Original Article

Determining Roads Asphalt Damage with Image

Processing Algorithms in MATLAB
Hamed Davoudi *, Mohammad Davoudi

M.Sc., Computer Science, Software, Roshd Danesh Higher Education Institute, Semnan, Iran.

Abstract
In recent years, a new application making image processing so significant is the use of image processing techniques (IPTs) in urban planning
and road construction to facilitate work and enhance the rate of detecting asphalt damage on road surfaces using the identification of these
damages and bettering them prior to further damage to these important communication infrastructures. The paper tried to reduce the existing
challenges by properly training of the software and assigning tasks to the machine automatically on the one hand, and on the other tried to
extract better answers from machines compared to humans by reducing human interference in identifying and determining the extent of road
surface damage using extraction of various features by IPTs and proper analysis in MATLAB software. Regarding this, firstly, 824 images
were taken from various places with different conditions (sunny, cloudy, shady, and so on) using a high quality camera. Then the images
received in the previous step were stored in a database and the color images were converted to binary (black and white) ones. Then, after
edge detection, mean filter was done with Sobel default and the images were divided into some sub-blocks, and the features were examined
after applying k-means algorithm. Finally, 199 features of each image were examined and the final pattern was obtained after training the
software. The results indicated that the accuracy of the output in terms of percentage was more than 85%, which based on the previous studıes,
shows acceptable detection by the method presented in this study. The accuracy of the method presented in the classification output was much
higher than the percentage output and in almost more than 95% of cases were placed in the correct class that the experts have identified,
showing the high accuracy of the proposed method in the method in classification.

Keywords: asphalt damage rate, image processing, learning networks, classification.

dataset, one cannot make an accurate and correct evaluation

INTRODUCTION of the proposed methods, as in some cases the methods
One of the indispensable needs of humans after the advent of theoretically seeming efficient do not work well in practice.
the computer and its advancement in all areas is meeting
many of human needs and being well accountable. One can Although recognizing objects in the image is an old issue and
state a stronger reason that is the computer can perform much effort has been made to improve it, it continues to be a
human actions continuously in the long run, after learning, complex issue for many researchers. This is because despite
with a low damage rate and sometimes without errors. The many advances, existing methods still lack enough accuracy
continuation of machine (computer) work process has and speed to be used in many real complex applications and
reached the state where it can decide, see and recognize there is still a long way to develop a system, like the human
instead of humans. The implementation of human perceptual system, that can recognize and simulate shapes
mechanisms in machines has a significant effect on
advancing human goals. Designing object recognition
systems according to machine vision algorithms so that they Address for correspondence: Mr. Hamed Davoudi, M.Sc.,
can be resistant and efficient to different changes in the image Computer Science, Software, Roshd Danesh Higher Education
like rotation, transmission and scale is a very significant issue Institute, Semnan, Iran
with many applications. This has been divided into several
branches with different approaches presented for it, each of This is an open-access article distributed under the terms of the Creative Commons
which has advantages and disadvantages that do not allow Attribution-NonCommercial-ShareAlike 3.0 License, which allows others to remix,
them to be used in all applications, and the appropriate tweak, and build upon the work noncommercially, as long as the author is credited
and the new creations are licensed under the identical terms.
method according to the specific application should be
selected [1].
How to cite this article: Davoudi, H., Davoudi, M. Determining
One of the basic topics in different applications, especially Roads Asphalt Damage with Image Processing Algorithms in
pattern recognition and machine learning applications is MATLAB. Arch Pharma Pract 2020;11(S1):125-31.
having a set of appropriate data for evaluation, analysis, and
comparison of the proposed methods. Without a proper

