AUTOMATED PAVEMENT IMAGING PROGRAM (APIP) FOR

PAVEMENT CRACKS CLASSIFICATION AND QUANTIFICATION

Othman Che Puan1, Mushairry Mustaffar2, Tung-Chai Ling1,*

1
Department of Geotechnics and Transportation,Universiti Teknologi Malaysia, 81310
Skudai, Johor, Malysia
2
Department of Environmental Engineering,Universiti Teknologi Malaysia, 81310 Skudai,
Johor, Malysia

*Corresponding Author: tcling611@yahoo.com

Abstract: This paper describes the development of an Automated Pavement Imaging Program (APIP) for evaluating
pavement distress condition. The digital image processing program enables longitudinal, transverse, and alligator
cracks to be classified. Subsequently, the program automatically predicts types of cracks and estimates the crack
intensity which can be used to rate pavement distress severity. Results obtained by this technique are compared with
the conventional manual method to check accuracy. The algorithm developed in this study is capable of identifying
types of cracks and the severity level at the accuracy about 90%, i.e. similar to the accuracy obtained by the manual
method.

Keywords: Pavement Distress; Crack; Classification; Quantification; Image Processing.

Abstrak: Kertas ini membincangkan pembangunan Aturcara Imej Pavemen Otomatis (APIP) untuk menilai keadaan
kerosakan pavemen. Aturcara pemprosesan imej digital membolehkan retak jenis memanjang, lintang dan buaya
diklasifikasikan. Aturcara seterusnya secara otomatis mengkelaskan jenis-jenis retak dan menganggar keamatan retak
yang digunakan sebagai asas penilaian tahap kerosakan pavemen. Keputusan yang diperoleh dari teknik ini
dibandingkan dengan kaedah manual konvensional bagi menilai kejituannya. Algoritma aturcara yang dibangunkan
dalam kajian ini didapati mampu mengenal pasti jenis-jenis retak dan tahap kerosakannya dengan kejituan 90%, iaitu
menyamai kejituan yang diperoleh melalui kaedah manual.

Katakunci: Kerosakan pavemen; Retak; Pengkelasan; Penilaian; Pemprosesan Imej.

1.0 Background of Study

The evaluation of pavement condition is an important part in pavement management. Accurate

evaluations would result in a better chance that resources will be distributed normally, thus,
yielding a better service condition (Kim, 1998). Pavement can be evaluated through the different
types of distress experienced, such as cracks, disintegration and surface deformation. At present,
there are various methods of conducting distress surveys, recording and analysing distress survey
data. For example, pavement engineers have long recognized the importance of distress
information in quantifying the quality of pavements. This information has been used to document
present pavement condition, chart past performance, and predict future pavement performance.

2.0 Problem Statement

Manual visual inspection of pavement surface is costly and time consuming. Visual observation of
pavement distress is the most common method for monitoring pavement surface condition. This
has been traditionally performed by trained engineers who work or drive along the road and counts
the distresses (Oh, 1998). However, this method of field inspection poses several drawbacks, such
as: (i) Slow, labour intensive and expensive, (ii) Subjective approach, thus generates
inconsistencies and inaccuracies in determining the of pavement condition, (iii) Inflexible and does
not provide an absolute measure of the surface, (iv) Has poor repeatability since the assessment of
given pavement section may be differ from one survey to the next, and (v) Could expose a serious
safety hazard to the surveyors due to high speed and high volume traffic.
There is a need to minimise these drawbacks by replacing manual data collection system with
automated systems that utilise imaging techniques. However, the size, shape, and variations of
each distress type, as well as the variations of the texture and colour of the pavement surface,
present a challenge to researchers (Li et al., 1991). Therefore the aim of this study is to develop an
automated imaging program using existing software. Subsequently, the performance of the
developed program is being assessed and compared with the conventional manual method to check
level of accuracy.

