Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems (ICICCS 2021)

IEEE Xplore Part Number: CFP21K74-ART; ISBN: 978-0-7381-1327-2

Convolutional Neural network based Online

Rail surface Crack Detection
CHITTURI AKHILA CHEEMALA ASTHA DIAMOND

B.E, Computer Science and Engineering, Sathyabama B.E, Computer Science and Engineering, Sathyabama
Institute of Science and Technology Chennai, India. Institute of Science and Technology Chennai, India.
2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS) | 978-1-6654-1272-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICICCS51141.2021.9432080

[email protected] m [email protected]

A. Mary Posonia
Associate professor, Computer Science and Engineering
Sathyabama institute of Science and Technology Chennai, India
[email protected]

Abstract- The proposed system utilizes CNN to detect frequent crack with double thresholds. Then the
the faults in the railway tracks with the help of images. algorithm based on an outline of the shape of the
Their characteristics are obtained to extract the default fracture is used to consider then, the characteristics
railway track. This model helps to reduce the manual
of the fracture and finally attain the exact crack
inspection work. While performing the task, it lowers
location[6]. Finally determining the crack faults
the time and cost utilization and increases the safety of
passengers. To train and validate the CNN algorithm, level based on the space occupied of cracks is put
the ground truth databases on images of masks are forward. The algorithm improves the accuracy of
utilized. To this end, this CNN maintains the both space railway crack detection effectively under
and time coherence of the defects in the tracks, and the convoluted conditions.
final output of the false prediction effectively.
This process is trained by the ground truth
Keywords: Convolution neural networks, ground truth, dataset by the CNN algorithm. By, training this
crack detection, deep-learning, prediction. algorithm reduces in the false prediction of railway
crack images and these will be divided mainly into
I. INTRODUCTION two categories – defective or non-defective CNN
algorithm is performed[3].
In many countries, the majority of people
are depending on the railways for transportation.
Railways were highly encouraged for rapid II. LITERATURE SURVEY
development. Although the frequent improvement
Convolution neural networks [5] for steel have
of railway technology decreases the occurrence of
proposed the progress in this field of neural
internal defects, if a crack or any defects in the
networks was maintained by many achievements.
railway tracks are not detected may lead to heavy
Enabling the architectures of deep learning by
loss in human life. So far, the defects are detected
increasing the power of computing and provide
by few methods namely automatic detections and
higher and more levels of representation. Data
characterization of cracks [1], manual inspection,
augmentation is a powerful tool for not allowing
physical detection, and automatic crack detection
the overfitting in this proposed method. Error rates
by using the computer vision method[2]. In this
can be able to reduce by committee methods by
proposed system, we first perform the histogram
several networks combination especially the
equalization mainly it is used to reduce the
uncorrelated in the individual predictions.
illumination effect on the images of railway tracks
Unsupervised methods for representations and
used in the detection process[4]. Following to the
learning features became very popular and solve
next step we use well defined Canny edge
problems with supervised training. The
algorithm for detection is used for drawing a
convolutional neural network (CNN) is mainly

978-0-7381-1327-2/21/$31.00 ©2021 IEEE 1602

Authorized licensed use limited to: Zhejiang University. Downloaded on November 18,2024 at 21:47:37 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems (ICICCS 2021)
IEEE Xplore Part Number: CFP21K74-ART; ISBN: 978-0-7381-1327-2

designed for recognizing. CNN is performed The type of the fault is determined by considering
outstandingly including both MNIST and SVHN the features from signals of variance for faulty
datasets. Broken Rail Detection System using RF images.
[2] Technology is proposed. Mainly this system is
used to detect the broken rail and provide the a) Advantages of Proposed system
information exactly about the tracks. To prove this It works well with high dimensionality along with a
model practically over the wooden plank the large dataset also reduces the consumption of
metallic track is simulated, and these will be hardware resources. Improve prediction accuracy
divided into fewer sections and these will be effectively.
numbered and joined with metal plates. If any of
these joints are removed and isolated, the Initially data is loaded; the process will start from
information will be quickly passed to the remote image processing. The image will be read, and size
unit. Firstly, the alarm is raised by the receiver for is adjusted, all denoise will be eliminated. Then
the staff alert after that the tracking number, which segmentation followed by morphology will take
was broken will be displayed by LCD with a place. Where segmentation is a process of dividing
microcontroller. The receiver is located near the image visuals into smaller segments and
station to alert the staff of the station and passes the morphology is a process of collecting image
information to the trains which are traveling in that structures. With all these information image will be
route. This entire process is maintained by railway filtered with necessary features and segmented
signalling. crack will be identified.

