Application of neural network method for road crack detection

TELKOMNIKA Telecommunication, Computing, Electronics and Control
Vol. 18, No. 4, August 2020, pp. 1962~1967
ISSN: 1693-6930, accredited First Grade by Kemenristekdikti, Decree No: 21/E/KPT/2018
DOI: 10.12928/TELKOMNIKA.v18i4.14825  1962
Journal homepage: http://journal.uad.ac.id/index.php/TELKOMNIKA
Application of neural network method for road crack detection
Yuslena Sari, Puguh Budi Prakoso, Andreyan Rizky Baskara
Faculty of Engineering, Universitas Lambung Mangkurat, Indonesia
Article Info ABSTRACT
Article history:
Received Jun 6, 2019
Revised Feb 27, 2020
Accepted Apr 3, 2020
The study presents a road pavement crack detection system by extracting
picture features then classifying them based on image features. The applied
feature extraction method is the gray level co-occurrence matrices (GLCM).
This method employs two order measurements. The first order utilizes
statistical calculations based on the pixel value of the original image alone,
such as variance, and does not pay attention to the neighboring pixel
relationship. In the second order, the relationship between the two pixel-pairs
of the original image is taken into account. Inspired by the recent success
in implementing Supervised Learning in computer vision, the applied method
for classification is artificial neural network (ANN). Datasets, which are used
for evaluation are collected from low-cost smart phones. The results show that
feature extraction using GLCM can provide good accuracy that is equal
to 90%.
Keywords:
ANN
Crack detection
Feature extraction
GLCM
Image
This is an open access article under the CC BY-SA license.
Corresponding Author:
Puguh Budi Prakoso,
Faculty of Engineering, Universitas Lambung Mangkurat, Indonesia
National Center for Sustainable Transportation Technology, Indonesia.
Email: puguh.prakoso@ulm.ac.id
1. INTRODUCTION
Evaluating road quality is an important task in many countries. For example, in France, national roads
are examined every three years to estimate reparations, required construction, and to estimate quality.
These aspects can be calculated to estimate crack on road pavement surface based on compliance with
regulations, micro texture, macro texture, and surface degradation. In Banjarmasin, Indonesia, the condition
of subgrade, which mostly consists of peat is the major origin of road crackings. Considering the vast wetlands
of South Kalimantan reaching 382,272 ha is a major problem to be tackled. All road evaluation inspections in
Indonesia are done manually, whereas in other countries automatic systems have been performed with
non-invasive techniques such as image processing [1-3].
There are several ways to detect road damage including analyzing road conditions based on images
taken on the road surface. Research on detection of road surface damage using image processing techniques
has been actively carried out, achieving very high detection accuracy. Many studies only focus on detecting
the presence or absence of damage. However, in a real-world scenario, when road managers from
the management agency need to repair such damage, they need to clearly understand the type of damage to
take effective action. In addition, in many previous studies, researchers obtained their own data using different
methods. Therefore, there is no uniform road damage dataset available openly, which causes no benchmark for
detection of road damage [4, 5].
This research will propose an intelligent system in image processing to detect cracks on road surfaces.
To examine road infrastructure efficiently, especially road conditions, several methods using laser technology or
image processing have been studied [6-9]. However, the technique of using a laser requires a high cost [10, 11].