3.0 Cracks

Cracks are fissures resulting from partial or complete fractures of pavement surface. Cracks of
road pavement surface can happen in a wide variety of patterns, ranging from isolated single crack
to an interconnected pattern extending over the entire pavement surface (Jabatan Kerja Raya (JKR),
1992).
Based on the Distress Identification Manual for the Long-Term Pavement Performance project
(SHRP-P-338, 1993) published by the Strategic Highway Research Program (SHRP), cracks for
asphalt concrete pavement can be classified into six types, namely fatigue crack, block crack, edge
crack, longitudinal crack, reflection crack at joint and transverse crack. Each type is classified into
three severity levels of low, moderate, and high. The extent of crack can be measured in terms of
the number of crack, crack lengths and crack areas. The general form of the various types of crack
is illustrated in Figure 1. A description of each type of cracks can be found in JKR (1992).

Figure 1: Types of cracks (JKR, 1992)

4.0 Pavement Evaluation Method

There are generally two pavement evaluation methods, manual and automated or a combination of
both. The manual method may consist of mapping the road section, walking the road while
performing the evaluation survey or driving and recording distress observation and combination of
these. The automated method may include the use vehicles to photograph the road surface, the use
of ultra sonic sensor, video and laser. There are also different types of automated pavement
evaluation system, a fast evaluation with sampling data being collected to obtain an overview of
the road network and detailed condition surveys required for research (Benson et al., 1988).
The types of distress data collection for pavement evaluation are typically roughness (ride),
structural (deflection), surface distress, and skid resistance. This study concentrates on surface
distress which can be defined as a measure of pavement fracture (cracks) and disintegration
(delamination and pothole). Measures include distress type and level of severity.

5.0 Development of Image Processing in MATLAB Environment

The applications of image processing techniques to the collection and analysis of road pavement
distresses have been discussed by numerous researchers such as Baker et al (1987), Hintz et al
(1989), Ritchie (1990), Li et al (1991), Grivas et al (1994), Jitprasithsiri (1997) and, Cheng and
Miyojim (1998).
The first step involved in the automated image processing is the acquisition and digitisation of
the image. In this study, the height of the digital camera (5.4mm) lens to the pavement was about
one meter. The digitised array size was 640 by 480 pixels, which resulted in 480 lines vertically
and 640 elements horizontally. The original image was a mathematical representation of a colour
image in a 24-bit per pixel size format. This colour image consisted of a combination of three 8-bit
arrays, and 8-bit arrays contained brightness value for red, green and blue, respectively. To
facilitate image processing and manipulation of the image, a brightness level of each pixel,
assigned at a value between 0 (black) and white (255) is needed to convert the colour image to gray
scale image. Thus, the 24-bit per pixel format was converted to an 8-bit per pixel format, and this
reduced the file space required for storage by two-thirds.
This study utilises full programming language software MATLAB 6.5.1 to enable a series of
MATLAB statements to be written into a file and then execute them with a single command.
Algorithms for this study, named as Automated Pavement Imaging Program (APIP), were
developed in ordinary text file. APIP is capable of identifying, classifying, and quantifying the
pavement distress types in terms of severity and extent. In the following sub section, each of the
image processing steps developed under the MATLAB environment is discussed in detail.

5.1 Image Enhancement

Image enhancement was applied in an attempt to remove noise in pavement images. Variable
enhancement techniques are available to improve an image, including histogram equalization,
neighbourhood averaging, media filtering, and linear filtering in MATLAB environment. Among
the filtering techniques, the median filtering performed the best in removing spot noise as
determined by observation (Jitprasithsiri, 1997).
Median filtering was therefore applied as pavement image enhancement technique in this
research. It is similar to using an averaging filter, in that each output pixel is set to an average of the
pixel values in the neighbourhood of the corresponding input pixel. The size of the neighbourhood
used for filtering is 3-by-3. However, the value of an output pixel is determined by the median of
the neighbourhood pixels, rather than the mean. The median is much less sensitive than the mean to
extreme values. Median filtering is therefore better able to remove these outliers without reducing
the sharpness of the image.

