2023 3rd International Conference on Neural Networks, Information and Communication Engineering (NNICE)

Dam crack detection based on deep learning cascade

detection algorithm
2023 3rd International Conference on Neural Networks, Information and Communication Engineering (NNICE) | 979-8-3503-3597-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/NNICE58320.2023.10105763

Qinchuan WANG1,a,* Bin Li2,b

POWERCHINA Chengdu Engineering Corporation Limited, POWERCHINA Chengdu Engineering Corporation Limited,
Chengdu; 610072, China Chengdu; 610072, China
a,*
1193718851@qq.com b
2011029@chidi.com.cn

Ziqian Deng3,c Wanting Zhao

POWERCHINA Chengdu Engineering Corporation Limited, POWERCHINA Chengdu Engineering Corporation Limited,
Chengdu; 610072, China Chengdu; 610072, China
C
2015017@chi.com.cn d
2012045@chidi.com.cn

Abstract—In order to improve the crack detection accuracy of information, and still has the problems of incomplete extraction
concrete dam surface, a cascade crack detection algorithm based of crack information and incomplete removal of interference,
on deep learning and prior information reinforcement is and does not take into account the location information of
proposed. In the algorithm, two neural networks are cascaded to cracks in the picture. Gao[3-5] et al. used the cascading network
detect cracks by a cascade detection method. Prior information structure to classify cracks first and then segment them,
strengthening includes standardization and edge strengthening. achieving the two purposes of classification and segmentation.
When making standardized crack slices, the crack is rotated to However, such cascaded approach can not classify fractures
the horizontal trend and the center of mass is placed in the with weak features or complex backgrounds well, and the final
section center. Meanwhile, slices belonging to the general
segmentation result is bound to have a large error with the
background are removed according to the gray gradient
similarity. The slices were processed with the prior information
actual fracture situation.
of crack edges to artificially highlight the features and reduce the In this paper, a deep learning cascade detection algorithm
difficulty of network training. The network results without prior for concrete cracks based on prior information reinforcement is
information reinforcement and cascade detection are compared proposed, which can achieve better classification effect only by
respectively to evaluate the network detection performance. The cascading classification tasks. First, the original images
experimental results show that the accuracy rate and recall rate obtained by the camera equipment were processed and
of concrete dam surface crack detection have been improved
screened in a standardized way to unify the fracture patterns
significantly, which proves the effectiveness of the algorithm.
and strengthen the position information. Then, the edge
Keywords- Cracks,；deep learning,；algorithms
information of the crack slices was strengthened to construct
the training data set. The algorithm cascades into two stages: in
the first stage, the neural network classifies whether there are
I. INTRODUCTION lines in the slices; In the second stage, the neural network
Crack is a common problem in concrete dams, which not classifies the slices that were previously judged to be thread-
only affects the structure function and bearing capacity of like. The cascade network trained by information enhancement
DAMS, but also causes quality accidents. Therefore, it is very slices can obtain ideal classification effect under complex
important to find and repair the crack of concrete dam in time. background, which proves the superiority of prior information
Traditional dam crack detection relies on manual close enhancement and cascade detection in crack detection.
observation. This method is not only inefficient, but also has
great difficulty in detecting the parts that are difficult to be II. THEORETICAL BASIS
reached by manual in the face of complex environment.
With the development of computer technology, researchers A. Cascading Classification Network Structure
have realized the automatic detection of cracks by manually The cascaded network filters the final result step by step
selecting the characteristics of cracks. At present, automatic through a weak to strong classifier. In view of the interference
crack recognition algorithms are mainly based on image of more concrete surface and high crack similarity, a two-level
processing algorithms and algorithms based on deep learning. network structure is designed. The identification process is
Wen[1] used small slices to train deep neural networks. Liang [2] shown in Figure 1.
de-noised the images, then extracted the features of the In the first stage, AlexNet network is used to classify the
corresponding regions by using the connected domain general background and lines. Due to the large difference
information of the de-noised images, and finally put the between the general background and the line, AlexNet with
processed images into the deep neural network for training. It is simple structure and few parameters is adopted as the
designed to highlight the crack information in the picture and classification network, and its network model structure is
dilute the irrelevant information. This information shown in Figure 2.
enhancement method only relies on the connected domain

