Crack Detection using Deep Learning

© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 1372
Crack Detection using Deep Learning
Dr. Hansaraj Wankhede1, Abhishikth Thul2, Deep Malviya3, Pratik Pathe4, Shivam Kuite5
1,2,3,4,5 GHRCE Nagpur, India
------------------------------------------------------------------------***-----------------------------------------------------------------------
Abstract— Concrete cracks are one of the primary
signs of a structure that would cause major harm to
the entire infrastructure. The conventional method of
addressing cracks is manual inspection. We discovered
during the survey that there are numerous existing
technologies, including approaches based on computer
vision and image processing. There are several
techniques for locating fractures on different surfaces
utilizing manual inputs. Different factors influenced
the performance of traditional models, which provided
results with variable degrees of accuracy. Because
there are many different strategies that can be used to
produce a good crack detection output, the
performance and efficiency were influenced by the
environment. Using a Convolutional Neural Network,
we have constructed a deep learning model for
fracture classification and segmentation. Convolution,
activation, and pooling layers make up the foundation
of convolutional neural networks. These layers
enhance the performance and generalization of neural
networks by extracting picture information,
introducing nonlinearity, and reducing feature
dimensionality. The observed model's accuracy was
between 97 and 98 percent. This model can be used to
keep track of the condition of the concrete surfaces of
bridges, tunnels, and other public transportation
infrastructure.
Keywords— Crack detection, Deep Learning, Image
Processing, Convolutional Neural Networks, Edge
detection, Image Segmentation, Feature Extraction.
Introduction
Finding surface cracks is crucial for maintaining
infrastructure like roads and buildings. Using any of the
processing approaches, crack detection is the process of
locating cracks in structures. Early identification enables
the implementation of preventative steps to stop potential
failure and damage [1]. The conventional method required
a lot of time and labor. By using a range of contemporary
cameras to record crack images, it is possible to get beyond
the lack of manual inspection and complete successful
maintenance. The issue of fracture identification has
recently been solved by modern technology, although it is
not particularly precise because cracks can differ from
structure to structure and be different on different surfaces.
Accurately recognizing various types of cracks is also
difficult because of the intricacy of crack topologies and
noise in collected crack pictures. Pavement crack detection
using image processing is now both affordable and effective
[2][3].
In practice, approaches for hand-crafted feature extraction
from sub-windows based on intensity thresholding, edge
detection, and other image processing techniques are
frequently used [4].Although the efficiency of these systems
in detecting pavement surfaces has been established, there
is still room for development, particularly in the
classification of different pavement fracture kinds. However,
both computer vision and non-computer vision techniques
have difficulties recovering fracture characteristics from
pavement with shadows and complicated background
pavement, as well as extracting cracks using low-level
image cues [5].
Convolutional Neural Networks (CNN), a prominent tool of
the artificial intelligence branch of Deep Learning (DL),
have demonstrated their effectiveness in object detection.
In contrast to conventional machine learning techniques,
deep learning (DL) offers end-to-end classifiers that
internally learn characteristics and can recognise objects
automatically. The recent advancement of graphics
processing units (GPU), which allows for extremely quick
computations, together with this characteristic of DL
algorithms have increased their use in a variety of sectors.
In the instance of crack detection from images, the user
need only supply various photos as input, and any detected
cracks in these photos are returned as output without the
need for any operator involvement [6]. Mendeley's
Concrete Crack Images for is the source of the information
in our model. Our deep learning model has a classification
accuracy of between 97 and 98 percent.
Crack detection is identified and segmented based on a
variety of characteristics because it is an intriguing issue
with numerous challenging parameters. There are different
ways to find cracks. There are several techniques that use
manual inputs to find fractures in a variety of surfaces.
Image processing methods using cameras.
This section provides a summary of the processing methods
used to identify cracks in engineering structures using
camera images. Many of the publications that were
reviewed here used camera input.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 05 | May 2023 www.irjet.net p-ISSN: 2395-0072
I. LITERATURE REVIEW