5.2 Image Thresholding

The spatial and light intensity information on the image is usually combined with a thresholding
technique for the improved image segmentation. The segmentation relates the threshold for a
given image to the mean and standard deviation value of the corresponding gray scale histogram
through an equation. The procedure for image segmentation is shown in Figure 2.
Detail calculations are as the following:

(i) Calculation of the mean value of the enhancement image.

ly
1 lx
G = ∑ ∑ G ij , n = lx × ly (1)
n i =1 j =1
(ii) Calculation of the standard deviation value of the enhancement image.

( )
l
1 lx y

∑ ∑
2
S= Gij − G (2)
n − 1 i =1 j =1
(iii) Determination of the optimal threshold value.
Th = G − A × S , A = Constant Value (3)
(iv) Comparison of each pixel value with the optimal threshold value, and assign new value
based on the following expression:
If Gij < Th then Outij = 0 (Black)
If Gij > Th then Outij = 255 (White)
Where
Gij = Input grey scale intensity at position i, j
Outij = Output grey scale intensity at position i, j

Figure 2: Procedure for image segmentation

Based on the experiment with the image data set, it was found that the mean and standard
deviation value were good estimators for predicting the threshold level. It was observed that an
A-value of between 1.6 to 2.2 isolated most of the distress features from the background. When A
is less than 1.6, too many noisy spots would appear. On the other hand, if A is greater than 2.2,
some short and thin crack would be discarded.
Once the optimal threshold value was determined, the pixels with grey level below the
threshold were referred to as distress pixels and pixels whose grey level value exceeded the
threshold were referred to as background.
 It is found that the simple image enhancement algorithm worked well in predicting the
presence or absence of distress features on the image as shown in Figures 3a, 3b, 3c show the
binary image obtained from Otsu’s method and proposed segmentation algorithm respectively.

Original Image
0
Otsu's Method

50

100

150

200

250

300

350

400

450
0 100 200 300 400 500 600

Figure 3a: Original image with a mark of distress Figure 3b: Binary image (Otsu’s method)

TCLING Thresholding Image

Figure 3c: Binary image (proposed algorithm) with distress but no shadow and paint striping

Figures 4a, b, c show images contained distress and paint striping. The algorithm removed
extraneous features that have higher pixel intensities than the mean pixel intensity in the image. In
this process, all the pixels representing paint striping and surface textures brighter than the optimal
threshold level are suppressed to the background.
Original Image
0 Otsu's Method

50

100

150

200

250

300

350

400

450

0 100 200 300 400 500 600

Figure 4a: Original image with distress and paint Figure 4b: Binary image (Otsu’s method) with distress
striping and paint striping
 TCLING Thresholding Image

Figure 4c: Binary image (proposed algorithm) with distress and paint striping

The system also worked well for most images with shadow. The algorithm correctly isolated
the distress features from the background for the image with both shadow and distress (see Figure
5a, b, c). The algorithm estimated the optimal threshold level so that the resulting binary image did
not contain any noise that was due to shadow.

Original Image
0 Otsu's Method

100

200

300

400

500

600

700

800

900

1000

0 200 400 600 800 1000 1200 1400

Figure 5a: Original image with distress and shadow Figure 5b: Binary image (Otsu’s method) with distress
and shadow

TCLING Thresholding Image

Figure 5c: Binary image (proposed algorithm) with distress and shadow
 To perform a sensitivity analysis of image segmentation, two versions of the APIP were
written. One program used 1.8 as a constant A-value and the other program used 2.0. Figures 6a, b
show segmented images of Transverse Cracks resulting from two different constant value of A.
Twenty-eight pavement images were selected and processed through the automated programs with
those two values of A. The accuracy of both results was compared with the manual approach.

A = 1.8 A = 2.0

Figure 6a: Segmented image with 1.8 constant value Figure 6b: Segmented image with 2.0 constant value

5.3 Morphological Closing

Figure 7, segmented image of Figure 3a, shows how some noise cluster occupies open space
corresponding to the background. Noise removal at the threshold of 10 pixels was applied,
followed by performing morphological closing with structural element of size 5, than producing
another binary image. The resulting aspect is shown in the sample of Figure 8, where it can be
noticed that some of the noises have been removed and some holes being filled with object pixels.