TELKOMNIKA Telecommun Comput El Control 
Application of neural network method for road crack detection (Yuslena Sari)
1963
Futhermore, at a lower cost, there were several studies using computer vision approach for solving this problem
in 11 years [12-17]. In computer vision, it is first necessary to extract features from the dataset to obtain
quantitative data from image data, then the next process is to detect road pavement cracks. Research conducted
by E. Zalama, G. Jaime, and R. Medina [4] demonstrated that using Bank Gabor filter for feature extraction
methods in detecting road pavement cracks, the accuracy obtained is 90%. The Bank Gabor filter method
represents the image in a one-angle orientation scale, while the GLCM represents it in 4 angles namely 00
, 450
,
900
, 1350
, so that the extraction of the resulting features will represent the extracted image more.
For the detection stage, monitoring the process is considered as one of the most important tasks in
the detection process. Many researchers utilize Supervised Learning in the detection process, especially
the ANN method. A road detection strategy using the ANN method for the classifier was introduced by
M. Mokhtarzade, H. Ebadi, and M. J. Valadan Zoej, where the dataset was satellite imagery [5]. Researches
done by I. Kahraman, M. Kamil Turan, and I. Rakip Karas also apply ANN in detecting road cracks with results
that the ANN method is able to detect road cracks with 93.35% success [18]. Referring to those researches,
this study proposes ANN method in detecting road crack and also applies GLCM to extract features from
images into quantitative data.
2. RESEARCH METHOD
To detect road surface cracks, features of a road cracking are required. The features are shape feature
and texture feature, where these features can be used to distinguish road conditions [19, 20]. Figure 1 shows
the stages of the road surface crack detection process. Further explanation of Figure 1:
Figure 1. Proposed road surface crack detection stages
a. Road image data
The collected data is labelled as a cracked road image. Road images are taken using low-cost
smartphone camera, where the distance between the road surface and the camera is 1 meter in front
of the camera. Data is collected from some roadway sections in Banjarmasin, which consists of two types,
namely cracked road and non-cracked road images. Figure 2 presents an example data of road surface
with cracks.
b. Data pre-processing
In the pre-processing stage, image data is segmented first. Segmentation is the process of separating
objects contained in an image, which aims to ease the processing of digital image on each object.
Then thresholding, which is the process of changing a grayscale image into a binary or black and white image.
The goal of thresholding is to see clearly which areas are included in the object and in the background of an
image. Next step is morphology. Morphology is a digital image processing technique which uses shapes as
a reference in processing the image. The value of each pixel in a digital image is obtained from the results
of a comparison between the corresponding pixels in the digital image and the adjacent pixels. Morphology
operations depend on the order of pixels and do not pay attention to the value of pixels; thus, this technique
can be used to process binary images and grayscale images.
c. Feature extraction
Feature extraction is applied to retrieve assessment information from the analysis and calculations
performed on digital images [21, 22]. The results of this extraction have a significant effect on the results
of the classification later. The feature extraction process is carried out using Matlab software, where the GLCM
method is applied in the feature extraction of road pavement crack detection. GLCM utilizes texture
calculations in the second order. Measurement of textures in the first order assigns statistical calculations based
on the pixel value of the original image, such as variance, and does not pay attention to neighboring pixel

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 18, No. 4, August 2020: 1962 - 1967
1964
relationships. In the second order, the relationship between pairs of two original image images is taken into
account [23].
For example, f (x, y) is an image of size Nx and Ny that has pixels with the possibility of L levels
and 𝑟⃗⃗ is a spatial offset direction vector. 𝐺𝐿𝐶𝑀 𝑟(𝑖, 𝑗) is defined as the number of pixels with 𝑗 ∈ 1, . . , 𝐿 that
occurs in offset 𝑟⃗⃗ to the pixels with values 𝑖 ∈ 1, . . , 𝐿, which can be stated in the (1) [24].
𝐺𝐿𝐶𝑀 𝑟(𝑖, 𝑗) = #{(𝑥1, 𝑦1), (𝑥2, 𝑦2) ∈ (𝑁𝑥, 𝑁𝑦) × (𝑁𝑥, 𝑁𝑦)|𝑓(𝑥1, 𝑦1) = 𝑗 𝑟
= (𝑥2 − 𝑥1, 𝑦2 − 𝑦1)⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ } (1)
In this case, offset 𝑟⃗⃗ can be an angle and/or distance. For example, the Figure 3 shows the four directions
for GLCM.
Figure 2. Road surface cracking image Figure 3. Example directions for GLCM
with angles 00
, 450
, 900
, dan 1350
d. Road detection
Road crack detection can be done after the extraction of road image features. The data will be divided
into two parts, namely training data and testing data. After that the classification of crack roads and good roads
will be carried out using a machine learning approach which is artificial neural network (ANN) method. ANN is
a processor that carries out large-scale distribution, which has a natural tendency to store a recognition that has
been experienced, in other words ANN has the ability to be able to do learning and detection of an object [25].
e. Result evaluation and validation
In performance measurement using confusion matrix is used to measure how well the detection
performance of the ANN method is to recognize tuples from different classes. TP and TN provide information
when the detection results are true, while FP and FN tell when the values are false [4, 26]. Then after
the confusion matrix obtained, Accuracy, Precision and Recall value can be calculated. The Accuracy value is
obtained by (2). The Precision value is obtained by (3). The Recall value is obtained by (4) [27, 28]:
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
𝑥100% (2)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
𝑥100% (3)
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
𝑥100% (4)
where:
 TP is true positive, which is the amount of positive data that is properly classified by the system.
 TN is true negative, which is the amount of negative data that is properly classified by the system.
 FN is false negative, which is the amount of negative data but classified incorrectly by the system.
 FP is false positive, which is the amount of positive data but is classified incorrectly by the system
3. RESULTS AND ANALYSIS
3.1. Data processing
Image data will be cropped first to uniform all data. The amount of data is 100, which is labeled crack
and no_crack. The data to be processed is a dataset of 256x100 pixels. Then feature extraction will be performed
to all images. Examples of dataset are depicted in Figure 4.

