Pavement Distress Detection and Classification using Feature Mapping
E. Salari and G. Bao
Department of Electrical Engineering and Computer Science
The University of Toledo
Toledo, OH 43606
Email: esalari@utnet.utoledo.edu
Phone: (419) 530-6002
Abstract----The detection of crack s and other degradations on pavement surfaces was traditionally done by human experts using visual inspection while driving along the s urveyed road. To overcome the limitations of the manual scheme, an automatic crack detection and classification system is proposed in this paper to both speed up and reduce the subjectivity of the process. After the pavement images are captured by a digital camera, regions corresponding to cracks are detected over the acquired image by local s egmentation and then repres ented by a matrix of s quare tile s. Since the crack pattern can be re
pres ented by the dis tribution of the crack tiles, standard deviations for both vertical and horizontal histograms are calculated to map the cracks onto a 2D feature s pace, where four crack types, namely, longitudinal, transversal, block, and alligator cracks can be identified. The experimental results, obtained by tes ting real pavement images over local as phalt roads, present the effectiveness of our algorithm for automating the process of identifying road distresses from images.
I.INTRODUCTION
Pavement distress refers to the visible imperfection on the surface of the pavements due to overloading, environmental conditions, and normal wear. Often the distresses are present in the form of surface cracking of various types. The analysis of the pavement crack data is one of the most important criteria in determining optimum pavement maintenance strategies. The simplest method is to visually inspect the pavements and evaluate them by subjective human experts. This approach, however, involves high labor costs and produces unreliable and inconsistence results. To overcome the subjective visual evaluation process, several attempts have been made to develop an automatic procedure.
Wang et al. [1] describes an automated system capable of real-time assessment. Using analytical desc
riptions of pavement stress, they compare the images under consideration with a pre-defined database of typical crack characteristics such as location and geometry, ultimately producing surface crack indices.  Lee et al. [2] concentrate on image preprocessing and representation for input to a neural network. After tiling the image, they use local statistics to tag tiles which contain cracks, thus forming a binary crack matrix. This matrix is then summed along the X and Y axes, forming two histogram vectors. These vectors are then presented to a multi layer perceptron (MLP) for classification.
Hsu et al. [3] described a moment invariant technique for feature extraction and a neural network for crack classification. The moment invariant technique reduces a two dimensional image pattern into feature vectors which can be used in the classification stage. The overall results of this study is claimed to be satisfactory and the classification accuracy of the proposed system is eighty-five percent.
Although many attempts [4-6] have been made to automatically collect pavement crack data, due to the non-uniform illumination effect and irregularities of the pavement surface, limited success was achieved in accurately detecting cracks and classifying the crack types. In addition, most existing systems require complex algorithms with high level computing power, thus better approaches are still needed to optimize the process. In this paper, we present a reliable automated pavement distress eval
uation system capable of detecting the cracks as well as classifying the crack types from digital pavement images. The proposed model consists of two major parts: crack detection and crack classification.
For the crack detection, a non-uniform illumination effect removal filter [7] followed by a local segmentation method with multiple threshold values provides a binary image. The binary image is then divided into sub-images called “tiles”, each of which stands for either crack or background. Based on the distribution of these crack tiles, the crack type is then identified after mapping to a 2D feature space. This tile-based system could significantly reduce the computational complexity relative to pixel-based approaches. In addition, it is less sensitive to the background noise since only a few noise pixels alone will not be sufficient for a tile to be classified as a crack tile. As a result, the proposed algorithm is able to provide more reliable classification results.
978-1-4244-6875-1/10/$26.00 ©2010 IEEE
II.
IMAGE ANALYSIS TECHN CRACK DETECTI A) Removal of Non-Uniform Illumination In order to provide a more uniform b an observed pavement image, a backgro
method based on morphological operatio in this section, ensuring the same back condition for all the pavement images.