TCLING Thresholding Image Noise Removal & Closing

Figure 7: Image segmented of Figure 3a before closing Figure 8: Image segmented of Figure 3a after closing
and noise removal and noise removal

5.4 Thinning

The binary image with clusters was not useful enough and had to be further processed by thinning.
The thinning made clusters in the binary image to become one-pixel wider and took the form of
polylines. After thinning, a generic source image had the appearance as shown in Figure 9. These
are the skeletons that are beneficial to distress quantification.
 Thinned Image

Figure 9: Image of Figure 8 after thinning

5.5 Cracks Classification

After segmentation, the types of cracks could be automatically detected by the developed computer
program through comparing several parameters in x and y-array. The parameters ASGLX and
ASGLY, which were the average values of summed gray level of distress zones in the x- and
y-arrays respectively, are clearly illustrated in Figures 10a, b.

Suspected Distress Location ASGLX/ASGLY > 2

ASGLX ASGLY

x-array y-array

Figure 10a: Distress recognition concept of Longitudinal crack at suspected distress zone

Suspected Distress Location 0.6 <AS GLX/ASGLY < 2

ASGLX ASGLY

y-array
x-array

Figure 10b: Distress recognition concept of Alligator crack at suspected distress zone

From the characteristics of summed gray level graphs for a number of pavement images with
the different types of distresses, the following was found to be true. Different distresses will give
distinctly different summed array profiles. For alligator crack, the distress zones in two arrays are
relatively wide. For a longitudinal and transverse crack, the narrow and sharp peak of summed
gray level is always found in the x-array and y-array, respectively, whereas the summed gray level
in the other array should not show noticeable peaks (see Figures 11 and 12).
 Standard deviations of the two arrays were determined as the intermediate background gray
level (IBGL). If the IBGL values of any arrays are greater than 5, value of 5 will be assigned as the
IBGL value of the array. The following shows the IBGL value obtained from transverse crack as
shown in Figure 8. Column standard deviation = 3.93, therefore IBGLColumn = 3.93 Row standard
deviation = 21.62, therefore IBGLRow = 5.00
Every summed gray level (columns and rows) that had values larger than IBGL were referred
to as the distress zones as shown in Figures 11b and 12b. After eliminating the distress zones, the
ratio of average values of the distress zones in x- and y-directions (ASGLX/ASGLY) could be
calculated for crack type identification. The summary of detecting rules based on this
ASGLX/ASGLY ratio is shown in Figure 13.

Summed Graylevel Graph (Column) Distress Zones

20 20

18
(a) 18
(b)
16 16

14 14

12 12

10 10

8 8

6 6

3.93 4 4

2 2

0 0
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700

Figure 11: Summed gray level graph (column) of Figure 8 (a) before, (b) after compared IBGLColumn value

Summed Graylevel Graph (Row) Distress Zones

150 150

(a) (b)

100 100

50 50

5.00 0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500

Figure 12: Summed gray level graph (row) of Figure 8 (a) before and (b) after compared IBGLRow value
 Figure 13: Distress type identification rule

5.6 Cracks Quantification

Another deterministic approach in APIP is to classify distress severity level based on the
manual published by JKR (1992). The manual explains how to measure and give the unit of
measurement for each type of crack. Alligator crack is measured in square meters.
Longitudinal and transverse cracks are measured in linear meters. The severity levels of each
type of crack are classified as low, moderate, or high. For example, a moderate severity of
longitudinal crack is described as “cracks with moderately severe spalling; meaning unsealed
crack width greater than 3 mm; sealant material in bad condition” (JKR, 1992).
The standard crack density concept can be readily and logically implemented in pavement
image processing (Lee and Oshima, 1994). The standard crack density is determined by
dividing the extent with the average crack width, i.e.
Crack Density (cm/cm2) = extent / (average crack width × 10) (4)
Where the extent is the total cracked within the total pavement area (expressed in percentage)
and the average crack width is the mean crack opening width of set of cracks (expressed in
mm).
The standard crack density concept is at an advantage for use in image processing analysis
since it takes into consideration both extent and width of cracks simultaneously. A simple
concept of determining length, area, average width and extent of crack is shown in Figure 14.
Area (A) of distress can be calculated from the total number of pixels from distress area and
multiplied by the pixel size. Length (l) will be determined based on the polylines of one pixel
wide. Then, the average crack width (CW) can be derived from area by dividing the length.
 Figure 14: Crack area with central polyline

