Hindawi

Scientific Programming
Volume 2021, Article ID 3137083, 7 pages
https://doi.org/10.1155/2021/3137083

Research Article
Concrete Crack Detection Algorithm Based on Deep Residual
Neural Networks

Xiuying Meng
Henan Vocational College of Water Conservancy and Environment, Zhengzhou 450000, China

Correspondence should be addressed to Xiuying Meng; mengxiuyinghb@163.com

Received 29 April 2021; Revised 22 May 2021; Accepted 26 May 2021; Published 24 June 2021

Academic Editor: Shah Nazir

Copyright © 2021 Xiuying Meng. 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.

Crack is the early expression form of the concrete pavement disease. Early discovery and treatment of it can play an important role
in the maintenance of the pavement. With ongoing advancements in computer hardware technology, continual optimization of
deep learning algorithms, as compared to standard digital image processing algorithms, utilizing automation of crack detection
technology has a deep learning algorithm that is more exact. As a result of the benefits of greater robustness, the study of concrete
pavement crack picture has become popular. In view of the poor effect and weak generalization ability of traditional image
processing technology on image segmentation of concrete cracks, this paper studies the image segmentation algorithm of concrete
cracks based on convolutional neural network and designs an end-to-end segmentation model based on ResNet101. It integrates
more low-level features, which make the fracture segmentation results more refined and closer to the practical application
scenarios. Compared with other methods, the algorithm in this paper has achieved higher detection accuracy and
generalization ability.

1. Introduction shortcomings of manual detection methods, a crack damage

detection method based on crack image is developed
Concrete structure [1] is the basic element of constructing [15–18].
urban environment, which needs to be maintained sys- The traditional detection of crack damage is based on
tematically. Crack [2] is a common damage form of concrete pictures. In the early stage, the detection method based on
structure, which is one of the important indexes to evaluate morphology is often low and its detection accuracy is often
the safety of concrete structure [3, 4]. For bridges [5–7], low and is greatly affected by light and noise. In the later
roads [8, 9], and other engineering structures [10, 40], it is stage, methods such as seed points and tensor voting are
necessary to carry out regular management and mainte- used. Better detection accuracy however calculations are
nance; crack detection [11–14] is an important part of frequently complex, and computation efficiency is low. For
management and maintenance. At present, the crack de- crack pictures with complex backgrounds, false detections
tection is mainly artificial detection, which has a lot of and missed detections may occur, so it is necessary to de-
problems, such as in the crack detection of the bridge, the velop other concrete crack detection methods.
staff need to observe the cracks on the bottom and side of the In recent years, with the continuous improvement of
bridge by building scaffolding, which requires a lot of Internet technology [19] and the support of related hard-
preparation work and low detection efficiency. The safety of ware, artificial intelligence technology with deep learning
staff working in high places is also difficult to be guaranteed. [20] as the core has developed rapidly. Convolutional neural
There are blind spots in the manual detection process, and network is a feed-forward neural network that applies deep
the inspection effect is often affected by the professional learning. It has significant advantages such as high paral-
accomplishment and experience level of the detection lelism, good robustness, and strong generalization ability.
personnel, and the detection accuracy is low. Due to these Due to the characteristics of weight sharing, the
2 Scientific Programming