The information regarding the fracture is gathered through
pre-processing, image segmentation, and feature extraction.
Following the smoothing of the acceptable input image, the
Threshold method of segmentation was applied. The size
and perimeter of the roundness index are determined for
picture justification. The presence of the crack in the image
is then assessed in comparison [7]. Many commercial
camera-based image processing methods simply require
pre-processing, however some methods focus on the
integration algorithm, which is where feature extraction is
doneThe faults are represented quantitatively in a model
that was created. The fracture quantification and detection,
neural network, and 3-D visualization models, respectively,
make up the integration model [8].
A stitching algorithm for images was created. For the
retrieval of the crack segments, that has been utilized,
which uses a feature-based registration and skeletonization
technique. The evaluation of the crack quantification model
served as the sole foundation for the width and length-
based crack detection. An integrated model that combined
fracture length and change detection with 3D crack pattern
visualization and neural network prediction of crack depth
was developed [1].
Image Processing Methods
Preprocessing is a phase used in image processing
techniques to reduce noise and improve crack
characteristics. To identify cracks, it uses threshold
segmentation. The morphological approach method finds
fracture formations by evaluating curvature and using
mathematical morphology. The neighbourhood pixels are
used to estimate the size of the crack in the percolation-
based approach to identify whether a focus pixel is a part of
it. While the automatic route-finder algorithm calculates
the length and width of the cracks in the practical technique,
the endpoints of the crack are manually identified. As the
initial stage in enhancing the crack detection algorithm's
performance, all IP approaches preprocess photos. There
has been a lot of research done on these IP approaches'
accuracy [1].
The majority of studies concentrate on non-intrusive
techniques like feature extraction and thresholding. The
lack of uniformity in particular IP techniques has been
highlighted, nevertheless. Thresholding, for instance, has
relied on extremely basic techniques like a fixed 0.5 or
halving the maximum image brightness; this differs with
more sophisticated techniques like Otsu and Niblack [9].
Edge detection
Concrete cracks are one of the primary signs of a structure
that would cause major harm to the entire infrastructure.
The conventional method of addressing cracks is manual
inspection. Manual procedures omit the quantitative and
qualitative analysis. Experiments with automatic fracture
identification utilizing image processing techniques yield
reliable results. In more recent publications, a wide variety
of image-processing approaches have been tested. For the
purpose of identifying concrete cracks, a thorough analysis
of several edge detection methods is done. Prewitt, Sobel,
Gaussian, Roberts, Canny, and Laplacian edge detector
performance is discussed [18].
A break in the intensity field of the greyscale is referred to
as an ideal edge. In order to discover image cracks more
quickly and effectively within a big picture dataset, edge
detection refers to the employment of filters in an image
processing algorithm to detect or enhance the cracks in an
image. Matrix representations of images are used in
mathematics [20].
A discontinuity in the intensity level is referred to as an
edge. Methods for crack identification use filters in the
spatial or frequency domain to locate edges. These
techniques outperform manual examination in terms of
reliability. The following stages are generally followed by
all edge detector-based fracture detection techniques:
 Smoothing
 Enhancement
 Detection
 Binarization
Types of Edge detector algorithm:
 Sobel Edge Detector
 Canny Edge Detector
 Prewitt
 Robert
 Laplacian of Gaussian (LOG)
 Fast Fourier Transform (FFT)
This suggests that visual assessment of concrete structures
by people is a more time- and money-intensive task.
Concrete structure cracks are automatically detected,
reducing the need for human intervention. This study
examines various edge detection techniques' analyses of
concrete crack detection. We investigate two frequency
domain-based edge detector approaches and four spatial
domain-based edge detector methods. The outcomes from
the edge detection techniques were fairly satisfactory. It
should be expanded using deep learning techniques in
order to achieve higher performance in conventional edge
detection techniques [18].

