary crack images to remove the fusing multi-scale context features. The overall model ar-
linear seams of the tunnel. Fan et al. [23] proposed a novel chitecture is shown in Fig. 5. The proposed three attention
CNN architecture based on image patch (as in Fig. 4) to modules work in parallel. Specifically, the feature pyramid
achieve pixel-wise classification of pavement images, which attention module is used to fully extract multi-scale feature
improves the accuracy of crack detection. In their method, representations. The spatial and channel attention modules
the input is 5×5 image patches, and the output is the classi- are used to capture the global spatial and channel dependen-
fication result of whether each pixel in the patch is a crack. cies. After that, different kinds of features are finally merged
for producing the crack segmentation results.
3.2 Crack detection based on FCN
Continuous convolution operations may lead to the loss of
Some crack labels in the dataset may be only a few pix- local information and decrease the segmentation accuracy.
els wide, making it difficult to segment. In order to ac-

Fig. 4: The novel convolutional neural network(CNN) architecture based on image patch proposed by Fan et al. [23].

7302
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 3.4 Crack detection based on other deep learning net-
works
There are other deep learning network structures that have
also been applied in the field of crack segmentation and
achieved good performance. As it is difficult to ensure uni-
form lighting conditions due to environmental changes when
capturing images, Schmugge et al.[40] proposed a crack de-
tection method with fusion analysis of multiple video frames
to tackle the above problem. They first used the SegNet[41]
with an encoder-decoder structure to process each video
frame for predicting the confidence score of each pixel. Then
they merged multiple video frame features of different light-
ing conditions to obtain the final segmentation results. This
can reduce the false positive prediction rate and improve seg-
mentation prediction accuracy.
The denoising method in deep learning can automatically
learn the difference of the statistical characteristics between
Fig. 5: The diagram of the crack segmentation network the signal and the background noise, which helps to enhance
structure that combines full convolutional network and at- the signal and eliminate noise. Wang et al.[42] proposed
tention mechanism [34]. an improved regularized least squares generative adversar-
ial network based on the idea of the generative adversarial
network(GAN)[43]. This helps to alleviate the phenomenon
of pattern collapse in the generative adversarial network and
effectively reduces the image noise and retains more details
of the crack information. Zhang et al.[44] found that the
To address this, Liu and Yao et al.[35] proposed a network crack labels of images in the crack segmentation dataset was
structure that combines the extended FCN and the Deeply too small to detect, leading to the all-black phenomenon in
Supervised Nets(DSN) to reduce information loss, and re- the output after network training. In order to eliminate all-
fined the final prediction results by applying the Conditional black outputs, they presented a crack-patch-only (CPO) su-
Random Field(CRFs) method. Xie et al. [36] proposed that pervised generated adversarial network to generate correct
the frequency domain information extracted by the wavelet crack labels and to encourage the network to retain the abil-
channel can compensate the loss of local image information ity of detecting cracks. Due to the lack of data acquisition
to a certain extent, and developed a dual-channel network and annotation of crack images, transfer learning has also
integrating the FCN and wavelet transform. been applied into crack segmentation [33, 45, 46]. Yang et
3.3 Crack detection based on U-Net al. [45] transfered the accumulated knowledge of the deep
U-Net is another basic network that is widely used in convolution neural network(DCNN) in other domains to de-
crack detection [37, 38]. In U-Net network structure, the tect crack. They proposed a new network framework to real-
residual connection is utilized between the encoder and the ize precise crack segmentation, which adds fully connected
decoder. In this way, the decoder can share the shallow layers after pretrained VGG model for learning the knowl-
feature maps that are obtained in the encoder to compen- edge of crack detection. Dung et al. [46] proposed a deep
sate for the spatial information loss during down-sampling convolutional neural network based on VGG16 and utilized
[38]. Based on this, Mei et al. [29] proposed a densely con- transfer learning method to fine-tune the top convolutional
nected network structure. Their method can achieve higher layer of VGG16 to achieve good results for crack detection.
accuracy by fully encoding low-level feature maps with rich In addition to that, some researchers also introduced ob-
detailed information and high-level feature maps with ab- ject detection algorithms such as Faster R-CNN [27] (as in
stract semantic information. Konig et al. [39] developed a Fig. 7), YOLOv4 [47], etc. for the first-stage crack detec-
fully convolutional U-Net which uses residual connections tion. Then, the second-stage pixel-level segmentation is per-
inside the convolution module. And they further introduced formed on the extracted region of interest containing cracks.
an attention-based gating mechanism between the encoder This two-stage method can improve the network’s ability to
and the decoder to activate only the crack-related features to segment fuzzy cracks.
achieve more precise crack segmentation.