© 2020 Archives of Pharmacy Practice 1 125

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB
and, in general, patterns accurately in spite of all the obstacles Here, we mention to some examples of what has been done
that may exist [2]. in recent years to determine the extent of road asphalt damage
in different parts of the world. Ramli et al. [6] have proposed
The gradient of the image has two parts - direction and a new method by retrieving information from images in
intensity - both of which can be used. For instance, first one which road asphalt has been damaged. This method involves
can create the desired mask through the intensity, all the areas selecting digital images from the road surface and providing
where the intensity is greater than t, and then use the derived the resolution of each image. Reza Shahabian Moghaddam et
information at that point in these areas. Neural networks must al. [7] have designed hardware similar to a digital camera to
perform backpropagation operations each time the algorithm capture high-quality images with controlled light conditions.
is executed to update different weights and biases. Each The hardware has eliminated the need for all lighting to
repetition of an algorithm known as iterate can update completely remove the ambient light from the tarmac and
weights and biases so that the neural network algorithm can provides artificial lighting so that certain distances from the
identify different features and different classes. Gradients can asphalt pavement surface for all impressions eliminate the
find a minimum point for error, which is of course not the need for software operations to improve and enhance the
minimum value in the total space, but seems reasonable. This images. Rajab and Saif [8] have used imaging techniques to
reasonable point is called a local minimum for error [3]. measure the dimensions of road disruptions. Using ImageJ
software, analyses, designs and using filters, geometric and
Wavelet transform is of the tools that have many applications mathematical operations, measurements like mean and
in engineering, especially machine learning and pattern standard deviations on images are done freely. The results
recognition. Wavelet theory is in fact a generalization to the show that the measurements and the determination of the
theory of transformations and Fourier series, and image damage rate are significantly closer to the manual
compensates for the weaknesses of Fourier analyses in local method measurements. Biglary et al. [9] have presented a
performance and modeling short-term behaviors. There are method to identify asphalt stripping based on image
many ways to remove noise from signals and images by processing methods where the level of asphalt stripping was
wavelet transform. In the methods of noise removal of the first evaluated manually and visually by two methods,
wavelet, noise is shown by wavelet coefficients at small Lattman and boiling water. The results of the model
scales and the coefficients on such scales involve the edge developed to validate this stripping index, based on the results
information as well [4]. of the ITS test, showed that using the proposed stripping
index can accurately determine the stripping of the asphalt
The purpose of applying image processing mixture with 68 to 73% accuracy. Kim et al. [10] have
methods proposed a method to evaluate the performance of asphalt
The purpose of image processing is to extract useful and with image processing methods. Image analysis can
intended data from the input image [5]. The overall steps for automatically provide the configuration to work in a
doing this are generally as follows: At first, it is tried to continuous information access mode and analysis. Moreover,
enhance or reduce the input image; for instance, sometimes the software provides a two-dimensional imaging technique
we need to reduce the image noise. One way to do this is to with the best balance between speed, accuracy and cost.
use low-pass filters. The next step is the location operation, Although the proposed method is at its conceptual stage,
where the exact area of the target object is identified, which when this method is installed in a vehicle with high-resolution
usually creates a silhouette or mask image from the area image recording devices, it provides a feasibility study for
where the object is located. The next step is usually feature accurate and quantitative evaluation of cracking operations.
extraction from the image. This feature can be color, a dot, a Moohan et al. [11] presented a method for detecting and
line, or any other desired feature, which depends on the field determining the extent of road surface damage using digital
you are working in. Then in the last step, the machine learning IPTs, during which two steps are performed. In the first stage,
techniques are used to identify and classify the features of the a video is prepared from the road surface, and in the second
image and to compare them with the features introduced to stage, the analysis is done to identify the damaged points. In
the system in the training phase. this approach, the first 50 frames are used to prioritize the first
three Gaussian modes in the mixed model. Then the method
Past findings uses morphology to enhance noise and fill gaps in detection
With the ever-increasing progress of road construction, one objects. Then it filters the detected backgrounds to highlight
of the significant issues considered by the officials and the identified barriers. After learning and determining the
managers of these areas is timely and accurate identification asphalt damage of road surfaces, one can help the authorities
of damage to roads and in return, timely repair and in the correct timing and care and improvement of the
improvement of these damages. One of the important damaged axes. This will work better than the visual
applications of image processing is the timely identification recognition of asphalt damage by experts, which is costly,
of these damages using various methods, which can lead to inefficient and difficult. This is the first important point in
economic and manpower save and help maintain road prioritizing machine recognition over man, constantly doing
infrastructure. it without tiresome. Now, if we make these observations as
images taken by even ordinary people with medium cameras,
and each time it is recognized after being sent to the machine,