Traditional ML methods:
Crack evaluation has always been done manually using
human field surveys. These manual survey methods,
however, have low repeatability and reproducibility,
require a lot of time and labour, and put surveyors in
danger. Additionally, due to subjectivity of the raters, the
data collected may differ. There have been significant
research efforts to develop automated crack survey
methods to address the drawbacks of manual surveys. Data
collection, crack detection, and crack diagnosis make up the
automatic crack evaluation.
The crack detection method receives two- or three-
dimensional (2D or 3D) input data from data capture. Crack
detection is the process of identifying cracks in the 2D
and/or 3D data provided by automated data capture, with
the least amount of human participation. Crack diagnosis
requires key inputs from crack detection, including
classification data like type, severity, and extent that are
critical for planning maintenance. Through the
management process, crack diagnosis can be utilized to
choose the best time and intensity of therapy [11].
Before training the models, traditional ML techniques use a
predetermined feature extraction stage. Support vector
machines (SVM), artificial neural networks (ANN), random
forests, clustering, Bayesian probability, and naive Bayes
fusion are the most popular conventional ML techniques.
Both crack identification and image processing with the
best parameters have been accomplished using
conventional ML algorithms. These are also frequently
employed for the extraction of predetermined features.
They are unable to handle massive datasets, though.
Traditional ML techniques have a drawback in that they are
unable to learn more complicated features and cannot
handle the complex data included in photos, such as
backdrop pavement with varying lighting [9].
In our approach, crack segmentation is formulated as a
binary image labelling problem, where "Positive" and
"Negative" stand for, respectively, "crack" and "non-crack."
If there is a crack in the output image, deep learning will be
used to further segment it. A task like this necessitates both
high-level features and low-level clues. Convolution,
activation, and pooling layers make up the foundation of
convolutional neural networks (CNNs). These layers
enhance the performance and generalization of neural
networks by extracting picture information, introducing
nonlinearity, and reducing feature dimensionality.
Layers used in convolution, activation, and pooling are the
foundation of CNNs. These layers enhance the performance
and generalisation of neural networks by extracting picture
information, introducing nonlinearity, and reducing feature
dimensionality. Convolution layers in CNNs are used mostly
for feature extraction, which also maintains the spatial
relationship between pixels in the source pictures. The
output of the convolution kernel or filter is computed by
sliding the image by pixels, performing element-wise
multiplication, and adding the pixels. The feature map is the
name of the output matrix. A collection of 2D arrays of
weights is what makes up a convolution kernel or filter in
deep learning. Each 2D array of weights is used to generate
additional channels by applying it to all of the input
channels from the preceding layer.
While feature maps of higher layers include more semantic
meaning and less location information, lower convolution
layers primarily extract structural elements and location
information from an image. An error backpropagation
technique can be used to learn and improve the convolution
kernels with weights during training. A CNN is one of the
several kinds of artificial neural networks that are used for
different applications and data sources. It is also a sort of
network architecture for deep learning algorithms and is
specifically utilised for tasks like image recognition and
pixel data processing. An activation function is used to
provide nonlinearity to neural networks after the
convolution process. Rectified Linear Units are frequently
used activation functions.
CNN Architecture:
There are two main parts to a CNN architecture
 Feature extraction is a procedure that uses a
convolution tool to separate and identify the
distinct characteristics of an image for study.
 There are numerous pairs of convolutional or
pooling layers in the feature extraction network.
 a fully connected layer that makes use of the
convolutional process's output and determines the
class of the image using the features that were
previously extracted.
 This CNN feature extraction model seeks to
minimize the quantity of features in a dataset. It
generates new features that condense an initial set
of features' existing features. The number of CNN
layers is numerous, as seen in Figure 1.
II. METHODOLOGY
Predefined feature extraction requires image-processing
procedures, and diverse features have been applied in
various research projects, including statistical values of
images, vertical and horizontal projections of feature maps,
and characteristics of specified crack objects. The primary
issue with conventional ML approaches is that they only
use superficial learning strategies. Those approaches can't
handle the complicated information in the photos without
learning higher-level features. For instance, lighting and the
environment have a significant impact on the background
of photos of pavement.

