Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)

IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

Railway Bridge Inspection using CNN

Lakshmi Narasimham Chennareddy Sai Vamsi Gandabathula Vivek Vardhan Jasthi
Dept.of Information Technology Dept.of Inforamtion Technology Dept.Of Information Technology
Velagapudi Ramakrishna Siddhartha Vleagapudi Ramakrishna Siddhartha Velagapudi Ramakrishna Siddhartha
Engineering College Engineering College Engineering College
Vijayawada,India Vijayawada,India Vijayawada,India
[email protected] [email protected] [email protected]

Fathimabi Shaik
2023 7th International Conference on Computing Methodologies and Communication (ICCMC) | 978-1-6654-6408-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCMC56507.2023.10083695

Dept.of Information Technology

Velagapudi Ramakrishna Siddhartha
Engineering College
Vijayawada,India
[email protected]

Abstract— The key issue for the railway department has requirements. Prev iously, such problems were solved using
been to examine and monitor railway bridges, as urbanization various methods and technologies outside of India.
expands, the availability of railways grows, and the railway
1.2 Basic Definitions and Background
system has greatly expanded throughout the nation. The
1.2.1 CNN:
expense of maintaining railroad bridges and associated costs
with personnel have been a burden on the railroads. To ensure
CNN (Convolutional neural networks) is a deep learning
transportation safety, concrete bridge crack detection is algorith m used to identify the image by comparing with the
critical. Deep learning technology has made it possible to other related images or previous images. The algorith m used
automatically and accurately detect faults in bridges. The here is the CNN sequential model. Different layers will be
present methods are not accurate and they require a large size added to the model and then the model will be trained with
of dataset for model training and they require a high the labelled data.
computational power model training. The proposed model is a
convolutional neural network (CNN) based end-to-end crack
detection model. The proposed model achieved a 95%
detection accuracy.

Keywords— Remote sensing, OpenCV, Keras,

TensorFlow.

I. INTRO DUCTIO N
This research work provides a general overview o f
its objectives, origins, and applications. It outlines the
requirements and scope of the research work and includes Fig 1.1 CNN Algorithm
stoppers, wh ich specify the conditions under which the 1.2.2 Image processing:
research work is appropriate. With technology such as Initially, the image is processed using grey and negat ive
remote sensing, it is now possible to monitor bridges even image p rocessing. It is observed that the cracks are
without physically present and receive results promptly. The identified in both the image processing. Image p rocessing
analysis focuses mainly on model-based finite element plays a crucial ro le in identification o f cracks in the
update techniques, non-model-based (data-driven) defect bridges. It is necessary to avoid those types of railway
detection techniques such as artificial neural networks, and problems. Image processing helps in the identification o f
structural health monitoring techniques based on Bayesian cracks in the structure.
belief networks. The main focus is on issues related to 1.2.3 VGG 19:
bridge cracking. VGG19 is a deep learning arch itecture developed by
Cracks on concrete surfaces are one of the earliest signs of Oxfo rd Un iversity in 2014. It features 19 layers, including
structural degradation, which is crucial fo r maintenance. 16 convolutional and 3 fully connected layers. Th is model
Continued exposure can lead to significant environ mental is often used fo r image classificat ion tasks and has been
damage. The rev iew p rovides analysis based on image pre-trained on the ImageNet dataset for t ransfer learning .
processing techniques, objectives, accuracy levels, erro r VGG-19 is recognized for its high accuracy in image
levels, and image datasets. Finally, this paper present recognit ion and its use of s mall convolut ional filters,
various research issues that can be useful for future research resulting in a large nu mber of parameters. Although its
on crack detection. accuracy is impressive, the model requires a substantial
1.1 Origin of the Problem amount of co mputational power, making it challenging to
Railways play a crucial role in the GDP o f the deploy in resource-limited env iron ments. Despite these
Indian economy. It has been identified that the Indian limitat ions, VGG-19 remains a favored choice fo r image
Railways recently mentioned a problem in their portal, classificat ion because of its performance and the
which included both technical and non-technical availability of pre-trained weights.