979-8-3503-3597-2/23/$31.00 ©2023 IEEE 131

Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on July 22,2023 at 19:50:03 UTC from IEEE Xplore. Restrictions apply.
 It can be seen from Figure 2 that the AlexNet network has Table 1 Network model structure of Inceptionv2
two upper and lower parts, which correspond to two GPU Layer Name Patch size/stride Output size
respectively for computing. This setting only uses two GPU to
Conv1 3×3/2 299×299×3
improve computing efficiency. When one GPU is used for
Conv2 3×3/1 149×149×32
training, there are 5 convolutional layers in the structure, and
the size of the convolutional kernel ranges from large to small Conv3 3×3/1 147×147×32
from 11×11, 5×5, 3×3, with 3 maximum pooling layers and the Pool1 3×3/2 147×147×64
last 3 fully connected layers in the middle. The input size of the Conv4 3×3/1 73×73×64
network is [224,224,3], and the output feature matrix size is [6, Conv5 3×3/2 71×71×80
256] after one convolution with step size of 4 and three Conv6 3×3/1 35×35×192
maximum pooled subsampling with step size of 2. Finally, the 3×Inception 35×35×288
number of nodes output by the fully connected layer is 1000, 5×Inception 17×17×768
and the number of output categories is 3, corresponding to the 2×Inception 8×8×1 028
lines, background and honeycomb pockmarked surface. Pool2 8×8 8×8×2 048
In the second stage of fracture classification network, the Linear 1×1×2 048
lines identified in the first stage are classified into cracks, non- Softmax 1×1×1 000
cracks and suspected cracks. As the defects here are all linear
with similar characteristics, Inceptionv2 network with larger B. Sample preparation with prior information
number of parameters is used as the classification network, and reinforcement
the model structure is shown in Table 1. The size of concrete pictures obtained in the project is about
3000×4000, and the training samples are divided into 128×128.
As can be seen from Table 1, Inceptionv2 network The conventional sliding method has the problem of unstable
decomposed common large-scale convolution into multiple fracture pattern, that is, the length and direction of the fracture
small-scale convolution, decomposed a 5×5 convolution into in the section are different. Training samples with unstable
two 3×3 convolution, and reduced the computational amount of patterns will increase the difficulty of network recognition. For
(3×3+3×3)/(5×5) = 18/25 on the basis of the same coverage. neural networks to obtain the ideal recognition effect, it is
The final output classification number is 3. Corresponding necessary to maintain the stability of the patterns within the
crack, non crack and suspected crack. training set and the difference of the patterns between the
classes. Therefore, this paper proposes a standardized operation
for crack location to enhance the prior information of crack
location. In this paper, edge was extracted from sections
traversed with fixed step size, and the centroid and principal
direction Angle of the sections were calculated. Then the
sections were re-cut with the centroid as the center, and the
new and improved sections were obtained by rotating the re-cut
sections according to the principal direction Angle. The slices
obtained by this method are shown in Figure 3. It can be found
that the cracks in these slices are basically in the position of
section axis, and the trend is controlled in the horizontal
direction as far as possible. The slices obtained in this way
have stable patterns and unified features, and the attitude and
position of cracks are controlled within a relatively fixed range,
which is conducive to the convergence of the training model
and reduces the difficulty of pattern recognition.
For an image with a resolution above 4k, about 4000 slices
can be obtained by using the above method. In order to
improve the quality of training samples, this paper adopts gray
gradient similarity to quickly screen crack samples and remove
Figure 1 Concrete crack identification algorithm flow background. In this paper, Pearson correlation coefficient is
used to quantify the correlation of gray gradient. The purpose
of using gray gradient similarity is to screen out samples with
almost no edge complementarity such as background, rather
than directly identify cracks, so 0.3 is selected as the threshold
of gray gradient correlation. That is, when the gray gradient
correlation of slices is less than 0.3, it will be considered as the
background and removed from the sample library. In addition,
in the final recognition process, when the gray gradient
Figure 2 AlexNet network structure correlation of slices is detected to be less than 0.3, it will not be
sent to the neural network for judgment, which improves the

