IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 65, NO.

3, MARCH 2016 583

Automatic Crack Detection and Measurement

Based on Image Analysis
Romulo Gonçalves Lins and Sidney N. Givigi, Senior Member, IEEE

Abstract— Crack detection and measurement in civil structures and, in addition, sometimes need special equipment to improve
has been a constant field of research. Conventionally, a technician the accuracy of the measurement, which may be time consum-
is responsible to detect and measure cracks in the field. In this ing and strenuous, consequently lowering their accuracy. Due
paper, a system based on machine vision concepts has been
developed with the goal to automate the crack measurement to these facts, their automated use is not technically viable.
process. Using this method with only a single camera installed in a In this paper, the goal is to create a method able to detect and
truck or even in a robot, a sequence of images is processed and the measure cracks using only images acquired by a camera. The
crack dimensions are estimated. The experimental results validate proposed approach combines techniques of statistical filtering
the application of the proposed method for real structures. (particle filters) to recognize cracks automatically in images
Index Terms— Automation, image analysis, machine vision, and techniques of machine vision to measure the crack width
maintenance, particle filter, structural engineering. and length.
I. I NTRODUCTION The main contribution and novelty of the proposed approach
lies in the use of only a single camera to identify and

C RACK detection and measurement is an essential task to

assess civil structures. Civil infrastructure systems rep-
resent a significant fraction of global assets, and due to many
measure cracks online and in real time. Moreover, the use
of a particle filter for the identification of the crack introduces
an improvement over previous methods. For the experiments
factors, cracks will appear in the surface of these structures. described in this paper, a computer with Reduced instruction
These cracks not only affect the visual appearance of the set computing technology was used. However, implementation
structures but also lead to steel corrosion, which accelerates in a digital signal processor (DSP) or a field-programmable
concrete aging affecting the normal use of these structures gate array (FPGA) is possible with many advantages over a
and leading to underuse of the infrastructure and a shorter life software program operating on a conventional computer. For
span. Besides this, many structures have collapsed abruptly instance, the robustness, time of processing, and reliability
due to these effects, for example, the Minneapolis (USA) may be considerably improved [3].
bridge collapse in 2007, the Tuo River bridge collapse in In order to test the proposed methodology in this paper,
Sichuan province (China) in 2007, and the Pan River bridge several experiments were conducted in a controlled environ-
collapse in China in 2013. These accidents have highlighted ment and their results were compared with other methods.
the tragical results of structural failures and the importance of In addition, experiments with real structures were conducted
periodical structural maintenance. to validate the method.
In order to prevent disasters, the structures can be repaired This paper begins by describing the related works in
using the technique of concrete structure reinforcement. Section II. Section III discusses the proposed architecture,
Basically, this method consists in recognizing defects in the the problem formulation, and the proposed algorithms. The
structure and reinforcing it using materials such as carbon experimental results and analysis are provided in Section IV.
fiber or epoxy [1]. Out of the many methods developed to Finally, Section V presents the conclusion.
detect and measure cracks, three techniques have emerged in
the literature: human inspection, in which an operator inspects II. R ELATED W ORK
the structure and measures an identified crack using tradi- In the past few years, many techniques to detect and mea-
tional measurement systems; microscopic inspection of cracks sure cracks automatically have been studied. These techniques
using special tools; and machine vision inspection, in which are used to assess and monitor cracks in order to provide
cracks are identified and measured automatically by image safety for the users as well as to reduce the cost of main-
analysis [2]. The first two methods are performed manually tenance. Among these proposed techniques, approaches based
Manuscript received June 1, 2015; revised October 29, 2015; accepted on image analysis, neural networks, and pattern recognition
October 31, 2015. Date of publication January 8, 2016; date of current version have emerged as the most efficient ones.
February 5, 2016. The Associate Editor coordinating the review process was For example, local binary pattern, a technique based on the
Dr. Amitava Chatterjee.
The authors are with the Department of Electrical and Computer Engi- gray-scale and rotation invariant operator from images has
neering, Royal Military College of Canada, Kingston, ON K7K 7B4, Canada been used for crack detection. This method is widely used
(e-mail: [email protected]; [email protected]). for texture classification [4]. Another similar method is based
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. on classifying local neighbors into smooth and rough areas
Digital Object Identifier 10.1109/TIM.2015.2509278 of an image. Then the areas with coarse texture are further

