See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/346714622

Crack Detection of Structures using Deep Learning Framework

Conference Paper · December 2020

DOI: 10.1109/ICISS49785.2020.9315949

CITATIONS READS
24 2,929

4 authors, including:

Amit Kumar Avinash Kumar Jha

Delhi Technological University Delhi Technological University
2 PUBLICATIONS 74 CITATIONS 8 PUBLICATIONS 315 CITATIONS

SEE PROFILE SEE PROFILE

Ashutosh Trivedi
Delhi Technological University
77 PUBLICATIONS 710 CITATIONS

SEE PROFILE

All content following this page was uploaded by Avinash Kumar Jha on 13 February 2021.

The user has requested enhancement of the downloaded file.

 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

Crack Detection of Structures using Deep Learning Framework

2020 3rd International Conference on Intelligent Sustainable Systems (ICISS) | 978-1-7281-7089-3/20/$31.00 ©2020 IEEE | DOI: 10.1109/ICISS49785.2020.9315949

Arunish Kumar1, Amit Kumar1, Avinash Kumar Jha1, Ashutosh Trivedi1

1
Department of Civil Engineering, Delhi Technological University, India
{kumararunish7, amitvijaranyia, avinash.jha.6696}gmail.com, [email protected]

ABSTRACT cracks occur, it should be first detected and acted upon

suitably to remove it.
The transportation system has drawn attention due
to the growing threats of cracks to the roads and bridges. 1.2 Crack Detection Algorithm
To accurately detect cracks was made possible by the The detection of cracks can be performed by both visual
introduction of deep learning. In this paper, we propose inspection and automatic survey. Although both methods
the modified LeNet 5 model for detecting cracks in roads present good distress analysis results, the automatic crack
and bridges. We summarize our work in the following detection system is quicker, lower costing than traditional
points: - 1. We worked on the three datasets, i.e., human vision detection [15]. Therefore, automatic crack
Automated Bridge Crack Detection Dataset, Concrete detection has attracted much attention from scientific and
Crack Images for Classification Dataset, and Asphalt technical corporations in recent years [7]. For the crack
Crack Dataset. 2. We applied a modified version of detection purpose, we have modified the LeNet-5 model, and
LeNet-5 model on Automated Bridge Crack Detection used it. We have also applied Principal Component Analysis,
Dataset, Concrete Crack Images for Classification while predicting the cracks on the various materials that we
Dataset, and Asphalt Crack Dataset. 3. Then, we analyzed have used, to reduce model training time without
our results and compared the same with Principal compromising much with the accuracy. Trained model can
Component Analysis and without using Principal also be used to locate the cracks in an image by highlighting
Component Analysis. We are also highlighting the region it with red color. In the end, we are comparing the results that
of crack and non-crack using green and red color, we have got for the materials with using Principal Component
respectively. Our proposed model takes both the factors Analysis and without using Principal Component Analysis.
of time and accuracy into consideration while producing
the output. 2. LITERATURE REVIEW
Keywords— modified LeNet, principal component
analysis (PCA), deep learning, crack detection Eisenbach et al. [1] have examined different assessment
strategies applied on the pavement surface and have sorted
1. INTRODUCTION them into three gatherings: - crack image thresholding, patch-
based classification, and depth-based algorithms. Their
Machine learning has been used widely these days to solve paper's shortcomings were that the pictures captured by them
various real world problems [19]. From solving vast were distinct in 64*64 size, which does not give much
engineering problems to finance and healthcare, machine information. Yokoyama [2] built up the identifier that
learning is used everywhere [20-22]. In this paper, the consequently identifies cracks from the photos of solid
practical application of deep learning for crack detection has surfaces utilizing CNN. However, for the stained concrete,
been discussed. the detection rate was meager. Cha et al. [3] have utilized
1.1 Inspiration CNN classifiers that are less affected by clamor brought
Cracks are common distresses in concrete,rock about by lighting, shadow, projecting to identify cracks from
masses[15], asphalt as well as bridges [16]. Road surface, images and different sorts of auxiliary harms like voids,
aging, climate are the various reasons for the occurrence of spalling, and corrosion of cement in steel structures. These
different types of cracks [7]. Asphalt cracks are mainly due techniques probably will not have the option to treat the non-
to stress and friction. Friction value increases with the homogenous solid surface concerning color and texture.
increasing value of coefficient of friction [17]. The cracks can Pauly et al. [4] illustrated the potential of deep learning by
reduce the efficiency of materials. The water, by penetrating using deeper networks for the pavement crack detection.
the deeper layers of the materials through cracks, decreases However, when the training and testing images are taken
the compaction of the material, which results in the decrease from different areas, the network does not perform well.
in load-bearing capacity of the whole structure. Hence it is Hoang [5] has preprocessed the Otsu method's image
obvious to maintain the structures from time to time. If any threshold, which is an intensity adjustment method called