4 Datasets and metrics for crack detection algo-
Since only the final prediction result of U-Net is used to
rithms
calculate the segmentation loss, the network tends to be dif-
ficult to converge, and the segmentation results are less sat- With the rapid development of pavement crack detec-
isfied. To address this, Zou et al.[12] proposed a novel net- tion, many crack data sets have been proposed. Here we
work, DeepCrack. As shown in Fig.6, it connects the feature introduce the 6 most commonly used public datasets for
maps of each convolution stage in the encoder and decoder, crack detection, CrackTree200 [48], CFD [49], Aigle-RN
and calculates the pixel-wise prediction loss at each stage. [14], GAPs364 [50], G45 [51] and DeepCrack [35]. Crack-
The multi-scale feature maps are spliced and merged to gen- Tree200, CFD and Aigle-RN are the crack datasets pro-
erate the final segmentation map. posed earlier, and they contain fewer images than the lat-

7303
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 Fig. 6: Schematic diagram of the structure of DeepCrack proposed by Zou et al. [12].

Fig. 7: The two-stage strategy based on Faster-RCNN [27].

est datasets. CrackTree200 is composed of 206 pavement Moreover, another popular metric for semantic segmenta-
images with various types of cracks. CFD has 118 images tion, IoU, measures the percentage of overlap between the
collected from urban road surfaces in Beijing, and Aigle-RN ground-truths and the predictions, which can be defined as
only contains 38 French pavement images with the size of follows:
991 × 462. Therefore, these data sets are usually integrated Area of Overlap
IoU = (4)
together to be used as a validation set for evaluation. An- Area of U nion
other crack dataset, GPAs364 [50], contains 1969 gray-scale Traditional crack detection algorithms can achieve good
images with multiple classes of pavement distress such as performance on early datsets, however, they cannot well han-
cracks, patholes, inlaid patches, etc. And Liu et al. [51] pro- dle crack images with complex backgounds. Due to the fact
posed collected G45 dataset, which contains 122 gray-scale that deep learning algorithms have stronger generalization
images of pavement crack on the G45 highway in China. capabilities, with the rapid development of deep learning
The images of the G45 have a resolution of 2048 × 1536, technology, more and more advanced deep learning-based
and some images contain oil stains, speckle noise, and lane algorithms have been proposed to tackle the problem of
lines, which makes crack detection difficult. The DeepCrack pavement crack detection in order to achieve better perfor-
dataset proposed in [35] contains 537 RGB color images. mance.
There are various types of crack with different backgrounds
in these images. The percentages of cracked pixels and non- 5 Problems and prospects
cracked pixels in these data sets are unbalanced, which is
also a challenge for pavement crack detection. Although many research results have been made in pave-
As for the metrics for the evaluation of the algorithms, ment crack detection, there are still the following problems:
precision, recall and F1-Score(as formula (1-3)) are the most 1) At present, most of the work is done in laboratory con-
common ways. ditions, and automatic pavement crack detection technology
TP has not been widely promoted; 2) Pavement conditions are
P = (1) complex and changeable, and the data collection process
TP + FP
can be easily disturbed by factors such as illumination and
TP shadow. Therefore, the robustness and generalization of the
R= (2)
TP + FN algorithm still need to be improved; 3) The automatic pave-
ment crack detection algorithms need to continuously im-
2P R
F1 = (3) prove the accuracy while also taking the efficiency into ac-
P +R count. Only in this way can the algorithm better meet the re-
Where T P , F P and F N denote the numbers of true posi- quirements of real-time segmentation in certain scenarios; 4)
tive, false positive and false negative, respectively. At present, only a small number of public datasets are avail-

7304
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 able and more crack images need to be collected in order to actions on Intelligent Transportation Systems, 17(10):2718–
train more powerful models. 2729, 2016.