126 Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB
for sure it is cost effective in saving time, long distances and
limited communication with experts in this field. PROPOSED METHOD
Determining the extent of road asphalt damage is a The purpose of presenting this method is to determine the
challenging task because of the complexity of the visual level of asphalt damage on road surfaces with image
patterns and different experiences of experts. . The paper tried processing algorithms based on images received from road
to reduce the existing challenges by properly training of the surfaces and cracks in them. Prior to anything, the research
software and assigning tasks to the machine automatically on data must be collected and introduced to the system, after
the one hand, and on the other tried to extract better answers which the program is designed in two different phases. As the
from machines compared to humans by reducing human database size is high, the images must be read properly and
interference in identifying and determining the extent of road then be processed, so that we finally reach a percentage of
surface damage using extraction of various features by IPTs accuracy. Figure 1 indicates the framework of the proposed
and proper analysis in MATLAB software. In many cases, method
this has happened, and the output of machine analyses on
images has been much more accurate than that of humans
with a lower error rate.

Figure 1: Proposed method framework

1. Data collection In this study, 578 images were used for software training, and
Firstly, 824 images were taken with a high quality camera 246 images were used for testing.
from various locations with different situations (sunny,
cloudy, shady, and so on). After that, by visiting five experts Data collection and storage steps were performed to detect
of the Iran Ministry of Roads & Urban Development, each of road asphalt damage and then the data size and characteristics
the images was reviewed by each of the experts. In the next of each image were reduced by reducing the volume of data
step, using previous experience and knowledge, each expert and applying pre-processing methods.
re-examined each of the images and considered a number for
each image as the value and severity of the damage in The application of various classification or clustering
percentage. In the last step, an average of the opinions of 5 algorithms in the paper was done to discover the pattern,
experts was taken for each image, each of them had assigned whose purpose in discovering the pattern was to create a
a number to the desired image, and finally that number was machine to show the whole training dataset.
recorded in the database as the asphalt damage in the desired
image. 3. Evaluation
Test dataset were used to compare the accuracy of the
2. Pre-processing proposed method. The number of data in test datasets was 246
The data collected in the study is 824 items, with 99 features images. To this end, after training the software by training
examined for pre-processing of each image. In image dataset, the desired output was received in the test section by
processing science, training and test datasets are needed to introducing the number of each image to the software.
apply algorithms related to pattern discovery. The According to the number of images that was 824 in the
development method in this study was to create training and database, each cluster or K-means will eventually be a
test datasets. collection of 199 features. In these 199 features, almost 100
were related to the centers themselves, and 99 to the
Training: To create a pattern of data in the study differences between the centers. We were looking for a
Test: To estimate the accuracy of predicting the pattern relationship between quality and the increased difference
created from training data between centers and higher numbers of centers.

Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201 127

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB

4. Proposed network of learning in neural networks is iteration, which means the

The proposed neural network in the paper was a multilayer data of a data set is injected into the algorithm several times,
perceptron (MLP) network. Neural networks must perform and this algorithm can detect differences in the training data
backpropagation operations each time the algorithm is by increasing or decreasing W and B. Using this technique, a
executed to create different weights and biases. Each iteration minimum point is found for features and errors. Now that
of the implementation of an algorithm known as “iterate” can training has taken place, the network has definite W and B.
update the weights and biases so that the neural network From now on, the network can process an image by entering
algorithm can identify different features and different classes. the previous training, and present the best response with the
As is seen in Figure 2, MLP neural network tries to learn least error.
through changes in weights (W) and bias (B). The principle