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 526

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

Min-Max Gray Level Discrimination (M2GLD). This The dataset possesses images of concrete. The dataset
strategy's downside was that the clients need to fine-tune has two categories, i.e., positive and negative. The positive
boundaries changing proportion (RA) and edge boundary (τ). category contains cracked images, and the negative has non-
By utilizing this examination, it is discovered that RA = 2 and cracked images of concrete. The dataset contains 40,000
τ = 0.5 can convey acceptable crack identification on the images of concrete of pixel size 227*227 [18].
building surface. Be that as it may, for various kinds of
pictures, such as black-top or Asphalt, these two qualities 3.1.3 Asphalt Crack Dataset
should be balanced. Zheng et al. [6], to boost and extract the This dataset consists of 400 images of Asphalt cracks
data features, they have suggested FCN, RCNN, and RFCN and Non-Cracks images, that are captured using Nikon
frameworks. Although the accuracy is improved with the help camera, and can be used for training crack detection models
of RCNN and FCN but the framework takes more time to [13].
train. Fan et al. [7] proposed U Hierarchical Dilated Network
(U-HDN) to detect cracks. U Hierarchical Network uses the 3.2 Deep Learning Approach
architecture of encoder-decoder along with hierarchical
feature learning. However, U Hierarchical Network although 3.2.1 Modified LeNet-5
produces acceptable results, but due to the neural network's Following Table 1 depicts the architecture of the
complicated structure, the computational cost is high, and the modified version of LeNet-5 model when applied on
efficiency is low. Li et al. [8] concocted a technique for Concrete Crack Images for Classification Dataset. Input
distinguishing the surface breaks for stacked coal with the shape to the first layer of the model is of size (*, 227, 227, 3),
help of a vibration failure process, which is based on a here * represents the total number of images used while
vibration failure system and SVM. However, they did not use training the model.
Principal Component Analysis to analyze further which
feature of the crack plays a major role so that they can reduce
TABLE I. ARCHITECTURE OF MODIFIED LE-NET5
the training time. Chen et al. [9] detected cracks for the
submerged metallic plane from nuclear inspection videos Layer 𝒑𝒂𝒕𝒄𝒉 𝒔𝒊𝒛𝒆
with the help of a deep learning model called NB-CNN. Output Shape Param #
(Type) 𝒔𝒕𝒓𝒊𝒅𝒆
However, they did not use Principal Component Analysis to
further reduce their training time. Zhou et al. [10] analyzed Conv2D (5 × 5)/1 (*, 223, 223, 6) 456
three-dimensional (3D) laser image range data with an
automated crack detection technique, which uses frequency Average
domain filtering and contouring analysis. However, the (2 × 2)/2 (*, 111, 111, 6) 0
Pooling2D
above-proposed method fails in case of detecting shallow
cracks. Lin et al. [11], with the help of a sandstone specimen Conv2D (5 × 5)/1 (*, 107, 107, 16) 2416
under three-point bending, determined the cohesive zone's
characteristics and cohesionless crack's length. However, the Average
primary outcomes need additional examinations because the (2 × 2)/2 (*, 53, 53, 16) 0
Pooling2D
results of some numerical modeling get challenged. Zhang et
al. [12] compared the Handcrafts method's crack detection Flatten _ (*, 44944) 0
performance with the deep learning model and concluded that
the latter performed better. The dataset that they worked upon Dense _ (*, 120) 5393400
was small, which consisted of just 500 images.
Dense _ (*, 84) 10164
3. METHODOLOGY AND EXPERIMENTATION
Dense _ (*, 1) 85
3.1 Datasets
Total Parameters # 5,406,521 (5.4 M)
3.1.1 Automated Bridge Crack Detection Dataset
The original dataset comprises of 2068 Bridge Crack Trainable Parameters # 5,406,521 (5.4 M)
Images, collected by the Phantom 4 Pro's CMOS surface
array camera, with a resolution of 1024*1024 [14]. Non-Trainable Parameters # 0