1965
(a) (b)
Figure 4. Example of dataset with (a) crack and (b) no crack
3.2. Feature extraction
Feature extraction utilizes the gray level co-occurrence matrix (GLCM), which applies five quantities,
namely angular second moment (ASM), contrast, inverse different moment (IDM), entropi, and
correlation [29-31]. Examples of the results from feature extraction can be seen in Table 1. The values from
the feature extraction will be the input parameters in detecting road cracks.
Table 1. Data of feature extraction result
No Label ASM Contrast IDM Entropy Correlation
1 crack 3,37E+11 1,78E+17 0.123868916860676 8.515.177.221.615.980 9,10E+11
2 crack 3,54E+11 1,46E+17 0.148823309991595 8.419.985.894.605.640 9,11E+11
3 no_crack 4,22E+11 1,51E+17 0.133652331513828 8.235.894.329.981.770 0.001376067169563
4 no_crack 8,06E+11 1,25E+17 0.170254651932088 7.773.661.510.499.480 0.001538142413451
5 crack 2,37E+11 4,37E+17 0.084879178580005 8.854.083.850.684.690 7,78E+11
3.4. Result and evaluation
Before training the data, the learning rate is set 0.01 with momentum 0.9. The network architecture
model which is obtained from the training dataset produces a network architecture with 5 inputs, namely: ASM,
contrast, IDM, entropy, correlation; as well as with 5 hidden layers. The network architecture is shown
in Figure 5. As for the weights of hidden layer are shown in Table 2. These weights were generated from
the training data results of 500 epoch repetitions. The network architecture model brings about two groups
of output, namely crack and no crack. Table 3 shows the resulting weight outputs, meanwhile Table 4 exhibits
the assessment results.
Figure 5. The best network architecture from training datasets
Table 2. The weight of hidden layer
Hidden Layer Threshold ASM Contrast IDM Entropy Correlation
1 -0.088 0.064 0.002 -0.149 -0.139 0.072
2 -0.756 0.132 0.132 -0.470 -1.541 -2.232
3 0.569 -0.638 0.085 0.394 1.307 1.696
4 -0.214 0.289 -0.093 -0.563 -1.011 -0.609
5 0.708 -0.847 0.030 0.298 1.496 2.627