convolutional neural network has a much smaller amount of convolutional neural network (DCNN) AlexNet, the correct
calculation than the BP neural network, which reduces the rate of labeled samples reached 99%. The network pair in the
burden on the computer and does not require manual migration learning mode can correctly detect 86% of the
feature extraction. These all make it have a wide range of crack samples and can detect cracks with a width greater
applications in the field of computer vision, especially in than 0.04 mm, while the network in the fully trained mode
image classification and detection. When there are enough can also detect cracks greater than 0.08 mm. In terms of time,
samples, the classifier of the trained convolutional neural if the training process of the network is not considered, the
network can accurately identify concrete damage. Nowa- calculation time of DCNN is also shorter than the most
days, there are many convolutional neural network models effective edge detection algorithm. These all show the su-
with more efficient training and more accurate classification periority of using convolutional neural network to detect
every year. The detection of concrete cracks through con- concrete cracks.
volutional neural networks helps to reduce the subjective Kim and Cho [24] used concrete pictures on the Internet
influence of people and is more efficient and economical. It to construct a dataset, which included five categories, such as
is an excellent solution to the problem of crack detection. cracks and plants. AlexNet is trained by transfer learning,
Following are the main innovative points of this paper: that is, the network parameters of the pretrained AlexNet are
fine-tuned by using the existing dataset of concrete images.
To propose an improved fully convolutional neural
In the test process, two kinds of sliding windows with 50%
network model based on ResNet101, using ResNet101
overlap rate were used to scan the image so that the cracks
as the backbone network for feature extraction and
appearing on the edge of window 1 appeared in the central
then fusing the extracted features of different scales
area of window 2, so as to avoid the cracks being missed due
with the up-sampling and restoring feature maps.
to appearing on the edge of window, and the detection effect
Because it combines shallow position information and
was enhanced by adjusting the threshold value of classifi-
high-level semantic information, it makes the image
cation layer.
edge detail segmentation more refined.
The classification network locates the cracks based on the
To use the concrete crack detection method based on crack samples and cannot accurately describe the shape of
deep residual neural network proposed in this paper is a the cracks. Semantic segmentation is to classify each pixel in
nondestructive detection technology, which has urgent the image. Therefore, when the semantic segmentation
needs and extremely high application value in the field. network is applied to crack detection, the precise location
The rest of this paper is organized as follows: Section 2 and shape description of the crack can be realized at the
shows the related work of the paper. Section 3 discusses the same time.
methodology of the paper. The experiments and results are Ye et al. [25] used full convolutional networks (FCN) to
shown in Section 4. Section 5 shows the conclusion of the identify cracks in concrete bridge images at the pixel level
paper. and compared them with the results of edge detection
methods. The results show that the edge detection method is
2. Related Work greatly affected by noise, and when there is enough training
data collected from the real situation, the full convolutional
In recent years, convolutional neural networks have been network (FCN) can eliminate a lot of noise interference, the
widely used in the field of civil engineering. For cracks and performance is reliable, and the position of the crack in the
other structural damage problems, scholars usually use image is displayed, and the ability of the path, its detection,
classified convolutional neural networks to locate the has stronger robustness. Du et al. [26] proposed a new model
damaged parts and use semantic segmentation networks to based on the improvement of the original DeepLab v2 to
describe the damaged parts at pixel level to achieve the adapt to the particularity of crack damage detection.
description of the damage. Compared with the previous method, this method can
Classical convolutional neural networks include AlexNet obtain very high-precision output, and a new method of
[21] and GoogLeNet [22] which have specific requirements marking cracks is proposed, which is beneficial to the
on the size of the input image. When using these networks, measurement of crack length and width. Hoskere et al. [27]
the image to be tested needs to be divided into small samples proposed a structural damage detection method based on
of specific sizes. The network is trained to classify these multiscale pixel-level deep convolutional neural network
samples, determine the position of the crack sample on the [28–31]. This method utilizes two networks [32], derived
picture to be tested, and realize the location of the crack in from VGG19 and ResNet45, for structural damage classi-
the picture. fication and semantic segmentation, respectively. By parallel,
Dorafshan et al. [23] compared the detection effects of the two networks can simultaneously process multiple types
common edge detectors and deep convolutional neural of damage classification and pixel-level segmentation. Due
networks on concrete cracks. In this article, six edge de- to the high accuracy of damage detection in the classification
tection methods, Roberts, Prewitt, Sobel, Laplacian of network [33, 34], the segmentation network only needs to
Gaussian, Butterworth, and Gaussian, are used to detect perform semantic segmentation [35] on the images judged as
concrete crack images. The final binary image produces damage and does not need to identify the nondamage parts,
residual noise, and the best method can only detect cracks which reduces the probability of false detection. Hoskere
with a width greater than 0.1 mm. When using the deep constructed a dataset of 1695 images from 250 different
Scientific Programming 3

structures, including six types of damage. On this dataset, and learning rate. Finally, this experiment completed the
the applicability of the method to civil infrastructure was network model training under the Keras deep learning
verified, and the damage classification was achieved at the framework.
pixel level.