3.1.2 Concrete Crack Images for Crack Classification 3.2.1.1 Convolutional Layer
Dataset

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 527

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

The convolutional layers are the primary building function and AdamOptimiser as the optimiser. The formula
blocks of a convolutional neural network structure. It is the of BinaryCrossEntropy loss function:
convolutional layers that learn the features that are suitable
for differentiating between a crack and non-crack image.
Initial convolutional layers learn the local features required
(3)
for this, whereas the deeper layers learn from the global
features required for differentiating cracks and non-cracks
[4]. The stride is the additional hyperparameter of the layer. where, CE is Binary Cross Entropy, C1 and C2 are are the
The number of receptive field columns and rows covering the two classes[0, 1](in our case, it is crack or non-crack), t1
input width and height is defined by stride. Fewer receptive and s1 are groundtruth and score for C1 class respectively
field applications and smaller output sizes are due to larger and t2(1-t1) and s2(1-s2) are groundtruth and score for C2
stride, which reduces computational cost at the price of loss class respectively.
of some input data features.

3.2.1.2 Pooling Layers

The pooling layer reduces the number of parameters
and computation time by down-sampling the representation.
A pooling layer usually takes a patch of adjacent pixels and
tries to reduce it to a single pixel. A pooling layer is generally
included between two consecutive convolutional layers.
There are two types of pooling layers, Average Pooling layer Figure 1. Non-Linear Activation function
and Max Pooling Layer. We have used average pooling layer
between two successive convolutional layers.
3.3 Machine Learning Approach
3.2.1.3 Activation Layer 3.3.1 Principal Component Analysis
The typical Artificial Neural Network is given non- Principal Component Analysis helps us to reduce
linearity with the help of sigmoidal functions. However, Nair dimensionality. If there are vast numbers of features, there
and Hinton stated that the computations are slowed down by will be a high chance that they are co-related. Here we reduce
saturation of nonlinearities. A non-linear function called the data, and also we want to keep similar kinds of
RELU was introduced recently [3]. The example of a information with us. In Principal Component Analysis, if we
nonlinear function is depicted in Figure 1. Apart from its have a considerable number of features, we try to reduce it
negative input values, the RELU gives no bounded output into fewer features that would also explain our data very well.
than other nonlinear functions, generating bounded output Although we might be losing some amount of information
values. RELU's gradients are always zeros and ones. Due to while reducing the dimensions, the benefits of using it
these factors, RELU has faster calculations and better eclipses the disadvantages to a great extent. The advantages
accuracy than sigmoidal functions. We have used the of using Principal Component analysis are lesser training
Sigmoidal function in the output layer and in all other layers, time, lesser memory requirement, and easy to visualize.
RELU function is used. The formula for RELU function is: Before using Principal Component Analysis, usually feature
scaling is applied.
(1)
3.3.2 Math behind Principal Component Analysis
Principal Component Analysis requires the directions in
The formula for Sigmoid function is: which the data changes the maximum, i.e., the variance. The
Principle Component is found out using the covariance
𝟏 matrix, which will tell how a particular feature varies with
𝑹(𝒛) = (2)
𝟏 + 𝒆−𝒛 each and every feature. Covariance matrix will be
symmetrical as covariance (f1, f2) = covariance (f2, f1). Once
3.2.1.4 Loss function & Optimisation function
the covariance matrix is calculated, the next step is to find the
Loss functions provide a quantity that is sought to be eigenvalues and eigenvectors. Main eigenvectors are defined
minimised during the training phase of any model. This as the direction in which the variance is maximum.
minimisation is done with the help of some optimization Eigenvectors and Eigenvalues will be founded using Singular
function. We have used BinaryCrossEntropy as the loss Value Decomposition. Eigenvectors will be a unit vector. The