A review on computer vision-based defect

detection and condition assessment of concrete and
asphalt civil are proposed [7]. This model presents
a synthesis of the art of computer vision this mainly
explains the categories of the methods in computer
vision methods, they are used to automate the
damage and process detection. The computer vision
is pipeline starts processing from a lower to a
higher level. These methods cant be considered
completely separately. Extracted features were
studied to support the process of classification
based on pattern methods. afterward, it is classified
how the methodologies are used, validated, tested
to find the unique defect and observing the damage
in patterns by remote and morphology in the
images of concrete surfaces, underground concrete
pipes, precast concrete tunnels, and asphalt
pavements.

III. PROPOSED METHODOLOGY

In the proposed system, the image processing and

deep neural network-based CNN model is used for
detecting the faulty images in the railway crack
detection [7]. This inspection process is performed Fig 3.1 Architecture of Proposed System
by collecting the images of railway tracks by Otsu
segmentation model mainly used in railway surface When the dataset is loaded the image processing
detection. When variance is calculated by rail will take place by reading, resizing and delete the
surface image the signal is obtained it will all noises in the images, divide into defined
determine whether there is any fault in the categories and finally predict whether there is any
particular image of the railway track surface [1]. default or not in the image of railway track[1].

978-0-7381-1327-2/21/$31.00 ©2021 IEEE 1603

Authorized licensed use limited to: Zhejiang University. Downloaded on November 18,2024 at 21:47:37 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems (ICICCS 2021)
IEEE Xplore Part Number: CFP21K74-ART; ISBN: 978-0-7381-1327-2

b) Module Description This relation between input image and the

processed output image can also be represented
as.
i) Image Preprocessing
s = T (r)
The railway surface is predicted by collecting
Where, r is actually the pixel value or grey level
various railway surface images from rail tracks. So intensity of f(x,y) at any point. And s is the pixel
here, for the railway surface detection the OTSU value or grey level intensity of g(x,y) at any point.
segmentation method is used.[2]. The signal
The basic grey level transformation has been
acquired by calculating variance of obtained rail discussed in our tutorial of basic grey level
surface image. By considering the obtained signal it transformations.
will determine whether there is any fault on railway
surface. This faults being extracted the features
from faulty images variance signals [1].
ii) Feature Extraction
The rail surface is detected by collecting the rail
surface images from the railway. Otsu Feature extraction provides two functions, they are
segmentation method is used in the rail surface data compression, invariance. Feature vector
detection process [2]. The signal is acquired by contains selected features, they are extracted from a
calculating the variance of the obtained rail surface sample set of measurements. The feature step is to
image. By taking the signal obtained, it is eliminate redundancy in representation. A
determined whether there is a fault on the rail cautiously selected vector should contain most of
surface. The type of fault is determined by the preferred information and it should be smaller
extracting features from variance signals belonging in dimensions compared to the master copy of the
to faulty images signal vector. This improves the accuracy of a
classification and also reduces the computational
Border detection is calculated by the grayscale of effort
depth image sum for every row and the disparity
ratio of two rows. When the ratio is greater than a iii) Prediction with proposed project
threshold, that row will be separated border table
and floor areas The principal of this methodology is to detect the
cracks on the railway track with pictures. This
The transformation in the brightness modifies the model is used for image processing. Mainly
brightness in the pixel also, the transformation segmentation is carry out in view of the density
completely depends on the properties of the pixel. pixels of the image. Considering K number of class
and (k-1) Threshold pixel value and the image pixel
Generally, brightness transformation includes
K is divided
brightness corrections and greyscale
transformations. Brightness correction takes pixel After all the segmented cracks in the data set are
brightness into account by its location or position trained and validated.
of the image and the original brightness and
modified. Grayscale transformations change the The algorithm trained an entire dataset is passed
brightness despite the image position. forward and backward until we get correct
predictions
Consider this equation
The class is the prediction model (CNN) When
G(x,y) = T{ f(x,y) } class is less than 0.5 it returns having fault in the
In this equation, particular railway track. When the class is more
than 0.5 then there is no fault in the railway track
F(x,y) = input image on which transformation
function has to be applied.
G(x,y) = the output image or processed image.
T is the transformation function.

978-0-7381-1327-2/21/$31.00 ©2021 IEEE 1604

Authorized licensed use limited to: Zhejiang University. Downloaded on November 18,2024 at 21:47:37 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems (ICICCS 2021)
IEEE Xplore Part Number: CFP21K74-ART; ISBN: 978-0-7381-1327-2

proposed system concept we can achieve a full

completed desired output by training the images
collected from the data set.

IV. RESULT AND DISCUSSION

Defected cracks are calculated by using the ratio of

the correct cracks found to the number of cracks
Fig 3.2 : Data set training images in values
detected. Recollect how many actual cracks are
identified. This will be calculated by the ratio of
From the fig 3.2, the data set was collected by correct cracks to the actual number of cracks.
using real time data image processing, Here, the
images are classified into defective and non- As we know that we had used the CNN model to
defective by using these images we are training the create this concept of prediction we should divide
machine.
the images into train and validation dataset at
which the images with faults and no faults. By
training the model several times we get correct
predictions whether the images of tracks containing
any faults or not and these return the image with
defective and non-defective.