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 18, No. 4, August 2020: 1962 - 1967
1966
Table 3. The weight output
Output
Layer
Node1 Node2 Node3 Node4 Node5 Threshold
Crack -0.280 -2.177 1.322 -1.056 2.057 -0.602
No_crack 0.274 2.158 -1.299 1.098 -2.072 0.588
Table 4. The Assessment results
Measurement Results
Accuracy 90.00%
Precision 93.50%
Recall 87.50%
The assessment results in Table 4 demonstrate that ANN Backpropagation method reaches 90%
accuracy for the detection of road damage, which can be categorized as high accuracy level. It can also be seen
that the precision value is higher than the accuracy value, namely 93.50%. Meanwhile, the recall value has
the lowest value, which is 87.50%. The highest accuracy is obtained with a value of 90%. This is because
the data is well prepared, and the data used is data from crack and non crack data images (good asphalt).
No other road damage data is presented. The value of precision is 93.50% which is higher than the accuracy.
This means that the accuracy of the detection area is higher than the accuracy of the detection. The recall result
is 87.50%. this result is categorized as good. This result obtained from total correct data detection each class
(crack and non crack data) divided by all data classified correctly.
4. CONCLUSION
The proposed approach of road crack detection is able to identify cracks with an accuracy of 90%.
The image processing technique applies feature extraction using the GLCM method. This method produces
image feature extraction from four angles namely 00
, 450
, 900
, and 1350
. The experimental results demonstrate
that a set of 4 angles, consisting of properties derived from projective integrals and crack object properties
cooperates to achieve the most accurate prediction. In addition, the inclusion of characteristics of cracked
objects such as ASM, contrast, IDM, entropy and correlation has been proven to provide more information for
classification. This fact is exhibited through the good experimental results. ANN is a supervised learning
approach that has been implemented to study the mapping function between the image input and output features
of crack and no crack classifications. Based on the experimental results, ANN can be concluded as a competent
classification method. Thus, the application of ANN integrated with GLCM feature extraction method is highly
recommended for the detection of road pavement cracks. Future work should also investigate the applications
of different algorithms and other feature extraction methods in order to improve prediction accuracy. Last but
not least, it is necessary to collect more image datasets to improve the ability of the current road pavement
crack detection model.
ACKNOWLEDGEMENT
This paper is supported by USAID through Sustainable Higher Education Research Alliances
(SHERA) Program Centre for Collaborative (CCR) National Center for Sustainable Transportation
Technology (NCSTT) with Grant No. IIE00000078-ITB-1.
REFERENCES
[1] T. H. Nguyen, A. Zhukov, and T. L. Nguyen, “On road defects detection and classification,” Supplementary Proceedings
of the Fifth International Conference on Analysis of Images, Social Networks and Texts, pp. 264–275, 2016.
[2] A. Ragnoli, M. R. De Blasiis, and A. Di Benedetto, “Pavement Distress Detection Methods : A Review,”
Infrastructures, vol. 3, no. 58, pp. 1–19, 2018.
[3] T. B. J. Coenen and A. Golroo, “A review on automated pavement distress detection methods,” Cogent Engineering,
vol. 4, no. 1, 2017.
[4] E. Zalama, G. Jaime, and R. Medina, “Road Crack Detection Using Visual Features Extracted by Gabor Filters,”
Computer-Aided Civil and Infrastructure Engineering, vol. 29, no. 5, pp. 342–358, 2014.
[5] M. Mokhtarzade, H. Ebadi, and M. J. Valadan Zoej, “Optimization of road detection from high-resolution satellite
images using texture parameters in neural network classifiers,” Canadian Journal of Remote Sensing, vol. 33, no. 6,
pp. 481–491, 2007.
[6] Z. Fan, S. Member, Y. Wu, J. Lu, and W. Li, “Based on Structured Prediction with the Convolutional Neural
Network,” arXiv:1802.02208, pp. 1–9, 2018.
[7] K. Kubo, “Pavement Maintenance in Japan,” Road Conference 2017 International Symposium, 2017.
[8] Cross-ministerial Strategic Innovation Promotion Program, “Infrastructure Maintenance, Renovation
and Management - Introduction; The R&D Project of Infrastructure,” 2017. [Online]. Available:
https://www.jst.go.jp/sip/dl/k07/pamphlet_2017_en.pdf
[9] S. C. Radopoulou and I. Brilakis, “Improving road asset condition monitoring,” Transportation Research Procedia,
vol. 14, pp. 3004–3012, 2016.