0018-9197 © 2016 Canadian Crown Copyright

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 584 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 65, NO. 3, MARCH 2016

explored. Moreover, artificial neural networks and wavelet 1) RGB is a colorspace originated from cathode ray tube
transformation for pavement distress image compression noise display application, when it was convenient to describe
reduction [5] can be used when coupled with the same color as a combination of three colored rays (RGB). It is
technique. one of the most widely used colorspaces for processing
A study called bidirectional empirical mode decomposi- and storing of digital images [11].
tion (BEMD) has been conducted for crack edge detection. 2) HSV colorspace was introduced when there was a need
Different techniques for edge detection such as Canny filter for the user to specify color properties numerically.
and Sobel methods have been compared in the BEMD study. It describes color with intuitive values, based on the
The final conclusion is that different edge detection techniques artist’s idea of tint, saturation, and tone. Hue defines
are more efficient for different types of surfaces. For instance, the dominant color (such as red, green, purple, and
it was reported that the Canny filter is more efficient to detect yellow) of an area, saturation measures its colorfulness
cracks on asphaltic concrete, while Sobel methods works better in proportion to its brightness, and the value is related
for Portland cement concrete [6]. to the color luminance [12].
Studies have been conducted using multifrequency [7], A digital image is represented by a matrix of n rows and m
wherein a fractional differentiation is used for high, medium, columns, where the number of columns represents the image
and low frequencies. Then, a wavelet transform is used for width and the number of rows represents the image height.
noise reduction and a difference histogram method is applied Moreover, each element (Px y ) = [R, G, B]T of the matrix is
to detect cracks on pavement. Finally, from the histogram, the the color intensity of the pixel at column x and row y of the
image can be segmented and the crack is recognized [8]. image.
Pattern recognition is another approach used to detect and The color distribution is discretized into m b. In the pro-
to classify cracks. To handle the high amount of input data, an posed approach, the histogram is calculated in RGB standard
image database was split into subsets. The system was then space using (8 × 8 × 8) b to represent the color spectrum, and
trained by human experts who manually selected the crack mathematically, the color histogram is represented by
regions for training. Three parametric and three nonparametric ⎡ ⎤
R
supervised classification strategies were introduced to detect
h(rgb) = C ⎣G ⎦ (1)
the cracks. The cracks were then classified as longitudinal,
B
transversal, and mixed by exploring the 2-D feature space [9].
that assigns the color at location (Px y ) corresponding to a
III. P ROBLEM F ORMULATION
matrix element (pixel).
In Section III-A, an explanation about the proposed archi- Color feature extraction is the technique to search for a
tecture is provided. Section III-B discusses the crack detection specified color composition in an image in comparison with
algorithm based on particle filter. Finally, Section III-C intro- a standard query color. Color feature extraction is based
duces the crack measurement algorithm to estimate width and on the observation of how often a color is used to encode
length of a crack using only a single image. functionality: grass is green, sky is blue, and so on [13].
Mapping the colors in an image L into a discrete color space
A. Architecture of the Proposed System containing a color h(rgb), each matrix element histogram
The use of cameras as the main sensor has several advan- h(Px y ) can be compared with the standard histogram h(rgb)
tages over traditional techniques [10]. However, a correct setup by Euclidean distance. Mathematically, this comparison is
of the components that make up the system is important to
Di = h(rgb) − h(Px y ) (2)
ensure the accuracy of measurements in the field.
Fig. 1 shows the block diagram of the crack measurement where nm is the matrix element position in an image L.
system proposed. The system is composed of an IP camera After obtaining the distances of each matrix element (Px y ),
to acquire images from the environment, a processing device weights are assigned to each particle using Gaussian
to process the crack detection and the crack measurement distribution
algorithms, a screen to show the results in real time, a remote  
1 Di2
station to receive data from the processing device through Wi = √ exp − 2 (3)
a communication channel, and a database used to create a 2πσ 2σ
measurement history. Although the database appears as part of where σ is the variance for the random distribution. In our
the architecture, it will not be used for comparison purposes case, it represents the color spread around a crack and it ranges
as our focus is on the instrumentation and measurement. The from [−2, 2].
database is used to keep the history of the state of the cracks Particle filtering was originally developed to track objects
for maintenance purposes. in clutter. The state of a tracked object is described by the
vector X t , while the vector Z t denotes all the observations
B. Crack Detection Algorithm {z 1 , . . . , z t } up to time t [14]. Particle filters are often used
The particle filter proposed to recognize cracks is based on when the posterior density p(X t |Z t ) and the observation
a color model context. Color models describe different aspects density p(zt |X t ) are non-Gaussian.
of an image. Two models have emerged in the literature: The key idea of particle filtering is to approximate
red, green, and blue (RGB) and hue–saturation–value (HSV). the probability distribution by a weighted sample set