First a morphological opening performing a local minimum filter (erosi a local maximum filter (dilatation) structuring element. This allows the elim
objects (cracks in the foreground) while large ones (the background). In this way t illumination of background can be succes Then, by subtracting the background fr gray scale image, the non-uniform illu can be fully removed. Following this s average intensity value of the backgroun
pixel of the image to obtain a pavement i uniformly distributed background. Figu
details of the above procedure.
Figure 1 (a) Original pavement image with non-u of background.
(b) Non-uniform illuminated background extracte opening.
(c) Intensity surface of the background in 3-D spac (d) Pavement image after removal of non-uniform
B) Image Enhancement by Wavelet Deno Wavelet denoising attempts to rem present in the gray image while prese sharpness. The common approach to denoising is to transform the signal in
domain, shrink the detail coefficients and to the image domain. A more precise ex wavelet denoising procedure can be gi Assume that the observed data is,
()()()
X t S t N t =+
NIQUES FOR
ION
n Effects
ackground from
ound subtraction ons is introduced kground lighting is applied by ion) followed by with the same
mination of small e preserving the the non-uniform ssfully extracted. rom the original umination effect tep, we add the
nd back to every mage with more
ure 1 shows the
uniform illumination ed by morphological ce.
illumination effect. oising
move the noise erving the edge wavelet based
nto the wavelet d transform back xplanation of the iven as follows.          (1)
where S (t ) is the uncorrupted
N (t ). Let W  and 1
W − denot wavelet transform operator denoising operator with th
denoise X  (t ) to recover  ()S
t process of the thresholdin consists of the following thre
Y Z D S ===The soft thresholding operato
()(|(,)    0        sign Y D Y λ­=®
¯An appropriate choice of fundamental to the effect
wavelet denoising procedur remove important parts of the
a threshold that is too small reconstruction.
(a)              Figure 2 (a) Image before wav (b) Image after wavel
C) Thresholding and Segmen The selection of an appro a very important role in the en the mapping of the distress fe In most cases, cracks only t entire image. Thus, acco distribution, some dark nois fallen leaves, and oil stains e as cracks. Therefore a sim
method is not practical for th Instead of processing the each tested image is d
non-overlapping blocks (sho purpose of local segmentatio important parameter for this t window shoul
d be large enou and background regions w local information.
Since image blocks conta
blocks” hereafter) tend to intensity but higher stand
non-crack blocks, we can d signal with additive noise te the forward and inverse s, and (~,)D λ denote the
hreshold λ. We intend to  as an estimate of S  (t ). The ng method for denoising ee steps: 1()
(,)()W X Y W Z λ−                (2) or (,)D Y λ is defined as:
|)  ||                Y if Y otherwise λλ−>
(3) the threshold value Ȝ is
tiveness of the described e. A large threshold might
e underlying signal, whereas l would retain noise in the
(b)
velet denoising.
let denoising. ntation
opriate threshold value plays ntire process since it defines features in the binary image. take a small portion of the ording to the histogram sy pixels, such as shadow, tc. are easily misrecognized
mple global segmentation his type of application.
3tiles
pavement image as a whole,
divided into a set of own in Figure 3 (a)) for the on. The block size is a very type of application. A single ugh to contain both the crack while representing only the aining cracks (called “crack present a lower average
dard deviation than those
n roughly distinguish the
crack-blocks from the non-crack blocks. Then a common thresholding method based on Otsu’s algorith
m [8] is applied to these crack blocks to obtain a set of local threshold values. Otsu’s method is one of the most successful methods for image thresholding based on the statistics of the gray level histogram. It selects the global optimal threshold by maximizing the between-class variance. The average of these local threshold values is selected to be the global threshold value for the segmentation. The process leads to lower threshold values which can effectively reduce the misclassification caused by low gray value noise. Once an appropriate threshold value is determined, the pixels with gray levels below the threshold are classified as distress pixels and pixels whose gray level values exceed the threshold are assigned to background. Figure 3 (b) shows the binary image obtained from the proposed segmentation algorithm.
(a) (b)
Figure 3 (a) Subdivide the pavement image into non-overlapping
blocks.
(b) Binary image obtained from proposed segmentation algorithm.