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 416

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)
IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

InSAR can be utilized to monitor transition zones' health

1.3 Problem Statement with Objectives and Outcomes conditions [2].
1.3.1 Problem Statement: The authors presented a methodology for
The population is growing, and the use of railroads identifying bridge cracks using remote sensing. For training,
as a service has increased dramatically in the modern world. they used satellite image data. In order to find usable,
The railway department has gained a lot of insight thanks to quantitative indicators of bridge condition, remote sensing
the railroad bridge inspection. In recent years, it has suffered technologies are helpful. The data gathered in this way can
numerous accidents and damages, which has become be included into an approachable, web-based DSS created
problematic. Th is paper presents a railway bridge inspection using free software tools that operate with Pontis data
management system that operates using deep learning. The structures, providing a framewo rk that any state
proposed model enables the railway to quickly and transportation agency can employ. The DSS presented in
accurately monitor the bridges using a reliable CNN. The this study is meant to serve as an illustration of a cutting -
goal is to overcome the d ifficu lties in identifying damaged edge asset management tool for bridge condition evaluation
bridges and monitoring the structure of Indian railway [3].
bridges. The authors proposed a deep learning model fo r
1.3.2 Objectives: automatic crack detection in concrete bridges. The model is
The objective is to develop an efficient and cost-effective based on Convolutional Neural Netwo rks (CNNs) and uses
model for identifying damaged railway bridges in India techniques such as atrous convolution, Atrous Spatial
using image processing and CNN sequential model. To Pyramid Pooling (ASPP) and depth wise separable
achieve the best results with high accuracy in order to convolution to improve performance. The proposed model
overcome current limitat ions and limitations of tradit ional was found to have a detection accuracy of 96.37% with out
methods. The current methods are less accurate and pre-train ing and to perform better than other classification
expensive. models. It can also be used as an effective feature extraction
1.3.3 Outcomes: structure in other CNNs [4].
By the end of this work, we aim to have a model that can In this study the researchers provide different
accurately detect cracks in bridges with high accuracy and automatic crack detection using various image processing
low cost. The proposed model will make it easy to detect methodologies. The automatic crack detection helps to
cracks in bridges. replace the slower trad itional human inspection processes.
1.4 Real-time Applications of Proposed work The various convictions faced in the image detection are the
1. Railway bridge inspection provides a clear p icture irregularity in the size of cracks and illu minating conditions,
about the problems that they occur in the railway shading in the images. The image processing methodology
bridges. And it also helps to provide a solution for provides accurate results in comparison with tradit ional
that problem. processes. The crack quantification model and the image
2. Hu man efforts can be a waste of money and the acquisition pre-processing are some of the methods. The
time during the times/conditions like floods above methods detect the cracks with the curvature
because cracks can’t be identified by the cameras evaluation and mathematical morphology techniques . The
etc. methods with hybrid sy stem are got an accuracy 96%
3. Some of them p roposed a decision supported [5].
system to monitor the railway bridge condition. For maintaining the structural health of a bridge, it
is important to detect the structural damages of the bridge
II Review of Literature over time to t ime. In this study they created an algorithm fo r
crack detection. They developed a STRUM classifier model
2.1 Description of Existing Systems for crack detection. They used robotic imaging fo r
The literature review focuses on the materials that collecting the data about the bridge deck. W ith STRUM
helped understand the image processing and classification classifier they got 95% accuracy in detection of cracks and
techniques research field. The research articles discussed the many typical image-based approaches only got 69%
use of various algorithms and frameworks for data accuracy [6].
classification. This section focuses on research publications,
including the main details of each paper, and presents the 2.2 Summary of Literature Study
findings and conclusions from each study.
They developed a deep learning model for auto matic crack
A fracture detecting model was described by the
detection for bridges. They used algorithms convolutional
authors. For the purpose of detecting cracks, they put forth
neural networks and Faster RCNN. They developed a bridge
an RCNN model. Three main stages of the methodology —
condition decision support system (DSS), a web-based asset
data gathering, 3D construction, damage detection, and
management tool for bridge managers and inspectors.
analysis—are suggested for studying bridge damage. The
Django framework is used to develop the web portal. These
newly created deep learning crack segmentation model and
measurements are able to detect mm-level changes in the
toolbox have proven useful for UA V-assisted bridge
track geometry with a bi-weekly measurement update and a
inspection. They achieved an accuracy of 90% [1].
high spatial resolution by using interfero metric synthetic
This study illustrates the viability of applying
aperture radar (InSAR) techniques.
InSAR to railway track changeover zones. InSAR
measurements were taken in transition zones close to a steel
bridge. The results that have been given have shown that

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 417

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)
IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