3. Methodology 3.2.1. Data Labeling and Data Augmentation.

Application scenarios of crack detection in this paper are
Some scholars have proposed a variety of methods based on mainly surface structures composed of concrete such as
digital image processing to detect cracks on the surface of bridges and roads, but currently there is no standard crack
concrete structures, but there may be many problems such as dataset specifically used for semantic segmentation. In
uneven illumination intensity and distortion in the actual order to obtain enough concrete crack images to train the
collected images, and the traditional image processing al- model, part of crack datasets was found in some papers, and
gorithms cannot solve these problems well. Due to its a large number of crack images were supplemented by
powerful feature extraction ability [36], deep learning manual photography. Before training the network model,
technologies [37–39] can better overcome the interference these collected crack images need to be annotated first. In
brought by external environmental factors, and it has better this paper, the professional semantic segmentation and
detection accuracy compared with traditional methods. annotation software Labelme is used to annotate the crack
Therefore, this paper proposes a concrete crack detection images. There is a significant difference in the target de-
algorithm based on deep residual neural network to achieve tection task dataset of annotation and semantic segmen-
pixel-level segmentation detection of concrete crack images. tation, as in the mark of target detection. You only need to
Next in this section, the proposed network model for crack paint the rectangular box in the target and the appropriate
segmentation and the whole training process are described categories, whereas annotation and semantic segmentation
in detail. are more difficult. In particular, this slender aim requires
In the maintenance of the surface of concrete structures, cracking along the margins of each point to indicate until
cracks are one of the most common and serious diseases and the contour map is complete. It forms an enclosed space.
are the early features of many other diseases, as shown in Figure 3 shows the process of labeling crack images with
Figure 1. Labelme software.
For semantic segmentation tasks, a large number of
training samples is the key factor to improve the segmen-
3.1. Segmentation Detection Algorithm Based on Deep Re- tation effect of network models, but sometimes it is not easy
sidual Neural Network. At present, most of the semantic to obtain a large number of datasets. An effective way to
segmentation network models based on deep learning adopt improve performance is to enlarge the existing data and
encoder-decoder network architecture, which is the main- expand the diversity of the existing data. This article uses
stream structure of semantic segmentation network models Augmentor, a very useful Python data enhancement tool, to
due to its simple structure, strong scalability, and good augment datasets. Its operation is very simple. According to
segmentation effect. In this paper, the encode-decoder ar- the original image and label, the program performs rotation,
chitecture is still adopted, and by combining the advantages of stretch, zoom-in, and zoom-out of the labeled image with a
FCN network model and U-Net network model, an improved certain probability to complete the expansion of the dataset.
full convolutional neural network (CrackNet model for short) After augmenting the data by Augmentor library, this paper
is proposed. CrackNet also contains the narrowing path for finally obtained 3000 fracture datasets, among which 2500
collecting image features and the expanding path for locating data were selected as the training set, and the remaining 500
and restoring features. The two are symmetrical structures were used as the test set.
and constitute a U-shaped network. The deeper and more
generalizable ResNet101 network is used in the CrackNet
encoder. Compared with VGG16 network, ResNet101 net- 3.2.2. Loss Function. The cross-direction loss function is one
work has fewer parameters and can extract deeper features. In of the most widely used loss functions in deep learning. This
the decoder, the advanced semantic features encoded are article adopts the cross-direction loss function, but it is
decoded and restored to the size of the original image by different from the cross-direction loss of the traditional
bilinear interpolation and convolution operation. Another image classification network in that the loss function in this
important feature of CrackNet is that it splices and integrates paper needs to predict and evaluate each pixel and then
the features of different scales extracted in the process of calculate the pixel average. If the crossover loss is greater, the
down-sampling with the feature images restored from up- corresponding gradient will be greater and the training
sampling, which further enhances the segmentation effect of speed will be faster. Since this crack image segmentation task
crack edge details. Figure 2 is the network structure diagram belongs to a two-class classification problem, the calculation
of this paper. equation of the cross-function is as follows:
L � y log y′ +(1 − y)log 1 − y′ 􏼁, (1)
3.2. Model Training. The training process of our model
mainly includes the selection of parameters such as fracture where y is the truly marked crack value, y′ is the predicted
dataset annotation, loss function, optimization algorithm, value of the model, L is the cross-child loss calculated according
4 Scientific Programming

Figure 1: Cracks on the concrete surface.