D)  Morp h ological Operations
In this step, morphological closing is first applied in order to link the disconnected crack pixels by filling small holes and bridging the thin gaps that appear in the binary image. A line structuring element (SE) of size 15 pixels was used for the closing operation.
(a) (b)
Figure 4 (a) Binary image obtained from local segmentation.
(b) Binary image after morphological operation.
After linking the crack pixels, a binary noise removal approach based on the size of the connected components is applied. Any connected components with pixel numbers less than a predefined threshold value is considered to be noise, and therefore, will be removed. The result of this process is shown in Figure 4, where it can be seen that most of the noise has been removed and the small gaps between cracks are filled with object pixels.
III.CRACK CLASSIFICATION
Cracks extracted through the process as explained in the previous section can be classified into different types. In our research, a projection-based pavement distress classification system is develope
d, taking into account four distress types: longitudinal cracks, transverse cracks, diagonal cracks, and alligator cracks.
A) Crack Representation
In order to save processing time and reduce memory storage requirements, the resulting binary image is then subdivided into square tiles, each of which stands for either a cracked tile or non-cracked tile. The decision to classify a tile as a crack tile is based on the percentage of crack pixels present in a tile. Any tile with more than 10% of its pixels being crack pixels is considered as a crack tile, therefore labeled ‘1’, otherwise it is labeled with ‘0’. The size of the tile depends on the width of the crack. The result obtained at this stage is shown in Figure 5.
(a) (b)
Figure 5 (a) Original pavement image.
(b) Pavement image represented by cracked tiles.
B) 2-D Feature Mapping
To represent the distribution of crack tiles, two types of histograms, vertical histogram and horizontal histogram [9], are recorded by accumulating the number of crack tiles presented in each column and row as follows,
1
[]_[,]
M
j
H i crack tiles i j
=
=¦, i =1, 2…N (4)
1
[]_[,]
N
i
V j crack tiles i j
=
=¦  , j=1, 2…M    (5) where V and H represent vertical and horizontal histograms, M and N denote the number of rows and columns, respectively.
These two histograms can demonstrate a clear pattern in the crack. If the crack is developed in a longitudinal direction, there would be a peak in the vertical histogram.  On the other hand, if the crack is
developed in the transversal direction, there would be a
peak in the horizontal histogram. If the alligator type, the peaks could be found and horizontal directions. For a block c
also be found in both histograms but magnitude than those of an alligator crac
Based on the above observation, t space used for crack classification is c standard deviations of the vertical (fe
horizontal (feature two) histograms,
V σ
H σwhere
μ represents the mean value of Figure 6 Crack tiles distribution The result of the 2-D feature space Figure 7. As can be seen from the graph, L2 partition the 2-D feature space into
region 1, region 2, and region 3, repres types of cracks. Once the mapping point this space, the crack is classified ac following rules:  y  If the feature point maps into regi feature space, the crack is classified
crack.
y  If the feature point maps into regi feature space, the crack is c
longitudinal crack.
y
If the feature point maps into regi feature space and the percentage of more than 20% of the entire imag
classified as an alligator crack.
y
If the feature point maps into regi
feature space and the percentage of
less than 20% of the entire imag
e crack is o
f an
in both vertical
crack, peaks can with a smaller
ck.
the 2-D feature omposed of the
eature one) and
(6)
(7)
the histogram.
n.
e is depicted in
the lines L1 and
o three regions, senting different t is positioned in ccording to the
on 1 of the 2-D
d as a transversal on 2 of th
e 2-D
classified as a
on 3 of the 2-D the crack tiles is
ge, the crack is on 3 of the 2-D the crack tiles is
ge, the crack is classified as a block cra The result of the above Figure 7. Since the feature p
crack is classified as a transv Figure 7 2-D feature sp
IV. EXPERIME In this section, 200 actual road surface are tested. Fig classification results extracte
Figure 8 Original images (left colum column); crack classification f
Two kinds of criteria ar system’s performance. One i other is the accuracy of th detailed test results of the pro ack.
example is also shown in
point falls into region 1, the versal crack.
pace for classification.