132

Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on July 22,2023 at 19:50:03 UTC from IEEE Xplore. Restrictions apply.
recognition efficiency and reduces the possibility of network TP
misclassification. R (2)
TP  FN
A. Improvement effect evaluation
The test photos all come from the four dams outside the
training project, which has enough objectivity. In order to
verify the performance superiority of the cascade classification
Figure 3 The fracture slices were obtained using the method proposed in this network proposed in this paper, a control experiment is
project designed to compare with the one-step crack classification
network and analyze the detection accuracy of the two methods.
For cracks, edge information is one of the most important First, Alexnet was taken as the basic model, sample base was
features, so neural networks are expected to focus on learning used as the input pre-training model and test was carried out
the edge of cracks. The existing technology only does the with the test set. The output results were directly classified into
traditional filtering denoising and gray-scale processing on the 5 categories, and the effects were shown in Table 2.
sample, which can not well highlight the crack edge. Therefore,
this paper proposes to strengthen the prior information through
the superposition of three kinds of fracture edge information Table 2 Single classification test results
images. Firstly, gray-scale image GR of sample slices was Output
made. The edge image CE of GR is calculated by canny Input Pseudo- Suspected
Crack Accuracy /%
operator. OSTU was used to segment GR binary image and crack crack
extract its contour to get a new binary image OC. GR, CE and Crack 105 32 16 68.6
OC were connected to obtain a new three-channel picture, that Pseudo-crack 14 11 1 42.3
is, samples of enhanced edge information. Operations at each Suspected crack 16 6 3 12
stage are shown in Figure 3. Then the cascade classification network proposed in this
paper is tested. In the first stage, the object to be tested is
divided into linear objects. In the second stage, the threads are
classified as cracks, pseudo-cracks, and suspected cracks.
Alexnet and Inceptionv2 network models were used to train the
two stages respectively. The output is shown in Table 3. By
comparing the two test results, it can be seen that the accuracy
of crack identification is improved from68.6% to 77.8% after
adopting the cascade detection method, with obvious
improvement effect.

Table 3 Multilevel classification test results

Output
Input Pseudo- Suspected Accuracy
Crack
crack crack /%
Crack 119 18 16 77.8
Figure 4 Operation at each stage Pseudo-crack 8 18 0 69.2
Suspected crack 5 3 17 68

III. EXPERIMENTAL RESULTS AND ANALYSIS To test and analyze the effect of the prior information
enhancement technology proposed in this paper, the original
In this paper, the accuracy rate P and recall rate R are used image sample library and the strengthened sample library are
as the criteria to evaluate the quality of the model. The respectively sent to the cascade network for training and testing.
accuracy rate, for our prediction, is how many of the predicted There is only a difference between the two sample libraries
positive samples are actually positive samples. The recall rate with or without reinforcement. The test results of the original
is for our original sample and represents how many positive sample are shown in Table 4 and Table 5, and the test results of
cases in the sample were predicted correctly. In the crack the strengthened sample are shown in Table 6 and Table7. It
detection task, set the number of cracks correctly classified as can be seen that the samples strengthened with edge
TP, the number of cracks erroneously detected as FP, and the information have better recall rate and accuracy rate in the two
number of background erroneously detected as FN. The stages of cascade after training than the original image sample
formula for calculating the accuracy rate P and recall rate R is library. The accuracy and recall rates of 91.62% and 89.9%
as follows: were obtained by using the original image sample library in the
classification of linear materials in the first stage, while the
TP
P (1) accuracy and recall rates of cracks were 91.87% and 95.14% in
TP  FP the classification of cracks in the second stage. When using the
enhanced sample bank training, the accuracy rate and recall