ResNet101
1
128
64
64
256256
256
512 256512
1024
Input image 2561024 Output image
2048
2048 2562048

14 × 14
28 × 28
56 × 56

112 × 112
224 × 224
448 × 448

ResNet101 Interpolate + Conv

Conv_x Concat
Pool

Figure 2: Network structure of the proposed model.

to the cross-child loss function, and the overall loss of the image without causing memory overflow, the batch-size of this
is equal to the average loss value of all pixel points. experiment is 4, the size of the input image is 448 × 448, the
loss function is the cross-saturated loss function, the opti-
4. Experiments and Results mization algorithm is the gradient descent method of
driving quantity, the learning rate is 0.001, and the mo-
4.1. Experimental Setup. All the experiments in this section mentum coefficient Y is 0.9. The number of iterations is 100
were performed on a Lenovo workstation with Nvidia epochs.
GeForce GTX 1080 Ti and PyCharm installed. In this paper,
the previously produced fracture dataset is divided into 2500
training sets and 500 test sets. Since the color information is 4.2. Evaluation Method. Precision and recall are two basic
of little significance to the concrete image and will take up quantitative evaluation indexes in semantic segmentation.
too much storage space, in order to reduce this redundant FP represents the total number of pixels that were wrongly
information, this paper converts the image of three-channel judged as cracks in the background, and TP represents the
concrete cracks captured into single-channel grayscale im- total number of pixels that were correctly extracted from the
age. In order to maximize the use of computer memory cracks. FN represents the total number of pixels that belong
Scientific Programming 5

Figure 3: Data labeling.

(a) (b) (c) (d) (e)

Figure 4: Visualization results of segmentation detection. (a) Input image. (b) Ground truth. (c) FCN-8s. (d) FCN-32s. (e) Ours.

(a) (b) (c) (d) (e)

Figure 5: Visualization results of segmentation detection. (a) Input image. (b) Ground truth. (c) FCN-8s. (d) FCN-32s. (e) Ours.

to the crack area but are misjudged as the background. The Table 1: Comparison results of crack segmentation algorithms.
calculation equations of fracture accuracy index P and recall
FCN Ours
rate index R are as follows:
P R P R
TP Sample 1 0.9261 0.9366 0.9452 0.9525
P� , (2)
TP + FP Sample 2 0.9312 0.9425 0.9569 0.9655

TP
R� . (3)
TP + FN
It can be seen from Figures 4 and 5 that although the
FCN segmentation method can overcome isolated noise
points, the effect of wider crack width is acceptable, but
4.3. Comparative Experiments. This paper conducts a because of the low-level features of its fusion, some cracks
comparative experiment with the classic semantic seg- are not as detailed as others, and the noise extract is not as
mentation model FCN. Figures 4 and 5 show the visuali- effectively split. The segmentation algorithm in this article
zation results of the comparative experiment. can still maintain a relatively stable and accurate extraction
6 Scientific Programming

100 segmentation model has better performance in segmenta-

tion accuracy and generalization ability.
90
Data Availability
80
The data used to support the findings of this study are in-
cluded within the article.
70