III PRO PO SED METHO D 3.3.1 CNN ALGORITHM

3.1 DESIGN METHOLOGY  CNN is a deep neural network algorith m. It is most
The model is trained using pre-processed photos fro m widely used to classify the images.
bridge image and other datasets with a 6:1 training to test  CNN model is trained with a labeled data. It is
dataset ratio. The architecture includes three convolutional trained by using image dataset. It extracts the
layers, optional batch normalization, t wo max pooling feature of every image in the dataset.
layers, and ReLU activation functions. The channel count in  In a Convolutional Neural Netwo rk (CNN),
the first convolution layer is increased fro m 1 to 64, and the features of an image are extracted through a series
second layer's channel count is raised to 128. The pooling of layers that process the input image.
layer helps in summarizing the features generated by the  Convolutional Layer: The convolutional layer
convolution layer. The next step is to determine the point performs convolution operations on the input
where a steady increase in epoch values leads to a rise in image to extract features.
loss value and stop training the model. The model is then  Pooling Layer: The pooling layer used to retain the
created and trained with increasing epoch values and most important features.
adjusting the checkpoints as needed. To assess the model's  Activation Layer: The activation function allowing
accuracy, the number of correct predictions is counted and it to learn more complex features. Ex: Relu
the percentage of accurate predictions is calculated. The  Dense Layer: The dense layer is used for
metrics used for model analysis are precision, recall, f1 - classification. It takes the features extracted by the
score, and support. The accuracy measures the frequency of previous layers and learns to map them to the
correct predictions, while recall measures the number of desired output.
positive samples the model can identify. ReLU and Dense  The feature extract ion will be done by the
layer activation functions are commonly used in CNN repetitive steps of the algorithm.
models.  It takes image as a input and extracts the feature of
The CNN Mathematical Notation as follows: the image.
lout=[lin+2p-k]/s  The output is derived based on the feature value of
where lout is output features, lin is input features, p is the given input image.
padding size, s is stride size. 3.3.2 Algorithm steps
• Pre-processing of Image dataset (Image to array
and image is resized to (224,224)). And splitting
3.2 SYSTEM ARCHITECTURE the data into trained and testing data.
A CNN model is used for railway bridge inspection • Creat ing a convolutional base with 3 convolutional
to detect cracks in images. The model is trained using an and 3 pooling layers.
image dataset available on GitHub. In data preprocessing, • A convolutional base is a component of a
the images will be resized to a width of 224 and a height of Convolutional Neural Net work (CNN) that
224. The preprocessed data will be used for train ing the performs convolution operations on input image
model. During image processing, the image is div ided into data.
smaller parts with a width of 224 and height of 224. If the • It consists of mu ltip le convolutional layers,
image size is too large, it will be compressed. activation functions, and pooling layers.
The parts of the original image are passed to the • The convolutional base ext racts feature fro m the
model, wh ich detects cracks in the image. If any part input image through the convolutional layers and
contains a crack, the output will be labeled as "cracked." applies activation functions to produce feature
The parts that contain cracks are displayed in the output maps.
along with the corresponding negative image. The original • The pooling layers then reduce the spatial
image is also displayed in the output with its corresponding dimension of the feature maps to reduce
negative image. computational complexity.
• The extracted features are then passed on to the
fully connected layers for classification o r
regression.
• The image is passed to the convolutional layer.
• The output when the image passes through a
Convolutional layer: -
Convolution Output dimension = [(I - F +2 *P) / S]+1 x
D
where I->dimensions of the image
F->size of the kernel/filter
S->Strides
P->Padding
Fig 3.2. System Architecture D->Depth
• Next the image is passed to the pooling layer.
3.3 DESCRIPTION OF ALGORITHMS • The output when an image passes through a
Pooling (Max) layer: -

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 418

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)
IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

For a pooling layer, one can specify on ly the filter/kernel

size (F) and the strides (S).
Pooling Output dimension = [(I - F) / S] + 1 x D
• Next image is passed to relu layer.
• The output when an image passes through a relu
dense layer: -
ReLU formula is: f(x) = max (0, x)
• Model is co mpiled with loss
sparse_categorical_crossentropy, optimizer adam,
metrics accuracy.
To get the accuracy the training and testing of the model is
fitted using batch size and validation data.