Automatic pavement crack detection algorithms require [15] E. W. Dijkstra. A note on two problems in connection with
more in-depth research on accuracy, real-time and stabil- graphs. Numerische Mathematik, 1959.
ity. Real-time, efficient, and stable image-based automatic [16] Sylvie Chambon and Jean Marc Moliard. Auto-
matic road pavement assessment with image process-
crack detection will be one of road maintenance in the fu-
ing: Review and comparison. International Journal of
ture, which has good application prospects. Geophysics,2011,(2011-10-26), 2011:1–20, 2011.
[17] Bo Peng, Yangsheng Jiang, Shifan Han, and Nanxin Luo. The
References review of pavement crack image automatic recognition algo-
rithms(in chinese). Journal of Highway Communication Tech-
[1] Shanghai Municipal Highway Administration Office. Techni- nology, 31(7):19–25, 3 2014.
cal Specification for Maintenance of Highway Asphalt Pave- [18] Ronny Salim Lim, Hung Manh La, and Weihua Sheng. A
ments(in Chinese). China Communications Press, 2001. robotic crack inspection and mapping system for bridge deck
[2] Highway Bureau of Jiangsu Provincial Department of Com- maintenance. IEEE Transactions on Automation Science &
munications. Technical Specifications for Maintenance of Engineering, 11(2):367–378, 2014.
Highway Cement Concrete Pavement(in Chinese). China [19] Dejin Zhang, Qingquan Li, Ying Chen, Min Cao, Li He, and
Communications Press, 2001. Bailing Zhang. An efficient and reliable coarse-to-fine ap-
[3] Beijing daily. Daxing airport countdown, beijing is about proach for asphalt pavement crackdetection. Image and Vi-
to fly into the era of double hub(in chinese). [EB/OL], sion Computing, 57(jan.):130–146, 2016.
2019. http://www.xinhuanet.com/2019-06/29/ [20] Miguel Gavilán, David Balcones, Oscar Marcos, David F
c_1124686676.htm. Llorca, Miguel A Sotelo, Ignacio Parra, Manuel Ocaña, Pe-
[4] Thouraya Merazi-Meksen, Malika Boudraa, and Bachir dro Aliseda, Pedro Yarza, and Alejandro Amı́rola. Adaptive
Boudraa. Mathematical morphology for tofd image analy- road crack detection system by pavement classification. Sen-
sis and automatic crack detection. Ultrasonics, 54(6):1642– sors, 11(10):9628–9657, 2011.
1648, 2014. [21] Zou, Cao, QQ, Mao, QZ, and Wang. Crack tree: Automatic
[5] Hoang Nam Nguyen, Tai Yan Kam, and Pi Ying Cheng. An crack detection from pavement images. PATTERN RECOGN
automatic approach for accurate edge detection of concrete LETT, 2012,33(3)(-):227–238, 2012.
crack utilizing 2d geometric features of crack. Journal of Sig- [22] Haifeng Li, Dezhen Song, Yu Liu, and Binbin Li. Automatic
nal Processing Systems, 2014. pavement crack detection by multi-scale image fusion. IEEE
[6] Xingwang Guo and Vladimir Vavilov. Crack detection in alu- Transactions on Intelligent Transportation Systems, PP(6):1–
minum parts by using ultrasound-excited infrared thermogra- 12, 2018.
phy. Infrared Physics & Technology, 61(Complete):149–156, [23] Zhun Fan, Yuming Wu, Jiewei Lu, and Wenji Li. Automatic
2013. pavement crack detection based on structured prediction with
[7] Zhongcheng Gui and Haifeng Li. Automated defect detection the convolutional neural network, 2018.
and visualization for the robotic airport runway inspection. [24] Shuguang Wu and Yaokui Liu. A segment algorithm for crack
IEEE Access, PP(99):1–1, 2020. dection. In Electrical & Electronics Engineering, 2012.