Figure-1: CNN Architecture
[Source: Binary Image classifier CNN using TensorFlow -
Medium.com]
An overview of how to use a CNN to find cracks is shown in
Figure 2 below. The dataset for concrete cracks, modelling,
and testing the trained CNN classifier are the three
processes. Many raw photos of the concrete surface are
obtained in order to train this model. The gathered raw
images are reduced in size by cropping, and the reduced
images are then manually differentiated into photographs
with and without cracks. The training set and validation set
are then arbitrarily chosen from the dataset and loaded into
a CNN for training and validation, respectively. A CNN
classifier is created through training that can distinguish
between images with and without cracks. Cracks in photos
can be distinguished from others using the trained CNN
classifier.
Figure-2: Model Workflow
Model Training
 The number of epochs can be increased to improve
accuracy.
 Here, learning pace is equally crucial. Training at
various learning speeds can lead to better results.
However, this significantly lengthens the training
period.
 If a low learning rate is chosen, training moves
more slowly; if a high learning rate is chosen,
training moves faster but accuracy suffers.
Classification Report
 The function for creating classification reports
generates a text report with the key classification
metrics.
 The ratio of true positives to the total of true and
false positives is the definition of precision for each
class..
 Recall is determined for each class as the
proportion of true positives to the total of true
positives and false negatives.
 Since F1 scores incorporate precision and recall
into their computation, they are less accurate than
accuracy measures.
Tools /
Platform
Description
Miniconda
A free basic installer for conda is called
miniconda. It is a stripped-down,
bootstrapped version of Anaconda that
just contains conda, Python, and the
packages they require, as well as a select
few other important packages like pip, zlib,
and a few more..
CUDA
NVIDIA’s CUDA is a general-purpose
parallel computing platform and
programming model that accelerates deep
learning and other compute-intensive apps
by taking advantage of the parallel
processing power of GPUs.
Google
Colab
Google Research produces Google
Colaboratory, sometimes known as
"Colab." Colab is particularly useful for
machine learning, data analysis, and
education because it enables anyone to
write and run arbitrary Python code
through the browser.
Jupyter
The first web application for producing
and sharing computational documents was
called a Jupyter Notebook. It provides a
straightforward, efficient, document-
focused experience.
Command
Prompt
For the operating systems OS/2,
eComStation, ArcaOS, Microsoft Windows,
and ReactOS, it serves as the default
command-line interpreter.
VSCode
On your desktop, Visual Studio Code is a
quick yet effective source code editor that
runs on Windows, macOS, and Linux.
III. DATA COLLECTION & PLATFORMS USED