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 LINS AND GIVIGI: AUTOMATIC CRACK DETECTION AND MEASUREMENT BASED ON IMAGE ANALYSIS 585

Fig. 1. Block diagram of the crack measurement system proposed.

Algorithm 1 Crack Detection Algorithm TABLE I

C RACK C LASSIFICATION IN A CCORDANCE
Input: Images from cameras W ITH I TS G EOMETRICAL F ORM
Output: Cracks recognized on screen ⎡ ⎤
R
1: Target color identification using the h(rgb) = C ⎣G ⎦
B
2: Initialize the particle filter with N = 4000
3: Input the color model and measurement
4: Re-sample the weight π (n) of each particle according a
likelihood model map all the cracks. However, if larger cracks are expected, the
5: Position estimation of the target color multiplying the designer could increase this number.

N
particle’s position with their respective weight E = Wi x i C. Crack Measurement Algorithm
i=1
6: Show the particles detected on a computer screen The proposed method to estimate the width and length
7: Record the pixel location to apply them into Crack of a crack from digital images uses the pixel resolution
Measurement Algorithm multiplied by the number of pixels detected by the crack
detection algorithm. However, the geometric form of a crack
has determinant influence under the proposed technique.
S = {(s(n) , π (n) )|n = 1. . .N}, where N is the number of Usually, cracks can be classified in accordance with their
particles used. Each sample s represents one hypothetical geometrical form as vertical, horizontal, diagonal, or com-
state of the object, with a corresponding discrete sampling plex [15]. As shown in Table I, the incidence of complex
probability π and n = 1, . . . , N, where geometry refers to only 6% of all cases [15]. Therefore, we
propose an algorithm to find the normal direction of the cracks.
N
The proposed approach excludes the complex cracks geometry
π (n) = 1. (4) due to its relative rarity. However, with our proposed method,
n=1 we can measure 94% of all cases.
The evolution of the sample set is described by propagating The crack detected by Algorithm 1 is used in Algorithm 2
each sample according to a system model. Each element of the for the measurement of its width and length. At this point, it
set is then weighted in terms of observations and N samples is important to note that the angle can be calculated using a
are drawn with replacement, by choosing a particular sample straight line created using two points of a crack and applying
(n)
probability π (n) = p(zt |X t = st ). The mean state of an trigonometric concepts to work out its angle.
object is estimated at each time step as From data inputted by the crack detection algorithm,
Algorithm 2 classifies and estimates the width and length of
N
cracks. From the classification (complex, diagonal, vertical,
E[S] = π (n) s (n) . (5) or horizontal), if a crack is classified as complex, the algo-
n=1
rithm outputs a message (Unable to measure the geometry).
Algorithm 1 shows the sequence needed to implement If the crack is classified as vertical or horizontal, the width
the crack detection algorithm. The number of particles used and the length can be estimated using the number of pixels of
is 4000. This number was chosen as a tradeoff between quality the section multiplied by camera pixel resolution [16] and the
and processing speed. In our experiments, 4000 is adequate to algorithm outputs the final results.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 586 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 65, NO. 3, MARCH 2016