Figure 2: Perceptron network

Training samples are transferred to the network and the output of which will be classification, a pattern net network
output from the network is compared with the actual output. was used, which is used to identify the pattern in multi-group
This error is used to change the weight of the neurons so that class files.
the error gradually decreases. This is done using the
backpropagation algorithm, also called backprop. Normally, Accuracy percentage output:
data transmission over a network and weight update, so that Feed forward network is used for the second model based on
the error is reduced, is known as Stochastic Gradient Descent the accuracy percentage prediction. The features are a bit
(SGD). The rate of changes in weight is determined by a different, but for the prediction of the classification groups,
parameter called the learning rate. more wavelet features were used, and for the percentage
output, the gradient features provided a more appropriate
In this experiment, the programming environment used to answer. At the end of the output section, two programs must
implement and test the proposed algorithm is introduced. be executed, which for each image whose number is given by
MATLAB 2018 was used for programming. the user, the original image is found next to the edge detected
image and the category that the program recognizes is
RESULTS displayed simultaneously with the actual category, which in
Two types of output are provided by this proposed method. most cases are equal and this is an indication of the accuracy
In the first model, the outputs are classified and in classes 1 of the proposed method. In the percentage model, the original
to 4 and in the second model, it will make predictions based image is displayed next to the edge image, along with the
on accuracy percentage. correct percentage using the proposed detection method,
along with the number that actually existed (experts
Classification output: prediction).
As is see in Figure 3, in the classification model, after sorting
and preparing the information, the percentages that the Output analysis of the accuracy percentage
program predicts are in the range below 12 in group one, As Figure 1shows, in the percentage output, the task of the
between 12 and 20 in group two, between 48 and 20 in groups detection program with acceptable accuracy is the rate of
three and 48 and above are in group four. The reason for this asphalt damage using the training that has already been seen.
choice is that based on the samples given with this situation, Given the high accuracy of the method, the numbers detected
the test and training samples are almost the same, and about by the software for asphalt damage in each image are usually
200 images are placed in each group. For this model, the close to the damage reported by experts and recorded in the

128 Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB
database. In Chart 1 20, asphalt images are randomly this comparative chart, the detection rate of the method
selected, and their percentage damage, as percentage, is presented in this study is acceptable.
compared with the actual percentage reported by experts. In

80

60

40

‫واقعی‬ ‫درصد‬
Actual percentage 20

‫شده‬ ‫داده‬percentage
The ‫درصد تشخیص‬ 0

1 ‫تصویر‬
2 ‫تصویر‬
3 ‫تصیور‬
4 ‫تصویر‬
5 ‫تصویر‬
6 ‫تصویر‬
7 ‫تصویر‬
8‫تصیور‬
9‫تصویر‬
10 ‫تصویر‬
‫درصد واقعی‬

11 ‫تصویر‬
12 ‫تصویر‬
13 ‫تصویر‬
14‫تصویر‬
15 ‫تصویر‬
Actual percentage

16 ‫تصویر‬
17 ‫تصویر‬
18 ‫تصویر‬
19 ‫تصویر‬
20 ‫تصویر‬

Chart 1: Percentage output

Classification output analysis four classes are seen at the output: Class One, Class Two,
As Chart 2 shows, images are classified into 4 groups in the Class Three, and Class Four.
classified output. In the classification program, after sorting
the information and preparing it, the percentages that the As Chart 2 shows, 20 images were randomly selected and the
program predicts are in the range below 12 are in group one, classification output was obtained from them according to the
between 12 and 20 in group two, between 48 and 20 in group method described above. The chart well illustrates the
three and 48 and above in group four. This means that there accuracy of the proposed method in the study in terms of
classification.

‫واقعی‬ ‫کالس‬
Actual class
0
‫شده‬ ‫تشخیص داده‬
Detected class ‫کالس‬
1 ‫تصویر‬
2 ‫تصویر‬
3 ‫تصویر‬
4 ‫تصویر‬
5 ‫تصویر‬
6 ‫تصویر‬
7‫تصویر‬
8 ‫تصویر‬
9 ‫تصویر‬
10 ‫تصویر‬

‫واقعی‬class
‫کالس‬
11 ‫تصویر‬
12 ‫تصویر‬
13 ‫تصویر‬

Actual
14 ‫تصویر‬
15 ‫تصویر‬
16 ‫تصویر‬
17 ‫تصویر‬
18 ‫تصویر‬
19 ‫تصویر‬
20 ‫تصویر‬

Chart 2: Classification output

After training and test, the program was designed so that each more practical and for everyone to be able to use it in various
user can enter the image in any situation with different situations.
qualities and receive the desired output to make the program

Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201 129

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB
After running the program, the user opens a window and asks It has to be noted that, as explained in the previous section,
the user for an image path. After selecting the image, the the program provides output to the user in two ways. Figure
program automatically creates all past operations on the 5 shows that in this case, the output of the images is presented
image and provides the desired output to the user. to the user as classification and percentage.