60
Conflicts of Interest
The author does not have any possible conflicts of interest.
50

0 2 4 6 8 10 12 14 16
References
Epoch [1] S. W. Tang, Y. Yao, C. Andrade, and Z. J. Li, “Recent durability
studies on concrete structure,” Cement and Concrete Research,
Validation vol. 78, pp. 143–154, 2015.
Training [2] S. Kabir, “Imaging-based detection of AAR induced map-
Figure 6: P curve during training. crack damage in concrete structure,” NDT & E International,
vol. 43, no. 6, pp. 461–469, 2010.
[3] G. Peiying, D. Chang, and T. Lei, “Determination of local
damage probability in concrete structure,” Procedia Engi-
100 neering, vol. 28, pp. 489–493, 2012.
[4] L. Yang and H. Jianguo, “Comparison of safety level of
90 Chinese and foreign crack control of concrete members,”
Building Structure, vol. 2, 2011.
80 [5] P. L. Fuhr and D. R. Huston, “Corrosion detection in rein-
forced concrete roadways and bridges via embedded fiber
70 optic sensors,” Smart Materials and Structures, vol. 7, no. 2,
pp. 217–228, 1998.
60 [6] P. L. Fuhr and D. R. Huston, “Corrosion detection in rein-
forced concrete roadways and bridges via embedded fiber
50 optic sensors,” Smart Materials and Structures, vol. 7, no. 2,
pp. 217–228, 1998.
40 [7] S. Yehia, O. Abudayyeh, S. Nabulsi, and I. Abdelqader,
0 2 4 6 8 10 “Detection of common defects in concrete bridge decks using
Epoch nondestructive evaluation techniques,” Journal of Bridge
Engineering, vol. 12, no. 2, pp. 215–225, 2007.
Validation [8] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata,
Training “Road damage detection and classification using deep neural
Figure 7: R curve during training. networks with smartphone images,” Computer-Aided Civil and
Infrastructure Engineering, vol. 33, no. 12, pp. 1127–1141, 2018.
[9] M. Foedisch and A. Takeuchi, “Adaptive real-time road de-
for different types of cracks. It can be clearly seen from tection using neural networks,” in Proceedings. The 7th In-
Table 1 that the algorithm in this paper is superior to the ternational IEEE Conference on Intelligent Transportation
other two methods in all indicators. In addition, Figures 6 Systems (IEEE Cat. No. 04TH8749), pp. 167–172, IEEE,
and 7 also show the P and R curves of the algorithm in the Washington, WA, USA, October 2004.
[10] M. Z. Naser, R. A. Hawileh, and J. A. Abdalla, “Fiber-rein-
training process.
forced polymer composites in strengthening reinforced
concrete structures: a critical review,” Engineering Structures,
5. Conclusion vol. 198, Article ID 109542, 2019.
[11] C. V. Dung and L. D. Anh, “Autonomous concrete crack
In this paper, an image segmentation algorithm for concrete detection using deep fully convolutional neural network,”
cracks based on convolutional neural network is studied, Automation in Construction, vol. 99, pp. 52–58, 2019.
and an end-to-end segmentation model based on ResNet101 [12] A. Mohan and S. Poobal, “Crack detection using image
processing: a critical review and analysis,” Alexandria Engi-
is designed. In order to better train the network model,
neering Journal, vol. 57, no. 2, pp. 787–798, 2018.
Labelme software was used to manually annotate the col- [13] Q. Zou, Z. Zhang, Q. Li, X. Qi, Q. Wang, and S. Wang,
lected real scene fracture dataset and expand the data. Then, “Deepcrack: learning hierarchical convolutional features for
Keras deep learning framework was used to train the model, crack detection,” IEEE Transactions on Image Processing,
and the extracted feature map was visually displayed. Finally, vol. 28, no. 3, pp. 1498–1512, 2018.
the proposed algorithm is compared with the FCN method, [14] X. Yang, H. Li, Y. Yu, X. Luo, T. Huang, and X. Yang, “Au-
and the experimental results show that the proposed tomatic pixel-level crack detection and measurement using
Scientific Programming 7