Algorithm 2 Crack Measurement Algorithm

Input: Data from Crack Detection Algorithm in pixels
Output: Estimation the width and length of a crack (mm)
Input pixels position from crack detection algorithm
if the crack is recognized as complex then
Display the message (“Unable to measure the geometry”)
and close the application Fig. 3. Pixel map of a crack (a), the detection of the pixels of the crack by
Algorithm 1, and largest line in the curve that goes through p(x0 , y0 ).
else if the crack is classified as diagonal then
Determine the normal direction of crack in the curve
Estimate the width and length and show them on the screen
else
Display a message for the user (“the crack is classified as
vertical or horizontal”)
Estimate the width and length and show them on the screen
and close the application
end if

Fig. 4. Line l and its slope from least-square fitting.

Fig. 5. Sub-block T  resultant of the pixel rotation.

Fig. 2. Diagonal cracks with low cross section variation (a) and with high
cross section variation (b).
points with a least-squares algorithm (Fig. 4), we obtain a
line l; y = ax + b that contains the point p(x 0 , y0 ). Therefore,
However, to estimate the width of a diagonal crack, as now we have the slope of the curve.
shown in Fig. 2(a), or of a diagonal crack with a curve, as For example, in the sub-block T of Fig. 3(c), we find M = 9
shown in Fig. 2(b), we propose a technique that determines and |S P | = 25. All points are plotted in Fig. 4 and the
the normal direction of cracks in the curves. Usually, the line l is
curves are made up of some adjacent target pixels and they
can be described mathematically through the methods of y = 1.1481x − 0.0815. (6)
differentiation and difference [17]. In this paper, we determine Now, to estimate the crack width, we must apply a rotation
the normal direction of cracks by the coordinates of their cross in the sub-block T in order to linearize the curve. Assuming
section in pixels with the least-squares method. that point p(x 0 , y0 ) is the center of rotation, we find the angle
We start by establishing an arbitrary point p(x 0 , y0 ) in the normal to the curve at point p as θ = tan−1 (−1/a), where
crack curve. We then find the tangent to the curve at the a is the slope of the line l : y = ax + b.
selected point p and calculate its slope. From the slope, we With p(x 0 , y0 ) as the point of rotation, (X, Y ) ∈ T  is the
calculate the angle θ normal to the curve. rotation of sub-block (x, y) ∈ T by θ , that is
In order to calculate the slope, we zoom in the image around
the point p(x 0 , y0 ) including all the pixels that were detected X = (x − x 0 ) cos θ + (y − y0 ) sin θ + x 0
in Algorithm 1 [Fig. 3(b)]. Let us now define a block that Y = −(x − x 0 ) sin θ + (y − y0 ) cos θ + y0 .
includes the neighborhood of p(x 0 , y0 ). This neighborhood is
As result of each pixel rotation, we can plot a sub-block T 
centered in p(x 0 , y0 ) with a size M × M and we call it T . The as shown in Fig. 5 and using the proposed approach to rotate
dimension M is chosen based on the largest square at which a pixel sub-block, we can calculate the width of every cross
vertices and p(x 0 , y0 ) are placed on the curve [Fig. 3(c)]. Then section on the picture and adopt its average wcrack as the width
taking out the pixels sub-block T and recording all coordinates in pixels
of the crack pixels (establishing the top left corner of T as the 
origin (0, 0), and the down direction as the x-axis and the pcs
wcrack = (7)
left direction as the y-axis), we end up with a set of points n
S P = {(x i , yi ) ∈ c}, where c is the set of points that are where pcs is the number of pixels in each cross section
part of the crack. In the last step, by linear fitting of these and n is the number of cross sections.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 LINS AND GIVIGI: AUTOMATIC CRACK DETECTION AND MEASUREMENT BASED ON IMAGE ANALYSIS 587

Fig. 6. Experimental setup to test the proposed methodology.