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 528

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

first eigenvector will give first max variance, and the second Principal Component Analysis was applied and at 10th epoch
eigenvector will give second max and so on. Eigenvalues will when Principal Component Analysis was not applied. To
represent how much variance will be explained by a certain overcome overfitting, the orientation of every image is
dimension. In case the eigenvalue of a certain direction is changed before passing it to the model, so that the model
very less, which means it has very less variance or never sees the same image again, for all the datasets that we
information hence it can be removed. have used. Figure 3,Figure 4, Figure 5 shows predictiction of
model on Concrete Crack Images for Classification Dataset,
3.4 Training & Testing of the Model Asphalt Crack Dataset and Automated Bridge Crack
We have applied modified LeNet-5 on the Automated Detection Dataset respectively, without using Principal
Bridge Crack Detection dataset, Concrete Crack Images Component Analysis. Whereas Figure 6, Figure 7, Figure 8
Classification dataset, and Asphalt Crack dataset. Figure 2 shows the prediction of the model on Concrete Crack Images
shows the Detection of Concrete Cracks Flowchart. The idea for Classification Dataset, Asphalt Crack Dataset and
of using modified LeNet-5 was to check the performance of Automated Bridge Crack Detection Dataset respectively, by
LeNet-5 by detecting the cracks on different types of using Principal Component Analysis.
materials. We wanted to make sure that modified LeNet-5
produces consistent results for all types of materials that we 3.5 Predictions on Images
have chosen to work upon. The modified LeNet -5 was
implemented using the Keras framework. 3.5.1 Predictions on images without using Principal
Component Analysis
Figure 3 shows the predictions of the model on
Concrete Crack Images for Classification dataset without
using Principal Component Analysis. Figure 4 shows the
predictions of the model on Asphalt Crack dataset without
using Principal Component Analysis. Figure 5 shows the
predictions of the model on Automated Bridge Crack
Detection dataset without using Principal Component
Analysis.

3.5.2 Predictions on images using Principal Component

Analysis
Figure 6 shows the predictions of the model on
Figure 2. Detection Concrete Cracks Flowchart Concrete Crack Images for Classification dataset using
Principal Component Analysis. Figure 7 shows the
For the Concrete Crack Images for Classification
predictions of the model on Asphalt Crack dataset using
dataset, the modified LeNet -5 was trained on different sets
Principal Component Analysis. Figure 8 shows the
of training data with or without Principle Component
predictions of the model on Automated Bridge Crack
Analysis. We trained the modified LeNet-5 models with
Detection dataset using Principal Component Analysis.
8000, 16000, 24000, 32000 and 40000 training sets of
images. We are comparing the accuracy and time, first by
3.6 Locating the cracks in the images
applying Principal Component Analysis before modified
LeNet-5, second by using modified LeNet-5 alone, for all the We first loaded the image of some structure, which had
datasets we have used. We have used Stochastic Batch cracks. Then we resized the image to 1200*1200 pixels. After
Gradient Descent for training all the datasets with the batch reshaping the image, we divided the image into 144 small
size of 32. For the Concrete Crack Images Dataset, the min images, each of size 100*100 pixels. These small images then
delta's value was chosen as 0.01, and the patience value was were passed through our trained model for prediction, which
taken as 4. The Concrete Crack Images dataset was tested on we have implemented from scratch using Keras to detect the
ten epochs. Simultaneously, the Automated Bridge Crack crack and non-crack region. If the model is predicting cracks
Detection Dataset was also given the same value of min delta, for the image passed to it, then we are coloring the region as
patience, and epochs as the Concrete Crack Images Dataset. red; otherwise, it has been colored with green color. Then
Whereas for the Asphalt Dataset, the training was done these all 144 small images were recombined to make 1 full
without using the early stopping function, with the epochs image. Figure 9 below represents the raw and crack
value 30, and it was observed that the model was giving highlighted image.
maximum training and testing accuracy at 29th epoch when

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 529

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

Figure 10. True Positive Prediction

Figure 9. Raw images and crack highlighted image

4. RESULTS AND DISCUSSION

4.1 Metrics used for judging the performance of the model
4.1.1 Precision
It calculates the precision of the predictions with respect
to the labels. The formula for Precision is:
𝑇𝑃(𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛(𝑃) = (4)
𝑇𝑃(𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠) + 𝐹𝑃(𝑓𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠)
Figure 11. True Negative Prediction
4.1.2 Recall
It calculates the recall of the predictions with respect to
the labels. The formula for Recall is:

𝑇𝑃(𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠)
𝑅𝑒𝑐𝑎𝑙𝑙(𝑅) = (5)
𝑇𝑃(𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠) + 𝐹𝑁(𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠)

4.1.3 Accuracy
It calculates how often the predictions are correct. The
formula for Accuracy is:
𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦(𝐴) = (6)
𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
Figure 12. False Positive Prediction
TP, TN, FP, and FN are True Positive, True Negative,
False Positive, and False Negative respectively. In our case,
Positive is when the image doesn’t have cracks and Negative
is when the image does have cracks. True Positive is when
the image doesn’t have cracks and the model predicted it as
non-crack. True Negative is when the image does have cracks
and the model also predicted it has cracks. False Positive is
when the image doesn’t have cracks but the model predicted
that it has cracks. False Negative is when the image does have
cracks but the model predicted that it has no-cracks. True
Positive Prediction, True Negative Prediction, False Positive
Prediction, False Negative Prediction is shown by Figure
Figure 13. False Negative Prediction
10,11,12 13 respectively.
4.2 Training Time for judging the performance of the model