Fig 4.1 Graph between epochs and accuracy

3.4 DATASET DESCRIPTION
The dataset consists of total 4380 images and The model is trained with different sizes of train data. And
cracked images are 3726 and non-cracked are 1104. Th is is got the highest accuracy when the train and validation data
the crack data set used in the article "Automated Bridge ratio is 80:20.
Crack Detection Using Convolutional Neural Net works".
The authors placed the dataset in their GitHub. The dataset Table – 2 Different split ratios and accuracy
is split in the rat io of 80:20. The t rain dataset contains a
total of 3864 images and the validation dataset contains a
total of 966 images. The dataset used in the research gives
the best results for the desired model. And the dataset is
split into different sizes of train and validation data to
analyze the model. The dataset us ed in this paper is
available at https://github.com/tjdxxhy/crack-detection.git.

IV RESULTS AND OBSERVATIO N

4.1 Result analysis
CNN algorithm:
When the training and validation data rat io is 80:20, the
The noted observations are displayed in the below
train data contains a total of 3864 images and validation
table.These are met rics got with the in itial neural network
data contains a total of 966 images.
topology.
Table – 1 Epochs and accuracy values
VGG19 Algorithm:
VGG19 is a transfer learn ing model. It is a pre trained
model consisting of 19 layers and which quite complex.
And it requires high computational power and more time
for model training.

Initially trained the model with epoch value 8 and got an

accuracy 88%. Later by using the model analysis graph
trained the model with 16 epochs. And the accuracy is
increased to 95%. It is proved that when don’t know the
correct epoch value to train the model it is better to use the
model analysis graph.The suitable epoch value is needed in
order to acquire good performance.

Fig 4.2 Graph between epochs and accuracy

The VGG19 model is t rained with the same

dataset used to train the CNN. And by above graph,
conclude that the model loss it’s accuracy at epoch value

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 419

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)
IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

10. The model is trained with the train and validation data Fro m the observations conclude that CNN gave best results
ratio is 80:20. The t rain dataset contains 3864 images and in terms of accuracy and training time.
validation dataset contains 966 images.

Test case result: 4.2 Observations from the work

The path of the image is g iven as an input to the model and
the cracks are detected in the given input image. Using a The main focus is on detection of cracks. By this
CNN model achieved an accuracy of 95%. The proposed the inspection of the bridge will becomes easy and
model is accurate and fast compared to other transfer accurate. The model detects the cracks more accurate than
learning models. the human.
CNN has been more well-known as a tool in recent years
due to their proficiency in effect ively identifying and
classifying image flaws. The model had a good accuracy of
95% in observation. CNN inspection processes may easily
be automated, cutting down on manual inspection time. It
had demonstrated its reliability in producing outcomes.

V CO NCLUSIO N AND FUTURE WO RK

5.1 CO NCLUSION
A system based on deep learning was developed
for routine inspection of railway bridges to reduce the cost
and effort required for bridge inspections. The model uses a
Fig 4.3 Crack detection Convolutional Neural Network (CNN) algorith m to analy ze
the properties of the input dataset to produce accurate
results. This model outperforms the traditional radar remote
sensing model in terms of accuracy. The development of
this low-cost model offers a p ractical solution for railway
bridge inspection services.

5.2 Future Work

Without any effort fro m the user, a web portal is
used for railway bridge inspection utilizing deep learning to
look for bridge damage. The state of the b ridge is very
simp le for the user to check. By doing this, nu merous issues
Fig 4.4 CNN model accuracy
that are being brought about by poor management of the
state of the bridges would be resolved. The best field of
study for researchers using neural networks has proven to be
deep learning. In the future, able to extend this work by
leveraging the big data technology spark, which is faster and
has distributed datasets that are resilient.

REFERENCES

[1] Ayele, Yonas Zewdu, Mostafa Aliyari, David Griffiths, and Enrique
Lopez Droguett. "Automatic crack segmentation for UAV-assisted bridge
Fig 4.5 VGG 19 model accuracy inspection." Energies 13, no. 23 (2020): 6250.

In this research work, the co mparison of model performance [2] Wang, Haoyu, Ling Chang, and Valeri Markine. "Structural health
based on accuracy suggests that the CNN with 95% monitoring of railway transition zones using satellite radar
accuracy is a better model than the VGG19 with 80% data." Sensors 18, no. 2 (2018): 413.
accuracy. However, it is important to take into consideration [3] Endsley, Arthur, Colin Brooks, Devin Harris, T ess Ahlborn, and
other factors such as computation cost, training time, and Khatereh Vaghefi. "Decision support system for integrating remote sensing
generalization to new data. The simp ler and faster CNN is in bridge condition assessment and preservation." In Sensors and Smart
easier to train, deploy, and is well suited for research works Structures Technologies for Civil, Mechanical, and Aerospace Systems
2012, vol. 8345, pp. 1181-1195. SPIE, 2012.
that prioritize these aspects. On the other hand, VGG19,
with its deeper architecture, is better equipped to capture [4] Xu, Hongyan, Xiu Su, Yi Wang, Huaiyu Cai, Kerang Cui, and
complex features in data and excel in tasks that require high Xiaodong Chen. "Automatic bridge crack detection using a convolutional
accuracy. Ultimately, the choice of arch itecture should be neural network." Applied Sciences 9, no. 14 (2019): 2867.
based on a balanced evaluation of accuracy, co mputational
[5] Mohan, Arun, and Sumathi Poobal. "Crack detection using image
resources, and other relevant factors. processing: A critical review and analysis." Alexandria Engineering
Journal 57, no. 2 (2018): 787-798.