This procedure is repeated for several points in the curve in Fig. 7. Specimen model designed with its respective dimensions in
order to assess the average width of the crack. Alternatively, millimeters.
one could save the width of predetermined areas of crack
(borders, midsection, etc.) for future assessment.

IV. E XPERIMENTS
This section describes the experiment to assess the feasibil-
ity of the proposed methodology. In Section IV-A, an expla-
nation about the experimental setup is provided. Section IV-C
discusses the application of the proposed methodology.
Section IV-D shows the practical results obtained. Finally,
Section IV-E brings a discussion about the final results.

A. Experimental Setup Fig. 8. Sample of image used for camera calibration.

The experimental setup is composed of a Clearpath Husky
robot with an IP camera mounted on it. The robot is used to
ease the acquisition of images required by Algorithms 1 and 2. this paper, the only needed parameter is the pixel resolution
In addition, an experimental bridge was built in order to of the camera, which is Pr = 0.3776. However, note that
support the specimen. As shown in Fig. 6, while the robot this value is used without any loss of generality and if a new
maneuvers around the environment, images are acquired and camera was to be used, a new calibration procedure must be
processed at the same time with the objective of detecting and performed.
measuring the cracks located on the specimen. The specimens
are manufactured using 3-D printing technology to ensure B. Experimental Environment
accuracy of its dimension. Fig. 7 shows a specimen model
Apart from the experiments with fabricated structures, as
with its respective dimensions to be manufactured.
shown in Fig. 6, experiments with real structures were also
We notice that the experimental setup was created in order to
performed. The experiments for the fabricated specimens
compare a real and accurate specimen with the results provided
were indoors and controlled. As such, light was constant
by the proposed methodology. However, in Section IV-C
(fluorescent bulbs) and no direct sunlight exposure was
to assess the technical feasibility of the method, we have
allowed. This follows similar experiments in [19]. These
processed images from real concrete structures to complement
experiments were included in the study in order to provide
the final results.
a baseline for comparison, as we could manipulate the type
In order to measure the crack width and length, it is
and dimensions of cracks.
necessary to know the camera pixel resolution, and a procedure
The experiments with real structures were not controlled and
of calibration camera is necessary for its estimation. A camera
light was allowed to vary (natural light). These experiments
calibration routine for MATLAB is a convenient tool and was
were included in order to validate our approach in inspection
used for this task. This toolbox allows calibration of cameras
environments that could be used in actual applications.
and estimation of the intrinsic (Focal Length) parameter using
only a chessboard target [18]. From a sequence of images of
the chessboard in different positions, the toolbox returns the C. Application of the Crack Detection Algorithm
estimated parameter. Fig. 8 shows the toolbox processing of The experiment follows the methodology proposed
images used for calibration. in Algorithm 1. Three trial runs with different materials were
The routine outputs many parameters, such as the intrinsic performed, two of them in a controlled environment using
matrix of calibration, lens distortion, and others. However, for two different specimens and one more using real concrete

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 588 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 65, NO. 3, MARCH 2016

TABLE II
C RACK D IMENSION U SING THE C RACK
D ETECTION A LGORITHM ( IN P IXELS )

Fig. 9. Cracks recognized in a controlled environment.

if necessary. For the case Specimen 2 [crack cross section

variation (CCSV)], there were three measurements in the curve
due to its CCSV.