Figure 5: Output

DISCUSSION this end, the new proposed method in this paper was
Nowadays, the basic challenge in image processing is presented with MATLAB and the accuracy of the proposed
whether it can act like human eye or not. Human vision is so method was compared with other methods or programs for
complex that we have not yet fully understood our vision to evaluation. The accuracy of the output in terms of percentage
develop image processing. Image processing is on the verge is more than 85%, which shows acceptable detection by the
of making a fundamental difference in our world, and many method presented in this study according to previous studies.
studies are still needed to meet our expectation from image The accuracy of the method presented in this study in the
processing. Moreover, many operations have to be done to classification output is much higher than the percentage
process the visual information. Images usually contain noise output. Moreover, it is in the correct class the experts have
that distorts them. Although this problem is not considered identified in almost more than 95% of cases, which shows the
very serious for the human visual system, for machine vision high accuracy of the proposed method in the study in
systems these noises must be completely filtered and cleaned, classification method.
which is done using some algorithms.
REFERENCES
1. Sajjadi, H., Rashidi Baghi, M., Shir Alizadeh, M. Predicting the
CONCLUSION relationship between stock returns and information asymmetry using
Nowadays, different approaches have been proposed to artificial neural networks, Regional Conference on Accounting in the
identify damages in many areas, especially road construction, Information Technology Age, 2012.
but the methods described have to identify these damages in 2. Jadidi Osguyi, A. Investigation of the structure and function of SMA
asphalt for use in road paving, 4th International Congress of Civil
a timely and appropriate manner. One of the problems with Engineering, Architecture and Urban Development, 2016.
traditional approaches is the need for continuous monitoring 3. Beheshti Shirazi, A. Digital Image Processing Book, 2015.
of road construction experts at the road level and the 4. Ghanem, R., Romeo, F., A wavelet-based approach for the
experimental determination of the level of asphalt damage on identification of linear time-varying dynamical systems, J. Sound Vib.
200p; 234 (4): 555–576.
the road surface, which incurs high costs, and in many cases, 5. Shan, X., Burl, J.B., Continuous wavelet based linear time-varying
the damage and damage are not correctly identified. This is system identification, Signal Process. 2011; 91: 1476–1488.
less common in some developing countries. Sometimes a lot 6. Angreni, I., Adisasmita, S., Ramli, M., Evaluating the Road Damage of
of limitations are associated with going on long distances and Flexible Pavement Using Digital Image, International Journal of
Integrated Engineering, 2018.
sending experts to different axes on the roads. Efforts have 7. Shahabian Moghadam, R., Sahaf, M., Pourreza, H. Automatic
to be made to receive diagnostic data in the field of road diagnosis and classification of pavement damages based on image
construction, especially in Iran, with professional and tissue analysis in the field of location and conversion, Specialized
appropriate equipment, to reach an acceptable detection. To Letter of Transportation Engineering, 2017.

130 Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201

 Davoudi and Davoudi: Determining Roads Asphalt Damage with Image Processing Algorithms in MATLAB
8. Rajab, M., Alavi, M., Saif, M, Application of Image Processing to
Measure Road Distresses, Wseastransactions on Infrmation Science &
Application, 2008.
9. Bigleri, M., Karimi, A., Sharif Tehrani, Sh. Presentation of the
Numerical Index of Asphalt Mixtures Based on Digital Image
Processing, Transportation Infrastructure Engineering, Third Year,
12th in a row, 2017.
10. Kim, K., Kang, M., Linking the Effect of Aggregate Interaction to the
Compaction Theory for Asphalt Mixtures Using Image Processing,
pplied Sciences, 2018.
11. Moohan, M., Kuia, N., Digital Image Correlation Technique for
Measurementof Surface Strains in Reinforced Asphalt Concrete Beams
under Fatigue Loading, Journal of Materials in Civil Engineering,
2019.

Archives of Pharmacy Practice ¦ Volume 11 ¦ Issue S1 ¦ January-March 20201 131