[8] Farabet, Clement, Couprie, Camille, Najman, Laurent, Le- [25] Shu, Zhibiao, Guo, and Yanqing. Algorithm on contourlet do-
Cun, and Yann. Learning hierarchical features for scene la- main in detection of road cracks for pavement images. Jour-
beling. IEEE Transactions on Pattern Analysis & Machine nal of Algorithms & Computational Technology, 7(1):15–25,
Intelligence, 35(8):1915–1929, 2013. 2013.
[9] Long, Jonathan, Shelhamer, Evan, Darrell, and Trevor. Fully [26] Na Wei, Xiang Mo Zhao, Xiao Yu Dou, Hong Xun Song,
convolutional networks for semantic segmentation. IEEE and Tao Wang. Beamlet transform based pavement image
Transactions on Pattern Analysis & Machine Intelligence, crack detection. In International Conference on Intelligent
2017. Computation Technology & Automation, 2010.
[10] Young-Jin, Cha, Wooram, Choi, Oral, and Büyüköztürk. [27] Dongho Kang, Sukhpreet S Benipal, Dharshan L Gopal, and
Deep learning-based crack damage detection using convolu- Young Jin Cha. Hybrid pixel-level concrete crack segmen-
tional neural networks. Computer Aided Civil & Infrastruc- tation and quantification across complex backgrounds using
ture Engineering, 2017. deep learning. Automation in Construction, 118:103291,
[11] Nhat Duc Hoang, Quoc Lam Nguyen, and Van Duc Tran. 2020.
Automatic recognition of asphalt pavement cracks using [28] Xincong, Yang, Heng, Li, Yantao, Yu, Xiaochun, Luo, Ting,
metaheuristic optimized edge detection algorithms and con- and Huang and. Automatic pixel-level crack detection and
volution neural network. Automation in Construction, measurement using fully convolutional network. Computer-
94(OCT.):203–213, 2018. Aided Civil and Infrastructure Engineering, 2018.
[12] Qin Zou, Zheng Zhang, Qingquan Li, Xianbiao Qi, Qian [29] Qipei Mei, Mustafa Gül, and Md Riasat Azim. Densely con-
Wang, and Song Wang. Deepcrack: Learning hierarchical nected deep neural network considering connectivity of pixels
convolutional features for crack detection. IEEE Transactions for automatic crack detection. Automation in Construction,
on Image Processing, 2019. 110(February 2020):103018, 2019.
[13] Li Peng, Wang Chao, Li Shuangmiao, and Feng Baocai. Re- [30] Lei Zhang, Fan Yang, Daniel Zhang, and Ying Julie Zhu.
search on crack detection method of airport runway based Road crack detection using deep convolutional neural net-
on twice-threshold segmentation. In 2015 Fifth International work. In IEEE International Conference on Image Process-
Conference on Instrumentation & Measurement, Computer, ing, 2016.
Communication and Control (IMCCC), 2015. [31] A. Doulamis, N. Doulamis, E. Protopapadakis, and
[14] Rabih Amhaz, Sylvie Chambon, Jérme Idier, and Vincent A. Voulodimos. Combined convolutional neural networks and
Baltazart. Automatic crack detection on 2d pavement images fuzzy spectral clustering for real time crack detection in tun-
: An algorithm based on minimal path selection. IEEE Trans- nels. In 2018 25th IEEE International Conference on Image

7305
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.
 Processing (ICIP), pages 4153–4157, Oct 2018. level crack detection on concrete bridges. IEEE Transactions
[32] Seongdeok Bang, Somin Park, Hongjo Kim, and Hyoungk- on Industrial Informatics, pages 1–1, 2020.
wan Kim. Encoder–decoder network for pixel-level road [48] Qin Zou, Yu Cao, Qingquan Li, Qingzhou Mao, and Song
crack detection in black-box images. Computer-Aided Civil Wang. Cracktree: Automatic crack detection from pavement
and Infrastructure Engineering, 2019. images. Pattern Recognition Letters, 33(3):227–238, 2012.