fully convolutional network,” Computer-Aided Civil and In- [31] G. Krishna Durbhaka, B. Selvaraj, M. Mittal, T. Saba,
frastructure Engineering, vol. 33, no. 12, pp. 1090–1109, 2018. A. Rehman, and L. Mohan Goyal, “Swarm-LSTM: condition
[15] T. Nishikawa, J. Yoshida, T. Sugiyama, and Y. Fujino, monitoring of gearbox fault diagnosis based on hybrid LSTM
“Concrete crack detection by multiple sequential image fil- deep neural network optimized by swarm intelligence algo-
tering,” Computer-Aided Civil and Infrastructure Engineering, rithms,” Computers, Materials & Continua, vol. 66, no. 2,
vol. 27, no. 1, pp. 29–47, 2012. pp. 2041–2059, 2021.
[16] H. G. Moon and J. H. Kim, “Intelligent crack detecting al- [32] Q. Liu, L. Cheng, A. L. Jia, and C. Liu, “Deep reinforcement
gorithm on the concrete crack image using neural network,” learning for communication flow control in wireless mesh
in Proceedings of the 28th ISARC, pp. 1461–1467, Seoul, South networks,” IEEE Network, vol. 35, no. 2, pp. 112–119, 2021.
Korea, June 2011. [33] W. Cai and Z. Wei, “Remote sensing image classification
[17] B. Kim and S. Cho, “Image-based concrete crack assessment based on a cross-attention mechanism and graph convolu-
using mask and region-based convolutional neural network,” tion,” IEEE Geoscience and Remote Sensing Letters, In Press,
Structural Control and Health Monitoring, vol. 26, no. 8, 2020.
p. e2381, 2019. [34] R. Liu, “Multiscale dense cross-attention mechanism with
[18] Y. Ren, J. Huang, Z. Hong et al., “Image-based concrete crack covariance pooling for hyperspectral image scene classifica-
detection in tunnels using deep fully convolutional networks,” tion,” Mobile Information Systems, vol. 2021, Article ID
Construction and Building Materials, vol. 234, Article ID 9962057, 15 pages, 2021.
117367, 2020. [35] X. Zhang, Y. Yang, Z. Li, X. Ning, Y. Qin, and W. Cai, “An
[19] J. Chen, C. Du, Y. Zhang, P. Han, and W. Wei, “A clustering- improved encoder-decoder network based on strip pool
based coverage path planning method for autonomous het- method applied to segmentation of farmland vacancy field,”
erogeneous UAVs,” IEEE Transactions on Intelligent Trans- Entropy, vol. 23, no. 4, p. 435, 2021.
portation Systems, vol. 99, pp. 1–11, 2021. [36] X. Ning, W. Li, B. Tang, and H. He, “BULDP: biomimetic
[20] X. Zhang, Y. Yang, Z. Li, X. Ning, Y. Qin, and W. Cai, “An uncorrelated locality discriminant projection for feature ex-
improved encoder-decoder network based on strip pool traction in face recognition,” IEEE Transactions on Image
method applied to segmentation of farmland vacancy field,” Processing, vol. 27, no. 5, pp. 2575–2586, 2018.
Entropy, vol. 23, no. 4, p. 435, 2021. [37] Y. Tong, L. Yu, S. Li, J. Liu, H. Qin, and W. Li, “Polynomial
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet fitting algorithm based on neural network,” ASP Transactions
classification with deep convolutional neural networks,” on Pattern Recognition and Intelligent Systems, vol. 1, no. 1,
Advances in Neural Information Processing Systems, vol. 25, pp. 32–39, 2021.
pp. 1097–1105, 2012. [38] J. Zhang, J. Sun, J. Wang, and X.-G. Yue, “Visual object
[22] C. Szegedy, W. Liu, Y. Jia et al., “Going deeper with con- tracking based on residual network and cascaded correlation
volutions,” in Proceedings of the IEEE conference on computer filters,” Journal of Ambient Intelligence and Humanized
vision and pattern recognition, pp. 1–9, Golden, CO, USA, Computing, vol. 20, 2020.
June 2015. [39] X. Ning, Y. Wang, W. Tian, L. Liu, and W. Cai, “A biomimetic
[23] S. Dorafshan, R. J. Thomas, and M. Maguire, “Comparison of covering learning method based on principle of homology
deep convolutional neural networks and edge detectors for continuity,” ASP Transactions on Pattern Recognition and
image-based crack detection in concrete,” Construction and Intelligent Systems, vol. 1, no. 1, pp. 9–16, 2021.
Building Materials, vol. 186, pp. 1031–1045, 2018. [40] Z. Luo, “Application and development of electronic com-
[24] B. Kim and S. Cho, “Automated crack detection from large puters in aero engine design and manufacture,” ASP Trans-
volume of concrete images using deep learning,” in Pro- actions on Computers, vol. 1, no. 1, pp. 6–11, 2021.
ceedings of the 7th World Conference on Structural Control and
Monitoring, pp. 22–25, Qingdao, China, July 2018.
[25] X.-W. Ye, T. Jin, and P.-Y. Chen, “Structural crack detection
using deep learning-based fully convolutional networks,”
Advances in Structural Engineering, vol. 22, no. 16,
pp. 3412–3419, 2019.
[26] Z. Du, J. Yang, C. Ou, and T. Zhang, “Smallholder crop area
mapped with a semantic segmentation deep learning
method,” Remote Sensing, vol. 11, no. 7, p. 888, 2019.
[27] V. Hoskere, Y. Narazaki, T. Hoang, and B. Spencer, “Vision-
based structural inspection using multiscale deep convolutional
neural networks,” 2018, https://arxiv.org/abs/1805.01055.
[28] L. Nkenyereye, B. A. Tama, and S. Lim, “A stacking-based
deep neural network approach for effective network anomaly
detection,” CMC-Computers Materials & Continua, vol. 66,
no. 2, pp. 2217–2227, 2021.
[29] L. Zhang, X. Wang, X. Dong, L. Sun, W. Cai, and X. Ning,
“Finger vein image enhancement based on guided tri-
Gaussian filters,” ASP Transactions on Pattern Recognition
and Intelligent Systems, vol. 1, no. 1, pp. 17–23, 2021.
[30] K. Khan, J. Ali, K. Ahmad et al., “3D head pose estimation
through facial features and deep convolutional neural net-
works,” Computers, Materials & Continua, vol. 66, no. 2,
pp. 1757–1770, 2021.