1967
[10] C. C. Gu, H. Cheng, K. J. Wu, L. J. Zhang, and X. P. Guan, “A High Precision Laser-Based Autofocus Method Using
Biased Image Plane for Microscopy,” Journal of Sensors, vol. 2018, 2018.
[11] H. M. Shehata, Y. S. Mohamed, M. Abdellatif, and T. H. Awad, “Depth estimation of steel cracks using laser and
image processing techniques,” Alexandria Engineering Journal, vol. 57, no. 4, pp. 2713–2718, 2018.
[12] S. Agrawal and P. Natu, “Segmentation of Moving Objects using Numerous Background Subtraction Methods for
Surveillance Applications,” International Journal of Innovative Technology and Exploring Engineering (IJITEE),
vol. 9, no. 3, pp. 2553–2563, 2020.
[13] L. Zhang, F. Yang, Y. Daniel Zhang, and Y. J. Zhu, “Road crack detection using deep convolutional neural network,”
2016 IEEE International Conference on Image Processing (ICIP), pp. 3708–3712, 2016.
[14] M. Quintana, J. Torres, and J. M. Menéndez, “A simplified computer vision system for road surface inspection and
maintenance,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 3, pp. 608–619, 2016.
[15] M. M. Naddaf-Sh, S. Hosseini, J. Zhang, N. A. Brake, and H. Zargarzadeh, “Real-Time Road Crack Mapping Using
an Optimized Convolutional Neural Network,” Complexity, vol. 2019, 2019.
[16] S. Chambon and J. M. Moliard, “Automatic road pavement assessment with image processing: Review and
comparison,” International Journal of Geophysics, vol. 2011, 2011.
[17] B. F. Spencer, V. Hoskere, and Y. Narazaki, “Advances in Computer Vision-Based Civil Infrastructure Inspection
and Monitoring,” Engineering, vol. 5, no. 2, pp. 199–222, 2019.
[18] I. Kahraman, M. Kamil Turan, and I. Rakip Karas, “Road Detection from High Satellite Images Using Neural
Networks,” International Journal of Modeling and Optimization, vol. 5, no. 4, pp. 304–307, 2015.
[19] J. Zhao, H. Wu, and L. Chen, “Road Surface State Recognition Based on SVM Optimization and Image Segmentation
Processing,” Journal of Advanced Transportation, vol. 2017, no. 6, 2017.
[20] Y. Sari, P. B. Prakoso, and A. R. Baskara, “Road Crack Detection using Support Vector Machine (SVM) and OTSU
Algorithm,” 2019 6th International Conference on Electric Vehicular Technology (ICEVT), pp. 349–354, 2019.
[21] S. Jaya and M. Latha, “Diagnosis of Cervical cancer using CLAHE and SGLDM on RGB Pap smear Images through,”
International Journal of Innovative Technology and Exploring Engineering (IJITEE), vol. 9, no. 1, pp. 530–534, 2019.
[22] M. Malathy, C. Rajinikanth, V. Mohan, and T. Yuvaraja, “Hybrid Tile Based Feature Extraction and Support Vector
Machine Base Content-Based Image Retrieval System for Medical Application,” International Journal of Recent
Technology and Engineering (IJRTE), vol. 8, no. 3, pp. 3305–3310, 2019.
[23] M. Hall-beyer, “GLCM Texture : A Tutorial,” University of Calgary, Calgary, Canada, 2018.
[24] E. K. Sharma, E. Priyanka, E. A. Kalsh, and E. K. Saini, “GLCM and its Features,” International Journal of Advanced
Research in Electronics and Communication Engineering (IJARECE), vol. 4, no. 8, pp. 2180–2182, 2015.
[25] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, and B. Shuai, et al., “Recent Advances in Convolutional Neural
Networks,” Pattern Recognition, vol. 77, pp. 354-377, 2018.
[26] A. Jazayeri, H. Cai, J. Y. Zheng, and M. Tuceryan, “Vehicle detection and tracking in car video based on motion
model,” IEEE Transactions on Intelligent Transportation Systems, vol. 12, no. 2, pp. 583–595, 2011.
[27] R. S. Dewi, W. Bijker, and A. Stein, “Comparing fuzzy sets and random sets to model the uncertainty of fuzzy
shorelines,” Remote Sensing, vol. 9, no. 9, pp. 1–20, 2017.
[28] Forty-ninth Meeting of the Council Southeast Asian Fisheries Development Center, “Meeting agreed that the
SEASOFIA could be published on a regular basis,” Southeast Asian State Of Fisheries And Aquaculture
(SEASOFIA), 2017.
[29] D. A. Gustian, N. L. Rohmah, G. F. Shidik, A. Z. Fanani, R. A. Pramunendar, and Pujiono, “Classification of Troso
Fabric Using SVM-RBF Multi-class Method with GLCM and PCA Feature Extraction,” 2019 International Seminar
on Application for Technology of Information and Communication (iSemantic), pp. 7–11, 2019.
[30] E. E. Lavindi, E. J. Kusuma, G. F. Shidik, R. A. Pramunendar, A. Z. Fanani, and Pujiono, “Neural Network based on
GLCM, and CIE L*a*b* Color Space to Classify Tomatoes Maturity,” 2019 International Seminar on Application
for Technology of Information and Communication (iSemantic), pp. 1–6, 2019.
[31] Y. Sari, M. Alkaff, and R. A. Pramunendar, “Classification of Coastal and Inland Batik using GLCM and Canberra
Distance,” AIP Conference Proceedings, vol. 1977, no. 1, 2018.

Application of neural network method for road crack detection

More Related Content

What's hot (18)

Similar to Application of neural network method for road crack detection (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Application of neural network method for road crack detection