CLASSIFICATION IMAGES OF CONCRETE CRACK
Images of diverse concrete surfaces, both with and without
cracks, can be found in the datasets. For image classification,
the picture data are split into two categories: negative
(without crack) and positive (with crack). Totaling 40000
photos with 227 × 227 pixels and RGB channels, each class
comprises 20,000 images. 458 high-resolution photos are
used to create the dataset (4032x3024 pixel). High levels of
variance were discovered in high resolution photos [19].
Step-1: Setting the Dataframe
Images that have cracks are kept in a category called
"POSITIVE," whereas those without cracks are kept in
"NEGATIVE."
Step-2: Training Period
We have a fully convolutional model that has been trained
and fed the aforementioned dataset.
We installed the prerequisites, imported the required
libraries, and handled the project's dependencies to ensure
proper implementation. The dataset including images with
and without cracks is initially input to the model. The initial
dataset was split into 60 percent Training data and 40
percent Testing data, with the epochs set at 10. The model
successfully classifies the Test pictures into two groups
when each epoch is complete: "Crack Detected" and "No
Crack Detected."
Figure-3: [Images are classified into ‘Crack Detected’ and
‘No Crack Detected’]
Step-3: Model Summary
Using Keras, a neural network is created in this stage. ReLu
and sigmoid activation functions were used for these layers
because they produce more accurate results for the project
at hand.
Image Segmentation
The model has been trained and can now label the new
photos. We'll be dealing with photos that have cracks in
them. Any image can be used as an input to the model.
Figure-4: [Initial Image as an input]
For this, the images with cracks are saved into a folder
called positive. The model will now be transmitted with an
image that includes a crack. An appropriate size change will
be made to this image. Red dots, also known as scatter
marks in the matrix, will be used to label the whole image
after adjustments. In the next step, the red dots are reduced
so that the features, Crack marks are covered with red dots.
Finally, this marked part will be colored red.
The final output generated is an image that is classified into
a crack label with crack highlights.
Figure-5: [Image with crack highlights in red]
IV. DATASET
V. DESIGN, IMPLEMENTATION AND MODELLING

Image Processing technique:
a model that applies multiple image processing methods,
including grayscale image conversion, hue scale, and other
methods. These methods are employed to draw attention to
the entire image and extract the features. The next stage is
to use a clever edge detection method after image
processing. The multi-stage technique Canny Edge
Detection determines the Edge Gradient from the images to
mark the outlines.
𝐸𝑑𝑔𝑒 𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡 (𝐺) √𝐺𝑥
2
𝐺𝑦
2
This method had a problem with photos that were noisy,
overexposed to light, and blurry. The following is a list of
the results that the image processing techniques produced.
Figure-6: [Image with image processing technique]
In the testing phase, we used an image that contains a
crack. Example-
Figure-7: [Initial Image for testing]
The model will categorise this image as having been
cracked when it enters the first stage. This image is now
utilised in the following stage, where the image's crack will
be highlighted and indicated in red. Any image can be
provided to the model as an input. This image has a label
indicating whether it is cracked or not. The model that will
mark the crack is now given the image.
Figure-8: [Image with crack highlights in red]
Confusion Matrix
The figure below mentions the confusion matrix produced
using 10 epochs. The actual positive and negative numbers
are highest. The mistake rate is significantly reduced. After
being used as an input by the model, the testing starting
image is accurately classified as having a Crack.
The number of images that actually had no cracks and the
anticipated value were both correctly characterized, hence
the algorithm was able to estimate True Values accurately
for 1206 images.
However, 1147 of the photos genuinely had crack and
were accurately classified.
Figure-9: [Confusion Matrix]
VI. RESULTS