D. Application of the Crack Measurement Algorithm

Fig. 10. Cracks recognized in a real structure. To estimate the crack dimensions, the data from crack
detection algorithm is inputted into the crack measurement
structure of buildings in our campus. Then all data was algorithm. Using the width and length in pixels as shown in
processed in order to detect cracks using only images. Table II and the pixel resolution estimated by the procedure
Inputting a sequence of images Ste , where e is the trial of calibration discussed in Section IV-A, the algorithm is able
number and t is time, each sample is processed and the crack to process all data and outputs the crack measurement.
detection algorithm returns a figure showing (in red) a crack In order to verify the accuracy of the proposed methodology,
recognized. Fig. 9 shows crack detection in the controlled the cracks have been measured using a digital vernier caliper
environment, whereas Fig. 10 shows crack detection of a real with an uncertainty of ±0.01 mm. From the dimensions
structure. measured by the instrument, the error between the real dimen-
Besides the figure containing the identified crack (similar sions of the cracks and the results of the crack measurement
to Fig. 9 and Fig. 10), the algorithm returns the position algorithm can be calculated. For the length, the error is
(in pixels) of each particle of the detected crack. For cracks calculated as
classified as vertical, the width and the length can be estimated lvpm − lcm
in pixels as Le = ∗ 100 (8)
lcm
w p = x 2 − x 1 and l p = y2 − y1 where L e is the length error, lvpm is the length value measured
by proposed methodology, and lcm is the true value of length
where w p is the width in pixels, l p is the length in pixels, crack measured by the vernier caliper.
and the coordinates x 1 , x 2 , y1 , and y2 are returned by the Likewise, the width error is calculated as
algorithm. Fig. 9 shows a vertical crack detected and its points wvpm − wcm
extracted by the crack detection algorithm. In the case of We = ∗ 100 (9)
wcm
a crack classified as horizontal, we must invert the x- and
where We is the width error, wvpm is the width value measured
y-axis in order to standardize the width and length for the
by proposed methodology, and wcm is the true value of width
measurement to take place. Finally, when a diagonal or a curve
crack measured by the vernier caliper.
crack is detected, the algorithm outputs wcrack as proposed in
The results for all cracks are summarized in Table III.
Section III-C. A summary of the width and length of each
Table III includes the error calculated as introduced
detected crack is shown in Table II.
in (8) and (9). The next section discusses these results and
The crack detection algorithm then inputs this data
compares them with the literature.
(in pixels) into the crack measurement algorithm allowing it
to estimate the physical dimensions of the detected cracks.
For real structure cracks, the width and length are an aver- E. Analysis of the Results
age due to their geometrical form. Note that using the As shown in Section IV-C, the crack detection algorithm
proposed measurement, one could estimate the cracks’ area developed in this paper is able to recognize cracks in different

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 LINS AND GIVIGI: AUTOMATIC CRACK DETECTION AND MEASUREMENT BASED ON IMAGE ANALYSIS 589

TABLE III
C RACK D IMENSION P ROCESSED U SING THE C RACK M EASUREMENT A LGORITHM

materials such as concrete, plastic, and others. Using the TABLE IV

RGB color model combined with a particle filter, it was E XPERIMENTAL F EATURES
possible to analyze an image and search for cracks with great
accuracy. In terms of computational performance, along the
several trials accomplished, each image was processed on
average in approximately 6 s.
Section IV-D presents the results of the crack measurement
algorithm. From the geometry classification, the algorithm
is able to output the dimensions of a crack using only the
used in the experiments, the environment navigated when
pixels positions and camera resolution (Table III). The results
the experiments have been conducted, and the average error
demonstrate that the method can measure real cracks with
range (%) for each methodology. Even though the resolution
an uncertainty type A of ±5.48% of the measured value
of the camera used in the proposed methodology is much
according to the guide to the expression of uncertainty in mea-
lower, the results obtained are better than the ones reported
surement [20]. This uncertainty is caused by several different
in [21] and [22].
sources such as illumination of environment, inaccuracies in
In terms of computation, we were not able to find any esti-
the camera calibration, and others. In addition, the camera
mation of computational resources in the literature. Therefore,
resolution has affected directly the accuracy, and therefore,
we could not establish any comparison. However, in compar-
in order to improve the results, a high-resolution camera could
ison with manual techniques used nowadays, our proposed
be used. Finally, in terms of computational performance, from
method is faster and more efficient as it takes only on average
the cracks detected by the crack detection algorithm, each
6 s to detect and measure cracks.
image was processed on average in approximately 2 s.
Another important feature to be highlighted is the technique
to estimate the crack width in a curve. Although the technique V. C ONCLUSION
cannot be used to estimate the width of the crack in its total This paper presented a new methodology to detect and
extension, the results show that it can be used to estimate the measure cracks using only a single camera. The proposed
width of a section in small areas of the curve, simplifying methodology allows for automatic crack measurement in civil
the processing, while still resulting in good accuracy, even to structures.
measure a variable crack cross section. According to the methodology, a sequence of images is
To check the applicability of the proposed methodology, processed by the crack detection algorithm in order to detect
our method is compared with other vision methods. As shown the cracks. The algorithm receives images as inputs and
in Table III, the average error of the proposed methodology outputs a new image with red particles along the detected
was estimated at 8.59% for the length and 7.51% for the width. crack. The pixel positions of the particles are stored in a vector
In comparison with other methods in the literature, the pro- that is passed along to the crack measurement algorithm. With
posed methodology proved more accurate than the vision these pixel positions, the algorithm estimates the number of
system developed in [21], where the average error has been pixels in a cross section and outputs the crack dimension.
estimated at 13.9% as well as in [22], where the average error The statistical analysis demonstrates that the error is small
has been estimated at 15%. on average permitting the accurate measurement. Furthermore,
Other important features about the comparison are shown compared with other methods in the literature, the methodol-
in Table IV that list the camera resolution (in megapixels) ogy presented in this paper fares very well.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.
 590 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 65, NO. 3, MARCH 2016

