• No results found

Design and Analysis of Bridge Crack Detection using CNN

N/A
N/A
Protected

Academic year: 2021

Share "Design and Analysis of Bridge Crack Detection using CNN"

Copied!
5
0
0

Loading.... (view fulltext now)

Full text

(1)

Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited

The International Conference on Research Perspectives : IoT in Hybrid Grid Integrated Renewable Energy Sources In association with International Journal of Scientific Research in Computer Science, Engineering and Information Technology ISSN : 2456-3307 (www.ijsrcseit.com)

115

Design and Analysis of Bridge Crack Detection using CNN

Prof. Kanchan V. Warkar, Kalpana B. Lamsoge

M. Tech Department of Computer Science and Engineering, Bapurao Deshmukh College of Engineering, Sewagram, India

ABSTRACT

Even though there have been incidents in the past, it is critical to keep an eye on the bridges in our country or state. The reason for these tragedies is that there is no system in place that will alert people if a bridge is in poor condition when unexpected events such as floods or earthquakes occur. It indicates that the bridge is not in good repair. When this type of condition occurs, the bridge may collapse, resulting in a variety of losses such as accidents, human deaths, and so on. According to a 2016 report by the National Crime Records Bureau, Maharashtra had the second-highest number of deaths (4,237) due to structural failures between 2001 and 2015. All of these figures demonstrate that disregarding structure safety results in human lives being lost. Zig-Bee technology is employed in the present system, and the TCP/IP protocol was utilised, which is suitable for all sorts of bridges. In this article, we'll delve deeper into the methods for detecting bridge cracks.

Keywords : Convolutional Neural Network, Deep Learning, Digital Recognition, Bridge Crack, Image Processing.

I. INTRODUCTION

The density of highway networks has steadily expanded, and large-span bridges have continued to arise, owing to India's rapid development of the transportation industry. However, additional risk issues will undoubtedly arise as a result of the bridge project [1]. These risk factors are likely to have negative impacts on bridges and even cause bridge collapse, putting people's lives in danger and causing property damage. As a result, it is critical to conduct damage detection and early warning of the bridge structure, as well as to determine the health state of the bridge operation in a timely manner. Engineering and academia are presently debating how to diagnose the health of modern bridges. Many bridges require

immediate safety inspections, health assessments, and maintenance strengthening.

People have recognized the necessity of health

diagnosis since the 1950s, but they have been limited

in its use due to the backwardness of early detection

technologies. However, several countries and research

organisations have recognized the necessity and

necessity of studying bridge structural health

diagnostics in recent decades [2]. The use of efficient

methods for evaluating and assessing the health of

existing bridges, repairing and controlling damage,

and implementing long-term safety inspection,

vibration, and damage control systems.

(2)

II. STRUCTURE OF CNN: EXPLAINED

A. Convolutional Neural Network

1) The CNN's Convolution Layer is the most basic but also the most crucial layer. It essentially convolves or multiplies the pixel matrix generated for the supplied picture or object to build an activation map for that image.

2) The fundamental benefit of an activation map is that it stores all of an image's distinctive qualities while also minimizing the quantity of data that needs to be processed. A feature detector is a set of values with which the machine is compatible, and the matrix with which the data is convolved is a feature detector.

3) Distinct feature detector values are used to generate different versions of the image. Back propagation is also used to train the convoluted model in order to ensure that each layer has the least amount of inaccuracy. Depth and padding are set according to the lowest error set.

4) This entails convoluting the image data matrix and then using a feature detector to generate an activation map or a feature map. The values on identical points in the data and feature map, i.e. values with a value of 1 or more than 1, are maintained in convolution, while the rest are deleted. The picture data matrix is compared three times at a time. Depending on the type of CNN employed, the size of the feature detector changes. Convolution in some versions of CNN, for example, is done using 5x5 or even 7x7 scale filters.

Fig 1. Convolutional Layer ©SuperDataScience

B. POOLING

Pooling is an important step to further reduce the dimensions of the activation map, keeping only the important features while also reducing the special invariance. This in turn reduces the number of learnable features for the model.

This helps to resolve the problem of over fitting.

Pooling allows CNN to incorporate all the different dimensions of an image so that it successfully recognises the given object even if its shape is skewed or is present at a different angle. There are various types of pooling like max pooling, average pooling, stochastic pooling, spatial pyramid pooling. Out of them most popular is max pooling.

Max Pooling

Max Pooling takes the highest value from each sub matrix of the activation map and forms a separate matrix from it. Doing this ensures that the learnable features remain limited in number while also preserving the key features of any image. Max Pooling is usually done using a 2x2 filter.