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 420

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 7th International Conference on Computing Methodologies and Communication (ICCMC-2023)
IEEE Xplore Part Number: CFP23K25-ART; ISBN: 978-1-6654-6408-6

[6] Prasanna, Prateek, Kristin J. Dana, Nenad Gucunski, Basily B. Basily,

Hung M. La, Ronny Salim Lim, and Hooman Parvardeh. "Automated crack
detection on concrete bridges." IEEE Transactions on automation science
and engineering 13, no. 2 (2014): 591-599.

[7] Ali L, Alnajjar F, Jassmi HA, Gocho M, Khan W, Serh ani MA.
Performance evaluation of deep CNN-based crack detection and
localization techniques for concrete structures. Sensors. 2021 Mar
1;21(5):1688.

[8] Hammouch W, Chouiekh C, Khaissidi G, Mrabti M. Crack Detection

and Classification in Moroccan Pavement Using Convolutional Neural
Network. Infrastructures. 2022 Nov 10;7(11):152.

[9]Sathya, K., Sangavi, D., Sridharshini, P., Manobharathi, M., &

Jayapriya, G. (2020). Improved image based super resolution and concrete
crack prediction using pre-trained deep learning models. Journal of Soft
Computing in Civil Engineering, 4(3), 40-51.

[10] Chang, Ling, Rolf PBJ Dollevoet, and Ramon F. Hanssen.

"Nationwide railway monitoring using satellite SAR interferometry." IEEE
Journal of Selected Topics in Applied Earth Observations and Remote
Sensing 10, no. 2 (2016): 596-604.

[11] Li, Ruoxian, Jiayong Yu, Feng Li, Ruitao Yang, Yudong Wang, and
Zhihao Peng. "Automatic bridge crack detection using Unmanned aerial
vehicle and Faster R-CNN." Construction and Building Materials 362
(2023): 129659.

[12] Inam, Hina, Naeem Ul Islam, Muhammad Usman Akram, and Fahim
Ullah. "Smart and Automated Infrastructure Management: A Deep
Learning Approach for Crack Detection in Bridge
Images." Sustainability 15, no. 3 (2023): 1866.

[13] Li, Haotian, Hongyan Xu, Xiaodong T ian, Yi Wang, Huaiyu Cai,
Kerang Cui, and Xiaodong Chen. "Bridge crack detection based on
SSENets." Applied Sciences 10, no. 12 (2020): 4230.

[14] Xu, Hongyan, Xiu Su, Huaiyuan Xu, and Haotian Li. "Autonomous
bridge crack detection using deep convolutional neural networks." In 3rd
International Conference on Computer Engineering, Information Science
& Application Technology (ICCIA 2019), pp. 274-284. Atlantis Press,
2019.

[15] M. S. Saeed, "Unmanned Aerial Vehicle for Automatic Detection of

Concrete Crack using Deep Learning," 2021 2nd International Conference
on Robotics, Electrical and Signal Processing Techniques (ICREST ),
DHAKA, Bangladesh, 2021, pp. 624-628, doi:
10.1109/ICREST51555.2021.9331177.

[16] L. Zhang, G. Zhou, Y. Han, H. Lin and Y. Wu, "Application of

Internet of Things T echnology and Convolutional Neural Network Model
in Bridge Crack Detection," in IEEE Access, vol. 6, pp. 39442-39451,
2018, doi: 10.1109/ACCESS.2018.2855144

978-1-6654-6408-6/23/$31.00 ©2023 IEEE 421

Authorized licensed use limited to: ULAKBIM UASL - Anadolu Universitesi. Downloaded on September 17,2023 at 16:10:59 UTC from IEEE Xplore. Restrictions apply.