Where,
A = number of pixels of distress area × pixel size (5)

l = total pixel number of polylines × pixel size (6)

Crack Width = A / l (7)

To illustrate the computation of crack quantification in APIP, Figure 15 shows rule-based

classification of severity level. Severities of each type of crack are mainly referred to crack
density and average width.

Figure 15: Distress severity identification rule

A sample of TCLING-APIP (Notepad) output file containing the severity and extent
parameters for cracked image is presented in Figure 16. The output file also provides crack type
classification.
 Figure 16: Output report corresponding to the transverse crack of Figure 3c

6.0 Manual Cracks Assessment

The measurement of crack width is difficult, but it is important because it partly determines the
crack severity. In this research, the width of the cracks was measured with vernier calipers as
shown in Figure 17. This measurement technique is introduced by Long-Term Pavement
Performance (LTPP) program for measuring the opening crack. The widths of the cracks
usually vary within any block, and so it was the width of crack which was predominant that was
recorded.

Figure 17: Vernier Caliper

A rectangle was drawn with the back lines to cover the analysis of 414 tiles in the original
gray scale pavement image. The size of the rectangular area on the paper was 3.7 mm × 3.7 mm
of actual size (25 by 25 pixels). The image samples were printed on A4 (8.5 × 11 inch) paper.
The images were manually rated and the extent of each type of crack was computed. A data
collection form was prepared for each rectangular area rating. These pavement images were
filled out with a rating from 0 to 100 % based on an opinion indication on the amount of the
visible crack.

7.0 Results and Discussion

The data presented in this study requires analysis by statistical approach to evaluate the
performance of the pavement evaluation system. The accuracy of the system performance was
determined by using t-test. This is required in order to determine how close an observation
matches the accepted reference value or the assumed true value. For comparison of two
databases, visual field surveys were used as assumption of the true values or reference values.
Therefore, the observed data from the APIP could be compared with the data from the visual
field inspection.
The validation of APIP algorithms is discussed in this section. First, proper constant
values of A (1.8 and 2.0) were selected for the image segmentation (thresholding) algorithm.
After that, the crack width (CWs) and crack density (CDs) obtained from the automated rating
were compared with those from the manual rating. Twenty-eight image samples collected from
various road sections in Skudai, Johor were used for the analysis.

7.1 Crack Width and Cracks Density

The image samples were processed automatically through APIP to compute the CWs and CDs
for severity level identification. Distress type predictions resulted from APIP were compared
with the visually determined distresses descriptions. The CWs and CDs resulted from the 28
samples are shown in Table 1.