[33] Cao Vu Dung and Le Duc Anh. Autonomous concrete crack [49] Y. Shi, L. Cui, Z. Qi, F. Meng, and Z. Chen. Automatic road
detection using deep fully convolutional neural network. Au- crack detection using random structured forests. IEEE Trans-
tomation in Construction, 99:52–58, 2018. actions on Intelligent Transportation Systems, 17(12):3434–
[34] Yue Pan, Gaowei Zhang, and Limao Zhang. A spatial-channel 3445, 2016.
hierarchical deep learning network for pixel-level automated [50] M. Eisenbach, R. Stricker, D. Seichter, K. Amende, K. Debes,
crack detection. Automation in Construction, 119:103357, M. Sesselmann, D. Ebersbach, U. Stoeckert, and H. Gross.
2020. How to get pavement distress detection ready for deep learn-
[35] Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li. ing? a systematic approach. In 2017 International Joint
Deepcrack: A deep hierarchical feature learning architecture Conference on Neural Networks (IJCNN), pages 2039–2047,
for crack segmentation. Neurocomputing, 338(APR.21):139– 2017.
153, 2019. [51] Wenjun Liu, Yuchun Huang, Ying Li, and Qi Chen. Fpc-
[36] Yu Xie, Fangrui Zhu, and Yanwei Fu. Main-secondary net- net: Fast pavement crack detection network based on encoder-
work for defect segmentation of textured surface images. In decoder architecture. arXiv preprint arXiv:1907.02248, 2019.
2020 IEEE Winter Conference on Applications of Computer
Vision (WACV), 2020.
[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-
net: Convolutional networks for biomedical image segmen-
tation. In International Conference on Medical Image Com-
puting and Computer-Assisted Intervention, 2015.
[38] Jacob Konig, Mark David Jenkins, Mike Mannion, Peter Bar-
rie, and Gordon Morison. Optimized deep encoder-decoder
methods for crack segmentation. Digital Signal Processing,
108:102907, 2021.
[39] Jacob Konig, Mark David Jenkins, Peter Barrie, Mike Man-
nion, and Gordon Morison. A convolutional neural network
for pavement surface crack segmentation using residual con-
nections and attention gating. In 2019 IEEE International
Conference on Image Processing (ICIP), 2019.
[40] Stephen J. Schmugge, Lance Rice, John Lindberg, Robert
Grizziy, and Min C. Shin. Crack segmentation by leveraging
multiple frames of varying illumination. In 2017 IEEE Win-
ter Conference on Applications of Computer Vision (WACV),
2017.
[41] V. Badrinarayanan, A. Kendall, and R. Cipolla. Segnet: A
deep convolutional encoder-decoder architecture for image
segmentation. IEEE Transactions on Pattern Analysis & Ma-
chine Intelligence, 39(12):2481–2495, 2017.
[42] Kangwei Wang, Xin Zhang, Qiushi Hao, Yan Wang, and
Yi Shen. Application of improved least-square generative ad-
versarial networks for rail crack detection by ae technique.
Neurocomputing, 332(MAR.7):236–248, 2019.
[43] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing
Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
Yoshua Bengio. Generative adversarial networks. arXiv
preprint arXiv:1406.2661, 2014.
[44] Kaige Zhang, Yingtao Zhang, and Heng-Da Cheng. Crack-
gan: Pavement crack detection using partially accurate
ground truths based on generative adversarial learning. IEEE
Transactions on Intelligent Transportation Systems, page
1–14, 2020.
[45] Qiaoning Yang, Weimin Shi, Juan Chen, and Weiguo Lin.
Deep convolution neural network-based transfer learning
method for civil infrastructure crack detection. Automation
in Construction, 116:103199, 2020.
[46] Cao Vu Dung, Hidehiko Sekiya, Suichi Hirano, Takayuki
Okatani, and Chitoshi Miki. A vision-based method for crack
detection in gusset plate welded joints of steel bridges using
deep convolutional neural networks. Automation in Construc-
tion, 102:217–229, 2019.
[47] W. Jiang, M. Liu, Y. Peng, L. Wu, and Y. Wang. Hdcb-net: A
neural network with the hybrid dilated convolution for pixel-

7306
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on July 01,2022 at 05:08:24 UTC from IEEE Xplore. Restrictions apply.