Figure-10: [Training and Validation Loss over Time]
Classification Report
Figure-11: [Classification Report]
The model's accuracy was between 97 and 98 percent, and
it can recognize photos with cracks. The generated f1-
score is 98 percent. This indicates that the model is
prepared to assess additional photos.
The classification of photos into "Crack detected" and
"Crack not Detected" is successful. A random image
provided as input is reliably and precisely labelled. The
model correctly classified the image labels with an
accuracy of 97–98%. Further processing is applied to the
cracked images, and the cracks are segmented. Cracks are
first designated with a red color before being accentuated.
The output is a picture with a segmented crack as its final
form. A multi-class algorithm called Edge detection is
implemented as part of a similar strategy for image pre-
processing in order to identify the image's cracks. When
compared to image processing methods, deep learning
algorithms do better in segmenting images. Image
processing methods, blur, noise, and lighting conditions all
have downsides. To classify a crack and then separate the
crack from the image, a deep learning model learns how to
label various sorts of images.
[1] Crack detection using image processing: A critical
review and analysis Arun Mohan, Sumathi
Poobal Gurudeva Institute of Science and Technology
(GISAT), Kottayam, Kerala KCG College of Technology,
Chennai, India (2017)
[2] Wang X and Feng X 2011 Pavement Distress Detection
And Classification With Automated Image Processing
in Proceedings 2011 International Conference on
Transportation, Mechanical, and Electrical
Engineering, TMEE (2011)
[3] Broberg P 2013 Surface Crack Detection In Welds
Using Thermography NDT E Int.
[4] Zou Q Cao Y, Li Q, Mao Q, and Wang S 2012 Cracktree:
Automatic Crack Detection From Pavement Images
Pattern Recognit. Lett.
[5] Li B Wang K C P Zhang A Yang E, and Wang G 2018
Automatic Classification Of Pavement
[6] Dais, D., Bal, İ. E., Smyrou, E., & Sarhosis, V. (2021).
Automatic crack classification and segmentation on
masonry surfaces using convolutional neural networks
and transfer learning. Automation in Construction,
125, 103606.
[7] Zhang Yiyang. The design of glass crack detection
system based on image pre-processing technology, in:
Proceedings of Information Technology and Artificial
Intelligence Conference, (2014).
[8] R.S. Adhikari, O. Moselhi1, A. Bagchi, Image-based
retrieval of concrete crack properties for bridge
inspection, Autom. Constr. 39 (2014).
[9] Golding, V.P.; Gharineiat, Z.; Munawar, H.S.; Ullah, F.
Crack Detection in Concrete Structures Using Deep
Learning (2022).
[10] Image-based concrete crack detection in tunnels
using deep fully convolutional networks Yupeng Ren,
Jisheng Huang b, Zhiyou Hong c, Wei Lu b, Jun Yin b,
Lejun Zou a, Xiaohua Shen (2020).
[11] Hsieh, Y.-A., & Tsai, Y. J. (2020). Machine Learning
for Crack Detection: Review and Model Performance
Comparison. Journal of Computing in Civil
Engineering.
[12] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning
representations by backpropagating errors, Nature
323 (1986).
[13] V. Nair, G. Hinton, Rectified linear units improve
restricted Boltzmann machines, in: Proc. 27th Int.
Conf. Mach. Learn., (2010).
[14] A.L. Maas, A.Y. Hannun, A.Y. Ng, Rectifier
nonlinearities improve neural network acoustic
models, in: Proc. Int. Conf. Mach. Learn (2013).
VII. CONCLUSION
REFERENCES

[15] D. Scherer, S. Behnke, Evaluation of pooling
operations in convolutional architectures for object
recognition, in: K. Diamantaras, W. Duch, L.S. Iliadis
(Eds.), Artif. Neural Networks – ICANN 2010, Springer,
Berlin, Heidelberg,Thessaloniki, Greece, 2010.
[16] H. Simon, Neural network: a comprehensive
foundation, 2nd ed., Prentice Hall, 1998.
[17] Dimitris Daisabİhsan Engin BalaEleni Smyroua
Vasilis Sarhosisb. Automatic crack classification and
segmentation on masonry surfaces using
convolutional neural networks and transfer learning
(2021).
[18] A. Diana Andrushia, N. Anand and I. Antony
Godwin Karunya Institute of Technology and Sciences,
Coimbatore, India C. Aravindhan Bannari Amman
Institute of Technology, Sathyamangalam, India.
Analysis of edge detection algorithms for concrete
crack detection (2018).
[19] Özgenel, Çağlar Fırat (2019), “Concrete Crack
Images for Classification”, Mendeley Data, V2
[20] Dorafshan, Sattar; Thomas, Robert J.; Maguire,
Marc (2018). Comparison of deep convolutional neural
networks and edge detectors for image-based crack
detection in concrete. Construction and Building
Materials.

Crack Detection using Deep Learning

More Related Content

What's hot (20)

Similar to Crack Detection using Deep Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

Crack Detection using Deep Learning