Fig 2. Max Pooling of a Feature Map

©SuperDataScience

III. HOW CNN WORK

A. Input (The Training Data)

Input layer is represented in terms of 3 dimensions i.e.

width, length and height. It is commonly denoted as width*height*depth which is pixels for the image shown in form of a matrix. For example if input is (64x64x3) then width: 64px, height: 64px, depth: 3px.

Depth is mainly used to represent color pictures in

(3)

form of RGB. Commonly Input layer is even and can be divided by 2 multiple times.

B. Filter

Also called as kernels or feature detectors. Feature detection uses commonly uses a small matrix. For example, in ConvNet, first layer is of dimension 5x5x3.

C. Convolved Feature

Also called Activation Map or Feature Map. A filter of defined is moved through the image taking dot product of each sub matrix producing the Output volume.

D. Receptive Field

Region of input matrix similar to size of filter.

E. Depth

The number of filters.

F. Depth Column

Also called fiber, these are a group of neurons that point towards the same receptive field.

G. Stride

Sets the amount of distance stride will shift after each convolution. Having a larger stride gives a smaller output volume. For example a filter having a stride 2 will shift 2 columns after each convolution. It is necessary to set stride in such a way that the output volume is an integer. Smaller strides are used for better results.

H. Zero – Padding

Adding zeroes on the border of the input volume so as to maintain the size of input volume and output volume. Not doing so will result in loss of information on border of the image and dimension reduction leading to low performance,

I. Parameter Sharing

If a certain feature is common in many images, then it is beneficial to look for that feature in all images which is

Done through parameter sharing. However it is an uncommon practice. For example, in a face detection, there is no need to for facial features if the face is positioned at same location in each image.

J. Dilation

Dilation is filters which have spaces in its cells. If, we have one dimension filter W of size 3 and an input X:

Dilation of 0: w [0]*x [0] + w [1]*x [1] + w [2]*x [2].

Dilation of 1: w [0]*x [0] + w [1]*x [2] + w [2]*x [4].

K. ReLU Layer

ReLU layer is used to apply activation function element wise which converts all the negative values to zero Therefore converting the threshold to zero. It does not affect volume or hyper parameters.

L. POOL Layer

Pool Layer performs a function to reduce the spatial dimensions of the input, and the computational complexity of our model. Pool layer function reduces the dimensions of the input which helps in computational complexity of the

IV. Proposed Methodology

The Convolution Neural Network Algorithm accepts an image as an input and follow the steps to differentiate images one from other. The objects of the Convolution Operation are to extract the high- level features such as edges, from the input image.

The process is as follows:

STEP 1: Image preprocessing;

STEP 2: Set up convolution neural network model;

STEP3: Select the convolution kernel function and convolution parameter according to the size of the input picture, and determine the convolution layer;

STEP 4: 200 groups of input networks were randomly selected from three types of images for training;

STEP 5: The remaining 100 groups of pictures in each type of image input network to identify the classication of mine facies map. If the test results are satisfactory continue to

STEP 6. Otherwise return STEP 3, modify parameters;

STEP 7: Output classification Results.

The convolution layer of the network structure

contains 3 feature graphs, and the convolution kernel

size is 5X5. The output result is 3 classes. Before

(4)

training the sample, you need to adjust the processed fracture picture to 32X32 pixel size and convert it to a two-dimensional matrix.

Fig 3. Cracked Surface

EXAMPLE ANALYSIS

According to the bridge cracks, the image samples are divided into three categories: small cracks, large cracks and severe cracks, each of which has 300 images.200 images were randomly selected as input layers after image preprocessing.

The above convolution neural network model was used for training. The pixel was 32

32 3, and the

convolution kernel was 5

5 3. Using the

remaining 100 images as test samples, the system updates the weights every 30 samples. Through repeated tests, it is found that when the average time of a single iteration is 7 s and 200 iterations, the mean square error is less than 0.1. At this point, the accuracy of model classification is more than 90%, and the training results are better.

V. CONCLUSION

Based on the characteristics of the Convolution Neural Network algorithm and the structural characteristics of the bridge engineering, this paper analyzed the practical application value of the image processing technology in the crack identification of bridge structures, and established a crack

classification application system for bridge structures based on the Internet of Things technology. Aiming at the current situation that the bridge crack detection mainly relied on artificial visual inspection and the risk factor was extremely backward, a digital and intelligent detection method was studied to improve the efficiency of bridge safety diagnosis and reduced the risk factor. Firstly, image enhancement, wavelet denoising, image segmentation and other pretreatments will perform on the collected bridge crack images.