Table 1: CWs and CDs by the 1.8_APIP and 2.0_APIP vs. the manual inspections
Sample Average Width (mm) Cracks Density (cm/cm2)
Cracks Type
No. A=1.8 A=2.0 Manual A=1.8 A=2.0 Manual
1 Alligator 6.71 6.37 6.20 0.0955 0.0852 0.1113
2 Alligator 6.16 5.94 6.20 0.1174 0.1022 0.1103
3 Alligator 4.91 4.61 5.95 0.1200 0.1015 0.1185
4 Alligator 7.00 6.54 6.50 0.1053 0.0940 0.0895
5 Alligator 6.38 5.95 6.65 0.1034 0.0904 0.0962
6 Alligator 5.72 5.34 6.10 0.1112 0.1021 0.1069
7 Alligator 5.96 5.58 6.30 0.1151 0.1056 0.1219
8 Alligator 6.25 5.96 6.65 0.0800 0.0651 0.0630
9 Alligator 6.67 6.24 7.10 0.0967 0.0859 0.0613
10 Alligator 5.58 5.43 5.95 0.0747 0.0589 0.0524
11 Alligator 7.76 7.55 7.65 0.0604 0.0562 0.0607
12 Alligator 8.77 8.50 7.95 0.0562 0.0504 0.0307
13 Longitudinal 9.91 9.05 9.00 0.0300 0.0261 0.0137
14 Longitudinal 8.35 7.29 6.70 0.0311 0.0230 0.0130
15 Longitudinal 4.53 3.70 5.45 0.0132 0.0084 0.0084
16 Longitudinal 5.50 5.11 6.55 0.0224 0.0153 0.0092
17 Longitudinal 6.80 6.40 7.25 0.0165 0.0113 0.0097
18 Longitudinal 7.42 7.24 6.65 0.0221 0.0162 0.0098
19 Non-Crack 0 0 0 0 0 0
20 Non-Crack 0 0 0 0 0 0
21 Non-Crack 0 0 0 0 0 0
22 Transverse 6.91 5.99 6.10 0.0155 0.0144 0.0131
23 Transverse 5.63 5.27 6.45 0.0169 0.0150 0.0119
24 Transverse 6.84 7.13 7.45 0.0197 0.0143 0.0117
25 Transverse 5.34 4.98 7.30 0.0283 0.0215 0.0136
26 Transverse 7.59 6.91 7.45 0.0167 0.0153 0.0123
27 Transverse 8.10 7.97 6.65 0.0449 0.0378 0.0164
28 Transverse 9.46 9.53 7.95 0.0329 0.0269 0.0155
 To evaluate the APIP performance using the on-field measurement approach, the results
had to be analysed by using statistical approaches. The results of the paired t-test are
summarized in Table 2 and Table 3. The statistic t-values from t-test are listed in two categories
(CWs and CDs) for comparison.

Table 2: Statistic test for Manual and APIP (A=1.8) inspection

Constant Value A=1.8, Th = G − 1.8 × S
(a) Analysis of Average Crack Width
Technique Mean (mm) S.D t tα
Manual 6.077 2.271
0.363 1.297
APIP 6.080 2.503
(b) Analysis of Average Cracking Density
Technique Mean (cm/cm2) S.D T tα
Manual 0.0422 0.0428
0.841 1.297
APIP 0.0516 0.0415

Table 3: Statistic test for Manual and APIP (A=2.0) inspections

Constant Value A=1.8, Th = G − 1.8 × S
(a) Analysis of Average Crack Width
Technique Mean (mm) S.D t tα
Manual 6.077 2.271
0.363 1.297
APIP 6.080 2.503
(b) Analysis of Average Cracking Density
Technique Mean (cm/cm2) S.D T tα
Manual 0.0422 0.0428
0.841 1.297
APIP 0.0516 0.0415

As shown in Table 2, all statistic t-values were lower than the tα at the significance level α
= 0.20. Thus, t-distribution indicated that there were no significant differences between the
overall crack width and crack density for 1.8_APIP and the field inspections. In Table 3, test for
2.0_APIP also showed higher tα-values than the statistic t-values at the 0.20 significance level.
This indicates that the mean difference between each data was not significantly different.
For overall results with comparison to the manual inspection rating, the mean of the CWs
with 1.8_APIP performance is closer to the mean of manual crack width than the mean of the
CWs with 2.0_APIP. In contrary, the 1.8_APIP highly overestimated CDs while the 2.0_APIP
slightly overestimated CDs. This result showed that 2.0_APIP performed more accurately in
quantifying CDs than 1.8_APIP. Logically, there is no a need to test the 2.2_APIP since a high
degree of underestimated crack density would be expected.