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 530

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

It is the time taken by the model to learn all the

parameters. We have recorded the time taken by the model to
train for all the datasets which are discussed in section 3.1.
Figure 14, 15, 16 and 17 show the accuracy, precision, recall
and training timecurve for Concrete Crack Images for
Classification dataset, with and without using PCA. Figure 18
and 19 show the performance of the model, based on various
parameters (as discussed in section 4.1 & 4.2), for Asphalt
Crack Dataset and Auttomated Bridge Crack Dataset.

Figure 18. Performance on Asphalt Crack dataset

Figure 14. Accuracy Curve

Figure 19. Performance on Automated Crack Detection Data

5. CONCLUSION AND FUTURE SCOPE

Our modified LeNet model was performing well on
Automated Bridge Crack Detection dataset and Concrete
Crack Images for Classification Dataset. While for the
Figure 15. Precision Curve
Asphalt Crack dataset, it was performing average. The PCA
applied on these datasets has enabled us to get our datasets
trained in a shorter time without any considerable loss in
accuracy for the Concrete Crack Images for Classification
dataset and Automated Bridge Crack Detection dataset.
Although the Principal Component Analysis model reduces
training time for the Asphalt Crack dataset, we are also losing
a substantial amount of accuracy. Our modified LeNeT
model was only predicting whether there is a crack or not in
materials but it is not giving any further information about the
type of cracks or causes of the occurrence of cracks, which
Figure 16. Recall Curve contributes to our paper's limitation. Another limitation is that
the LeNet model’s performance was average in predicting
cracks in Asphalt. In our subsequent work, we will try to
eradicate these shortcomings of our paper.

6. ACKNOWLEDGEMENTS
The authors would like to thank the Department of Civil
Engineering, Delhi Technological University (DTU) and
Central Library, DTU for providing access to all the resources
used for our study. The financial support received under the
research project grant (F. No. DTU/IRD/619/2105) of DTU,
New Delhi is thankfully acknowledged
Figure 17. Training Time Curve

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 531

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

Figure 3. Predictions on Crack images for Classification dataset without using PCA

Figure 4. Predictions on Asphalt Crack dataset without using PCA

Figure 5. Predictions on Automated Crack Detection dataset without using PCA

Figure 6. Predictions on Crack images for Classification dataset using PCA

Figure 7. Predictions on Asphalt Crack dataset using PCA

Figure 8. Predictions on Automated Crack Detection dataset using PCA

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 532

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Sustainable Systems [ICISS 2020]
IEEE Xplore Part Number: CFP20M19-ART; ISBN: 978-1-7281-7089-3