ENTAL RESULTS l digital images taken from a gure 8 shows sample crack ed from the image database.
mn); crack detection results (center
feature space (right column).
re considered to assess the is computation time and the he system. Table 1 shows oposed system.
Region I
Region I
Region I
Region I
R egion I II
Re gion  II
Re gion  II
Re gion  II
Re gion  II
R egion I II
R egion I II
R egion I II
Table 1 Test results of the system.
Crack type Number
of test
images Computation
time
Classification
accuracy
Transversal 50    2.2
s 100% Longitudinal 50    1.9
s 100% Block 50
2.4
s 96% Alligator 50 2.5
s 98%
From Table 1 we can see that the system takes only
2.3 seconds on average on a computer equipped with
Intel(R) core(TM) 2Duo CPU to determine a crack type
for a pavement image of size 512*512 pixels.  This is
near real time and produces an average accuracy of 98.5%
for all classes of the detected cracks.
V.CONCLUSION
This paper presents a low-cost, user-friendly and fast pavement distress detection and classification method using advanced image processing techniques. It has been shown that the proposed pavement analysis system allows complete automation with near real-time evaluation of pavement distresses. More importantly, the accuracy of this system in identifying pavement distress meets the standards set out by the road authority for pavement management. The experimental results indicate that our proposed system produces highly reliable and accurate results from the 200 tested samples.
The application developed in this research was mainly focused towards distress detection and classification. Future developments will target the analysis of the crack properties, such as width, length and severity of the cracks.
ACKNOWLEDGMENT
This work was in part supported by a grant from the Michigan Ohio University Transportation Center (MIOH-UTC), U.S. Department of Transportation which
is greatly appreciated.
REFERENCES
[1]  C.P. Wang, Q. Watkins and S.R. Kuchikulla. “Digital Distress
Survey of Airport Pavement Surface”. Federal Aviation
Administration Airport Technology Transfer Conference. 2002.
v
[2]  B.J. Lee and H.D. Lee. “A Robust Position Invariant Artificial
Neural Network for Digital Pavement Crack Analysis”.
Technical Report TRB2003-000996, 2003.
www.ltrc.lsu.edu
[3]  C.J. Hsu, C.F. Chen, C. Lee, and S.M. Huang. “Airport
Pavement Distress Image Classification Using Moment
Invariant Neural Network”. 22nd Asian Conference on Remote
Sensing and processing (CRISP), Singapore, 2001.
[4]Kelvin C.P. Wang and W. Gong. “Automated Pavement Distress
Survey: A Review and A New Direction”. Pavement Evaluation
Conference, 21-25, Roanoke, Virginia., 2002.
[5]  D. Meignen, M. Bernadet and H. Briand, “One Application of
Neural Networks for Detection of Defects Using Video Data
Bases: Identification of Road Distresses”, Proceedings of Eighth
International Workshop on Database and Expert Systems Applications, pp 459 – 464, USA, Sep 1997.
[6]H. Cheng and M. Miyojim, “Automatic Pavement Distress
Detection System”, Journal of Information Sciences 108, ELSEVIER, pp 219-240, July 1998.
www1.maths.lth.se/bioinformatics/calendar/20031209/A
Bengtsson-MicroarrayImageAnalysis-20031209.pdf
[7]  A. Bengtsson. “Microarray Image Analysis: Background
Estimation using Region and Filtering Techniques”. Master’s
Thesis in Mathematical Sciences, Mathematical Statistics, Centre for Mathematical Sciences, Lund Institute of Technology,
2003.
[8]P.S. Liao, T.S. Chen and P.C. Chung, “A Fast Algorithm for
Multilevel Thresholding”. Journal of Information Science and
Engineering. V ol.17, pp713-727, 2001.
[9]H. Oliveira and P. L. Correia, “Identifying and retrieving
Distress images from road pavement surveys”, Proc. 1st ICIP.
Workshop on Multimedia Information Retrieval: new trends and
challenges, ICIP 2008, Las Vegas, USA, 2008.

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。