Fig 3.3 : Prediction using Defective Images

Figure 3.3 provides all the defective images and

training the machine for desired output. Here in the
fig 3.3 shows three bars blue bar indicates the
prediction and training was done only once, orange
bar indicates the prediction and training was done
ten times, finally the grey bar indicates that the
prediction and training was done thirty or more
times to get the desired output. The efficiency of
prediction is increasing every time.

FIG 4.1 Defect Identified – No Defect

Fig 3.4 : Prediction using Non-Defective Images

Figure 3.4 has taken all the non-defective images

from the dataset and training the machine for
desired output. In each step we can observe that the
training was done more times and we are getting
the expected output from the machine. By using

FIG 4.2 Defect identified - Defect

978-0-7381-1327-2/21/$31.00 ©2021 IEEE 1605

Authorized licensed use limited to: Zhejiang University. Downloaded on November 18,2024 at 21:47:37 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems (ICICCS 2021)
IEEE Xplore Part Number: CFP21K74-ART; ISBN: 978-0-7381-1327-2

http://www.sciencedirect.com/science/article/p
ii/S1474034615000208
6. K. Kang et al., “T-CNN: Tubelets with
Several images are collected in the data set and by convolutional neural networks for object
using the inception v3 and CNN model will return detection from videos,” IEEE Trans. Circuits
the images with or without defects. Syst. Video Technol., to be published, 2016.
7. O. Russakovsky et al., “ImageNet Large Scale
Visual Recognition Challenge,” Int. J. Comput.
Vis., vol. 115, no. 3, pp. 211–252, Dec. 2015.
V. CONCLUSION 8. . Soukup and R. Hu er- ork,
“Convolutional neural networks for steel
This paper has used the main detection model Mask surface defect detection from photometric
R-CNN to identify the cracks on the solid surfaces. stereo images,” in Proc. Int. Symp. Visual
This model is trained by using ground truth dataset, Comput., Dec. 2014, pp. 668–677.
9. L. Zhang, F. Yang, Y. D. Zhang, and Y. J.
where the main reference of this dataset permits the
Zhu, “Road crack detection using deep
picture data to be related to the actual features and convolutional neural network,” in Proc. IEEE
materials on the solid surfaces. And this model is Int. Conf. Image Process., Sep. 2016, pp.
trained and validated based on crack detection 3708–3712.
algorithms. In order to use this framework for 10. .-J. Cha, . Choi, and O. uy uk ozt urk,
different surfaces .and conditions in future it will “ eep learning-based crack damage detection
be trained again with large dataset, which helps to using convolutional neural networks,”
Comput.-Aided Civil Infrastructure Eng., vol.
reduce the false predictions.
32, no. 5, pp. 361–378, Mar. 2017.
11. S. J. Schmugge et al., “ etection of cracks in
nuclear power plant using spatial- temporal
VI. REFERENCES grouping of local patches,” in Proc. IEEE
Winter Conf. Appl. Comput. Vis., Mar. 2016,
pp. 1–7.
1. Matsumoto, T., et al. "Several image 12. Sireesha R, Ajay Kumar B, Mallikarjunaiah G
processing examples by CNN." IEEE and harath Kumar , “ roken Rail etection
International Workshop on Cellular Neural System using RF Technology”, in “SSRG
Networks and their Applications. IEEE, 1990. International Journal of Electronics and
2. Grosse, Roger, et al. "Ground truth dataset and Communication Engineering”, volume 2 issue
4- April 2015
baseline evaluations for intrinsic image
algorithms." 2009 IEEE 12th International
Conference on Computer Vision. IEEE, 2009.
3. Pizer, Stephen M., E. Philip Amburn, John D.
Austin, Robert Cromartie, Ari Geselowitz,
Trey Greer, Bart terHaarRomeny, John B.
Zimmerman, and Karel Zuiderveld. "Adaptive
histogram equalization and its
variations." Computer vision, graphics, and
image processing 39, no. 3 (1987): 355-368.
4. H. Oliveira and P. L. Correia, “Automatic road
crack detection and characterization,” IEEE
Transactions on Intelligent Transportation
Systems, vol. 14, no. 1, pp. 155–168, March
2013, doi:
https://doi.org/10.1109/TITS.2012.2208630.
5. C. Koch, K. Georgieva, V. Kasireddy, B.
Akinci, and P. Fieguth, “A review on computer
vision based defect detection and condition
assessment of concrete and asphalt civil
infrastructure,” vol. 29, no. 2, 2015, pp. 196 –
210, infrastructure Computer Vision. [Online].
Available:

978-0-7381-1327-2/21/$31.00 ©2021 IEEE 1606

Authorized licensed use limited to: Zhejiang University. Downloaded on November 18,2024 at 21:47:37 UTC from IEEE Xplore. Restrictions apply.