VI. ACKNOWLEDGEMENT

We would like to thank the authors, many people who have given their major contribution for doing research on Convolution Neural Network and related technology on bridge crack detection. Prof. K. V.

Warkar for help with coordinating across different time zones, and discussion with Image processing techniques.

VII. REFERENCES

[1]. L. Zhang, G. Zhou, Y. Han, H. Lin and Y. Wu,

"Application of Internet of Things Technology and Convolutional Neural Network Model in Bridge Crack Detection," in IEEE Access, vol. 6, pp. 39442-39451, 2018.

[2]. S. M. Khan, S. Atamturktur, M. Chowdhury, and M. Rahman, ``Integration of structural health monitoring and intelligent transportation systems for bridge condition assessment:

Current status and future direction,'' IEEE Trans. Intell. Transp. Syst., vol. 17, no. 8, pp.

2107-2122, Aug. 2016.

[3]. E. Mehrani, A. Ayoub, and A. Ayoub,

“Evaluation of fiber optic sensors for remote health monitoring of bridge structures,'' Mater.

Struct., vol. 42, no. 2, pp. 183 199, Sep. 2009.

[4]. S. Chen, F. Cerda, P. Rizzo, J. Bielak, J. H.

Garrett, and J. Kovacevic,``Semi-supervised

(5)

multi resolution classification using adaptive graph filtering with application to indirect bridge structural health monitoring,'' IEEE Trans. Signal Process., vol. 62, no. 11, pp. 2879- 2893, Jun. 2014.

[5]. C. R. Farrar, S. W. Doebling, and D. A. Nix,

“Vibration based structural damage identi cation,'' Philos. Trans. Roy. Soc. London A, Math. Phys. Sci., vol. 359, no. 19, pp. 131-149, Sep. 2015

[6]. https://adeshpande3.github.io/The-9-Deep- Learning-Papers-You-Need-To-Know- About.html

[7]. A Walk-through of AlexNet – Hao Gao - https://medium.com/@smallfishbigsea/a-

walkthrough-of-alexnet-6cbd137a5637

[8]. CNN Architectures: LeNet, AlexNet, VGG, GoogLeNet, ResNet and more - https://medium.com/@sidereal/cnns-

architectures-lenetalexnet-vgg-googlenet- resnet-and-more-666091488df5

[9]. Review: GoogLeNet (Inception v1)— Winner of ILSVRC 2014 (Image Classification) - https://medium.com/coinmonks/paperreview- of-googlenet-inception-v1-winner-of-ilsvlc- 2014-imageclassification-c2b3565a64e7

[10]. Understanding and Implementing Architectures of ResNet and ResNeXt for state-of-the-art Image.

https://medium.com/@14prakash/understanding -and implementingarchitectures- of-resnet-and- resnext-for-state-of-the-art-imagecf51669e1624 [11]. Review: ResNet — Winner of ILSVRC 2015

(Image Classification, Localization, Detection) - https://towardsdatascience.com/reviewresnet- winner-of-ilsvrc-2015-image-classification- localizationdetectione39402bfa5d8

[12]. Microsoft Presents: Deep Residual Networks - https://medium.com/@bakiiii/microsoft-

presents-deep residualnetworks-d0ebd3fe5887

References

Related documents

It was possible to improve the production process by applying the simulation technique, which shows that the general objective was achieved. Also, it was possible to

on deeply held misconceptions and fears of people who hold different beliefs, strategies of conceptual change theory, most often found in science education, were

Um fato interessante ´e que na base padr˜ao, com excec¸˜ao do classificador 2-NN, o menor resultado foi obtido utilizando o mesmo bag size (70%) e em todos os casos os pior

engineer can concentrate well on requirement gathering process. The collected requirement would help to make the fulfilled software. If any mismatch happened is

subspecies of Daucus carota were absent in Daucus crinitus. There is a certain variability in phytochemical composition between the two species.

Twenty-five percent of our respondents listed unilateral hearing loss as an indication for BAHA im- plantation, and only 17% routinely offered this treatment to children with

Based on the idea, we have put forward novel routing strategies for Barrat- Barthelemy- Vespignani (BBV) weighted network. By defining the weight of edges as

Whereas some typical drought-intolerant taxa almost disappeared in the cacao Table 3 – Similarity (ANOSIM) based on Sørensen’s similarity index of overall bryophyte