A few enhancements can be proposed to improve the [12] C.-C. Lai and Y.-C. Chen, “A user-oriented image retrieval system based
measurement accuracy. First, a camera with higher resolution on interactive genetic algorithm,” IEEE Trans. Instrum. Meas., vol. 60,
no. 10, pp. 3318–3325, Oct. 2011.
can be used in order to improve the accuracy. In order to assess [13] W.-T. Chen, W.-C. Liu, and M.-S. Chen, “Adaptive color feature extrac-
the resolution influence in the final result, a comparison of one tion based on image color distributions,” IEEE Trans. Image Process.,
of our specimens was performed. Using a 12 Mpixel resolution vol. 19, no. 8, pp. 2005–2016, Aug. 2010.
[14] G. G. Rigatos, “Particle filtering for state estimation in nonlinear
camera, the same crack (specimen 1), detected with a width industrial systems,” IEEE Trans. Instrum. Meas., vol. 58, no. 11,
of 3 pixels using the IP camera mounted on the robot, was pp. 3885–3900, Nov. 2009.
detected with 16 pixels using a high-resolution camera. This [15] S.-N. Yu, J.-H. Jang, and C.-S. Han, “Auto inspection system using
a mobile robot for detecting concrete cracks in a tunnel,” Autom.
is translated into an improvement of accuracy (the error was Construction, vol. 16, no. 3, pp. 255–261, 2007.
reduced from 8.12% to 5.27%) and the uncertainty is reduced [16] T. Edeler, K. Ohliger, S. Hussmann, and A. Mertins, “Super-resolution
to ±1.97%. However, in order to have fair comparisons with model for a compressed-sensing measurement setup,” IEEE Trans.
Instrum. Meas., vol. 61, no. 5, pp. 1140–1148, May 2012.
the literature, our images were collected with normal low- [17] J. Mukhopadhyay, P. Bhowmick, P. P. Das, S. Chattopadhyay, and
resolution cameras that are usually mounted on robots. Second, B. N. Chatterji, Digital Geometry in Image Processing. Boca Raton,
the crack detection algorithm can be optimized in order to FL, USA: CRC Press, 2013.
[18] R. G. Lins, S. N. Givigi, and P. R. G. Kurka, “Vision-based measurement
minimize processing time. Therefore, the algorithms can be for localization of objects in 3-D for robotic applications,” IEEE Trans.
ported to a dedicated DSP or an FPGA chip in order to be Instrum. Meas., vol. 84, no. 11, pp. 2950–2958, Nov. 2015.
used in the field. [19] S. Foix, G. Alenyà, and C. Torras, “Lock-in time-of-flight (ToF)
cameras: A survey,” IEEE Sensors J., vol. 11, no. 9, pp. 1917–1926,
Sep. 2011.
R EFERENCES [20] Joint Committee for Guides in Metrology, “JCGM 100:2008, evalua-
tion of measurement data—Guide to the expression of uncertainty in
[1] J. Sim, C. Park, and D. Y. Moon, “Characteristics of basalt fiber as a measurement, (GUM 1995 with minor corrections),” 2008.
strengthening material for concrete structures,” Compos. B, Eng., vol. 36, [21] K. Khalili and M. Vahidnia, “Improving the accuracy of crack
nos. 6–7, pp. 504–512, 2005. length measurement using machine vision,” Procedia Technol., vol. 19,
[2] M. L. Wang, J. P. Lynch, and H. Sohn, Eds., Sensor Technologies for pp. 48–55, Oct. 2015.