7.2 Cracks Type Prediction

The APIP was then tested to determine whether it could distinguish between several types of
crack and images without crack. Using additional three actual pavement images without crack,
the automated imaging algorithms correctly identified most of them. The results from the
1.8_APIP and 2.0_APIP for the 28 images are shown in Table 4.
Table 4: Summary of 1.8_APIP and 2.0_APIP algorithms performance in cracks type prediction
(APIP) Predicted As
Actual Total Alligator Longitudinal Transverse Non-Crack Accuracy (%)
Distress Type Sample
Number 1.8 2.0 1.8 2.0 1.8 2.0 1.8 2.0 1.8 2.0
Alligator 12 12 12 100 100
Longitudinal 6 1 5 6 83.33 100
Transverse 7 1 1 6 6 85.71 85.71
No Distress 3 3 3 100 100
Total 28 14 13 5 6 6 6 3 3 92.86 96.43

From Table 4, APIP worked very well in predicting the presence or absence of distress
features on the image. Of the 25 images with distress features, the 1.8_APIP correctly predicted
23 images that had distress features while 2.0_APIP correctly predicted 24 images. Only 2
images out of 25 images with distress features were incorrectly categorized by using 1.8_APIP.
However, 2.0_APIP had only miscategorised one image.
Of the three distress-free images, both 1.8_APIP and 2.0_APIP correctly categorised them
all as non-crack distress. Therefore, the overall predictions were found to be 92.86 % accurate
in 1.8_APIP and 96.43% accurate in 2.0_APIP.

7.3 Severity Level Classification

Table 5 shows the automatically severity level classification through the developed APIP which
was based on manual inspection.

Table 5: Summary of 1.8_APIP and 2.0_APIP algorithms performance in severity level prediction

Observed Observed Total Sample Number

Severity Severity High Moderate Low Accuracy (%)
Level Level 1.8 2.0 1.8 2.0 1.8 2.0 1.8 2.0
High 16 13 13 3 3 81.25 81.25
Moderate 9 9 9 100 100
Low 0 100 100
Total 25 13 13 12 12 0 0 88.00 88.00

Severity comparisons indicated that the APIP algorithms were mainly in the average of
88%. However, in a few cases, severity appeared to be lower by using the automated system.
The lower classifications of the severity indicated similar concerns that were present during any
of the survey at fields, which resulted from different interpretations of the severity level among
the evaluators.

8.0 Conclusions

In this study, an Automated Pavement Imaging Program (APIP) was developed to process
images with both noise and contrast variation problems. The image processing method
developed provides visual distress information and produces inspector quality interpretation.
The output from this APIP will be a report, detailing the type, extent and severity of the various
types of cracks. To date, the overall system has been found to be about 90 percent accurate. It
is still necessary to integrate the various components of both image-processing and analysis
systems before a goal of fully automated pavement distress analysis tool can be achieved.
 The APIP for pavement crack analysis involved six major steps:
(i) Image Enhancement Algorithm
A 3 × 3 median filtering was applied to reduce noise in the pavement images captured.
The 3 × 3 median filtering was found to be effective at reducing the noise, resulting
from the pavement aggregates in reasonable time.
(ii) Image Thresholding Algorithm
Image thresholding algorithm showed superior results in the image segmentation
process and can be used in non-uniform background illumination of asphalt pavement
images. The proposed algorithms used information from the source image itself had the
advantage of removing the effect of the background illumination, so that repeated image
acquisition of the same section of pavement at different times of day or illumination
conditions could render similar thresholded images, resulting in consistent
classifications for all the repetitions.
(iii) Thinning Algorithm
Crack length was determined using the thinning algorithm based on the central lines
with one pixel wide. Then, the average width of crack was obtained by dividing the area
and the length. These parameters were found to be very beneficial for cracks
quantification.
(iv) Morphological Closing
As a result of morphological closing, the edges of the objects were smoothed out by
eliminating the external jagged contours that might occur during the thresholding
operation. The closing operation also fills small holes and gaps produced by
thresholding that may cause inaccuracy in quantification.
(v) Distress Classification
A simple rule-based classification approach was developed, taking into account four
distress types: longitudinal crack, transverse crack, alligator crack and non-crack.
Subroutines were implemented with decision rules that compare the several parameter
in x and y-array extracted from the binary image. The accuracy of these subroutines is
close to 93 percent.
(vi) Distress Quantification
The standard crack density (CD) concept is based on the two primary dimensions of the
amount of cracks: extent and crack width. The advantage for its usage is in image
processing analysis, since it takes into consideration both extent and width of cracks
simultaneously. Thus, the CD calculated using APIP is an important data for local
pavement management system in maintenance and rehabilitation selection.
To develop a rule-based classification of severity level computed in APIP, crack density
and average width are chosen in this study to be the main parameters for severity level
classification. The APIP results showed that it was statistically not significantly different from
conventional method (JKR standard). Thus, the proposed program, which includes different
image processing algorithms, produces a reliable computer-based classification.
From the analysis of the results, assessment of the accuracy and reliability of the proposed
pavement evaluation system were obtained. The proposed evaluation method and field
inspection in this study showed some relationship. In general, most observed sections were
well matched by proposed method and field survey. However, some sections were not.
Even though there were some limitations, the system is expected to provide a user-friendly
approach to local transportation agencies as well as the government sector. An integration of
several image processing algorithm in one APIP has the capability of a) processing the image to
isolate the distress features, b) displaying the binary image, c) reporting the type and severity of
that image in an output report.
References