REFERENCES Data.mendeley.com<https://data.mendeley.com/data
sets/xnzhj3x8v4/2> (Sep. 27, 2020)
[1] Eisenbach, M., Stricker, R., Seichter, D., Amende,
[14] Xu, H., Su, X., Wang, Y., Cai, H., Cui, K., and Chen,
K., Debes, K., Sesselmann, M., Ebersbach, D.,
X. (2019). "Automatic Bridge Crack Detection Using
Stoeckert, U., and Gross, H. (2017) "How to get
a Convolutional Neural Network". Applied Sciences,
pavement distress detection ready for deep learning?
9(14), 2867
A systematic approach" 2017 International Joint
[15] Trivedi, A. (2015). "Computing in-situ strength of
Conference on Neural Networks (IJCNN)
rock masses based upon RQD and modified joint
[2] Yokoyama, S., and Matsumoto, T. (2017)
factor: Using pressure and damage sensitive
"Development of an Automatic Detector of Cracks in
constitutive relationship". Journal of Rock Mechanics
Concrete Using Machine Learning". Procedia
and Geotechnical Engineering, 7(5), 540-565
Engineering, 171, 1250-1255
[16] A. K. Jha, S. Adhikari, S. Thapa, A. Kumar, A.
[3] Cha, Y., Choi, W., and Büyüköztürk, O. (2017).
Kumar, and S. Mishra, “Evaluation of Factors
"Deep Learning-Based Crack Damage Detection
Affecting Compressive Strength of Concrete using
Using Convolutional Neural Networks" Computer-
Machine Learning” in 2020 Advanced Computing
Aided Civil and Infrastructure Engineering
and Communication Technologies for High
Engineering, 32(5), 361-378
Performance Applications (ACCTHPA), 2020:
[4] Pauly, L., Peel, H., Luo, S., Hogg, D., and Fuentes,
IEEE, pp. 70-74.
R. (2017) "Deeper Networks for Pavement Crack
[17] Tyagi, L., Butola, R., and Jha, A. (2020).
Detection" in 34th International Symposium on
"Mechanical and tribological properties of AA7075-
Automation and Robotics in Construction (ISARC)
T6 metal matrix composite reinforced with ceramic
[5] Hoang, N. (2018). "Detection of Surface Crack in
particles and aloevera ash via Friction stir
Building Structures Using Image Processing
processing". Materials Research Express, 7(6),
Technique with an Improved Otsu Method for Image
066526
Thresholding". Advances in Civil Engineering, 2018.
[18] Özgenel, Ç. (2020). "Concrete Crack Images for
[6] Zheng, M., Lei, Z., and Zhang, K. (2020) "Intelligent
Classification". Data.mendeley.com,
detection of building cracks based on deep learning"
<https://data.mendeley.com/datasets/5y9wdsg2zt/1>
Image and Vision Computing, 103, 103987
(Sep. 27, 2020)
[7] Fan, Z., Li, C., Chen, Y., Wei, J., Loprencipe, G.,
[19] A. Ghimire, S. Thapa, A. K. Jha, A. Kumar, A.
Chen, X., and Di Mascio, P. (2020) "Automatic Crack
Kumar, and S. Adhikari, “AI and IoT Solutions for
Detection on Road Pavements Using Encoder-
Tackling COVID-19 Pandemic”, in 2020
Decoder Architecture" Materials, 13(13), 2960
International Conference on Electronics,
[8] Li, C., and Ai, D. (2017). "Automatic crack detection
Communication and Aerospace Technology, 2020:
method for loaded coal in vibration failure process".
IEEE.
PLOS ONE, 12(10), e0185750
[20] A. Ghimire, S. Thapa, A. K. Jha, S. Adhikari, and A.
[9] Chen, F., and Jahanshahi, M. (2018). "NB-CNN:
Kumar, “Accelerating Business Growth with Big
Deep Learning-Based Crack Detection Using
Data and Artificial Intelligence”, in 2020 Fourth
Convolutional Neural Network and Naïve Bayes
International conference on I-SMAC (IoT in Social,
Data Fusion". IEEE Transactions on Industrial
Mobile, Analytics and Cloud) (I-SMAC), 2020:
Electronics, 65(5), 4392-4400
IEEE, pp. 441-448.
[10] Zhou, S., and Song, W. (2020). "Robust Image-Based
[21] S. Thapa, P. Singh, D. K. Jain, N. Bharill, A. Gupta,
Surface Crack Detection Using Range Data". Journal
and M. Prasad, “Data-Driven Approach based on
of Computing in Civil Engineering, 34(2), 04019054
Feature Selection Technique for Early Diagnosis of
[11] Lin, Q., Wan, B., Wang, S., Li, S., and Fakhimi, A.
Alzheimer's Disease”, in 2020 International Joint
(2019). "Visual detection of a cohesionless crack in
Conference on Neural Networks (IJCNN), 2020:
rock under three-point bending". Engineering
IEEE, pp. 1-8.
Fracture Mechanics, 211, 17-31
[22] S. Thapa, S. Adhikari, A. Ghimire, and A. Aditya,
[12] Zhang, L., Yang, F., Daniel Zhang, Y., and Zhu, Y.
“Feature Selection Based Twin-Support Vector
(2016). "Road crack detection using deep
Machine for the diagnosis of Parkinson’s Disease”, in
convolutional neural network". 2016 IEEE
2020 8th R10 Humanitarian Technology Conference
International Conference on Image Processing (ICIP)
(R10-HTC), 2020: IEEE
[13] A, J., G, T., M S, D., Nair, B., and D. S, H. (2020).
"Asphalt Crack Dataset".

978-1-7281-7089-3/20/$31.00 ©2020 IEEE 533

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on February 13,2021 at 05:17:56 UTC from IEEE Xplore. Restrictions apply.
View publication stats