Civil Infrastructures: Sensing Hardware and Data Collection Methods [22] W. Zhang, Z. Zhang, D. Qi, and Y. Liu, “Automatic crack detection and
for Performance Assessment, vol. 1. New York, NY, USA: Elsevier, classification method for subway tunnel safety monitoring,” Sensors,
2014. vol. 14, no. 10, pp. 19307–19328, 2014.
[3] S. Jin et al., “FPGA design and implementation of a real-time stereo
vision system,” IEEE Trans. Circuits Syst. Video Technol., vol. 20, no. 1,
pp. 15–26, Jan. 2010.
[4] Y. Hu and C.-X. Zhao, “A local binary pattern based methods for
pavement crack detection,” J. Pattern Recognit. Res., vol. 5, no. 1, Romulo Gonçalves Lins received the Ph.D. degree
pp. 140–147, 2010. in mechanical engineering from the State University
[5] P. Subirats, J. Dumoulin, V. Legeay, and D. Barba, “Automation of pave- of Campinas, Campinas, Brazil.
ment surface crack detection using the continuous wavelet transform,” He joined the Department of Electrical and
in Proc. IEEE Int. Conf. Image Process., Oct. 2006, pp. 3037–3040. Computer Engineering, Royal Military College of
[6] A. Ayenu-Prah and N. Attoh-Okine, “Evaluating pavement cracks with Canada, Kingston, ON, Canada, in 2014, as a
bidimensional empirical mode decomposition,” EURASIP J. Adv. Signal Post-Doctoral Fellow. His current research interests
Process., vol. 2008, no. 1, p. 861701, 2008. include signal and image processing, instrumenta-
[7] A. Bernieri, G. Betta, L. Ferrigno, and M. Laracca, “Crack depth tion, and robotics.
estimation by using a multi-frequency ECT method,” IEEE Trans.
Instrum. Meas., vol. 62, no. 3, pp. 544–552, Mar. 2013.
[8] R. Gutkin, C. J. Green, S. Vangrattanachai, S. T. Pinho, P. Robinson, and
P. T. Curtis, “On acoustic emission for failure investigation in CFRP:
Pattern recognition and peak frequency analyses,” Mech. Syst. Signal
Process., vol. 25, no. 4, pp. 1393–1407, 2011. Sidney N. Givigi (SM’14) received the Ph.D. degree
[9] H. Oliveira and P. L. Correia, Supervised Crack Detection and Classifi- in electrical and computer engineering from Carleton
cation in Images of Road Pavement Flexible Surfaces. Rijeka, Croatia: University, Ottawa, ON, Canada.
InTech, 2009. He is currently an Associate Professor with the
[10] S. Vidas, R. Lakemond, S. Denman, C. Fookes, S. Sridharan, and Department of Electrical and Computer Engineer-
T. Wark, “A mask-based approach for the geometric calibration of ing, Royal Military College of Canada, Kingston,
thermal-infrared cameras,” IEEE Trans. Instrum. Meas., vol. 61, no. 6, ON, Canada. His current research interests include
pp. 1625–1635, Jun. 2012. autonomous systems and robotics.
[11] X. Qu, S.-C. Wong, and C. K. Tse, “Temperature measurement technique
for stabilizing the light output of RGB LED lamps,” IEEE Trans.
Instrum. Meas., vol. 59, no. 3, pp. 661–670, Mar. 2010.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 04,2023 at 11:26:56 UTC from IEEE Xplore. Restrictions apply.