133

Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on July 22,2023 at 19:50:03 UTC from IEEE Xplore. Restrictions apply.
rate of linear classification in the first stage increased by 5.74% ASINVOS 0.6257 0.4875
and 5.93%, and that of crack classification in the second stage GF-CNN 0.8105 0.6365
increased by 1.55% and 4.46%. CNN-CRF 0.8128 0.9058
Algorithm of this paper 0.8875 0.9712
Table 4 Classification results of the first stage of the original image sample As can be seen from Table 8, the detection accuracy and
library recall rate of cracks of the proposed detection algorithm can
reach 88.75% and 97.12 %. The accuracy of traditional image
Output
Input processing algorithms CrackIT and CNN for crack detection is
background thread Accuracy /% less than 80%. GF-CNN and CNN-CRF algorithms, which are
Background 6 498 612 91.39 filtered and improved in structure on the basis of CNN,
thread 485 5489 91.88 increase the crack detection accuracy to 81.05% and 81.28%,
Recall rate /% 93.1 89.9 91.62 with a small gap between the two accuracy rates, but a 26.93%
gap between the recall rate. Compared with the CNN-CRF
Table 5 Classification results of the second stage of the original image sample algorithm with the best effect, the accuracy rate of the proposed
library algorithm is increased by 7.47% and the recall rate by 6.54%.
Output The results show that the prior information reinforcement and
Input Pseudo- Suspected cascade detection techniques adopted in this paper are more
Crack Accuracy /% accurate in concrete crack detection.
crack crack
Crack 1 956 121 52 91.87
Pseudo-crack 65 1 924 56 93.95 IV. CONCLUSION
Suspected crack 35 29 1 746 96.46 Aiming at the problem of surface crack detection of
Recall rate /% 95.14 92.77 94.17 94.02 concrete dam, a cascade crack detection method strengthened
by prior information is proposed in this paper, and the
Table 6 Classification results of the first stage of the enhanced sample library following conclusions can be drawn.
Output (1) Compared with the traditional image preprocessing
Input
background thread Accuracy /% operations such as filtering denoising and graying, the prior
Background 6856 254 96.43 information strengthening method proposed in this paper,
thread 142 5 832 97.62 which stabilizes the crack trend at the horizontal level and
Recall rate /% 97.97 95.83 96.97 strengthens the edge information, can obtain better network
training results.
Table 7 Classification results of the second stage of the enhanced sample (2) The cascade detection is used to first screen out linear
library
objects and then detect cracks from the linear objects. This step
Output can greatly reduce the influence of background interferences on
Input
Crack
Pseudo- Suspected
Accuracy /%
detection results. Combined with the prior information
crack crack strengthening operation in the first step, the accuracy rate and
Crack 1 989 95 45 93.42 recall rate of the algorithm for crack classification reach
Pseudo-crack 2 2 040 3 99.76 93.42% and 99.60% respectively. Basically meet the
Suspected crack 6 5 1798 99.39 requirements of engineering application
Recall rate /% 99.60 95.33 97.40 97.39
REFERENCES
B. Algorithm evaluation
[1] WEN. Concrete Crack Recognition Based on Deep Learning [D].
In order to evaluate the performance of the proposed Zhejiang University, 2019. DOI:10.27461/d.cnki.gzjdx.2019.002297.
algorithm in crack identification, the proposed algorithm is [2] Liang Xuehui, Cheng Yunze, Zhang Ruijie, et al. Recognition and
compared with various crack identification methods. The measurement of bridge crack based on Convolutional Neural Network
specific test results of the proposed algorithm are shown in [J]. Journal of Computer Applications,20,40(04):1056-1061.
Table 8. The recognition methods involved in the comparison https://kns.cnki.net/kcms/detail/51.1307.TP.20191118.1107.010.html
include CrackIT based on traditional image processing method, [3] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for
ASINVOS based on classic CNN pavement crack detection semantic segmentation, in: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2015, pp. 3431–3440.
model, GF-CNN based neural network based on Gabor filter, https://arxiv.org/abs/1411.4038
and detection algorithms based on CNN and CRF. The [4] Mei Q, Gül M (2020) A cost effective solution for pavement crack
experimental results of accuracy rate and recall rate are shown inspection using cameras and deep neural networks. Construction and
in Table 8. Building Materials 256:119397.DOI: 10.1016/j.conbuildmat.
2020.119397
[5] Gao J, Yuan D, Tong Z, Yang J, Yu D (2020) Autonomous pavement
Table 8 The crack detection model compares the experimental results distress detection using ground penetrating radar and region-based deep
learning. Measurement 108077,DOI:10.1016/j.measurement.
Project Accuracy rate Recall rate 2020.108077
CrackIT 0.4524 0.5297

134

Authorized licensed use limited to: b-on: UNIVERSIDADE DO PORTO. Downloaded on July 22,2023 at 19:50:03 UTC from IEEE Xplore. Restrictions apply.