Baker, J., Dahlstrom, B., Longenecker, K., and Buu, T. (1987) Video Image Distress Analysis
Technique for Idaho Transportation Dep. Pavement Management System. Transp. Res.
Record 1117.
Benson, K., Elkins, G., Uddin, W. and Hudson, W. (1988) Comparison of Methods and
Equipment to Conduct Pavement Distress Survey. Transp. Res. Record. 1196: 40-50.
Cheng, H. D., and Miyojim, M. (1998). Automatic pavement distress detection system. Journal
of Information Sciences. 108: 219-240.
Grivas, D. A., Bhagvati, C., Skolnick, M. M., and Schultz, B. C. (1994) Feasibility of
Automating Pavement Distress Assessment Using Mathematical Morphology. Transp. Res.
Record. 1435: 52-58.
Hintz, R. J., Karakadas, C., and Kang, J. M. (1989) Analysis of Pavement Cracking and Rutting
Using Close-range Photogrammetry. Photogrammetric Engineering and Remote Sensing.
55(2): 217-221.
Jitprasithsiri, S. (1997) Development of a New Digital Pavement Image Processing Algorithm
for Unified Crack Index Computation. Ph.D. Dissertation, The University of Utah.
JKR (1992) A Guide to Visual Assessment of Flexible Pavement Surface Conditions. Malaysia:
Jabatan Kerja Raya Malaysia.
Kim, J. (1998). Development of a Low-Cost Video Imaging System for Pavement Evaluation.
Ph.D. Thesis, Oregon State University:
Kittler, J and Illingworth, J. (1986) Minimum Error Thresholding. Pattern Recognition. 19(1):
41-47.
Lee, H. (1990) Evaluation of Pavedex Computerized Pavement Image Processing System in
Washington. Proceeding of the 1990 Automated Pavement Distress Seminar, Federal
Highway Administration, 205-224.
Lee, H., and Oshima, H. (1994) New Cracking Imaging Procedure Using Spatial Autocolletion
Function. Journal J. Transp. Engrg, ASCE. 120(2): 206-228.
Lewis, R. (1990) Practical Digital Image Processing. England: Ells Horwood Limited.
Li, L., C han, P., and Lytton, R. L. (1991) Detection of Thin Cracks on Noisy Pavement Images.
Transp. Res. Record. 1311: 131-135.
Oh, H. (1998). Image Processing Technique in Automated Pavement Evaluation System. Ph.D.
Dissertation, University of Connecticut.
Otsu, N. (1979) A Threshold Selection Method from Gray-Level Histogram. IEEE
Transactions on System, Man, and Cybernetic. SMC-9(1): 62-66.
Ritchie, S. G. (1990) Digital Imaging Concepts and Applications in Pavement Management.
Journal Survey Engineering, ASCE. 116(3): 287-298.
SHRP-P-338 (1993) Distress Identification Manual for the Long-Term Pavement Performance
project. Washington, D.C.: Strategic Highway Research Program.