You are on page 1of 10

Machine Vision and Applications (2011) 22:245–254

DOI 10.1007/s00138-009-0244-5

ORIGINAL PAPER

A robust automatic crack detection method from noisy


concrete surfaces
Yusuke Fujita · Yoshihiko Hamamoto

Received: 31 August 2008 / Revised: 30 September 2009 / Accepted: 24 December 2009 / Published online: 2 February 2010
© Springer-Verlag 2010

Abstract In maintenance of concrete structures, crack information can be used to decide the appropriate rehabil-
detection is important for the inspection and diagnosis of itation method to fix the cracked structures and prevent any
concrete structures. However, it is difficult to detect cracks catastrophic failure. In practice, cracks on concrete surfaces
automatically. In this paper, we propose a robust automatic are traced manually for diagnosis. In manual trace, cracks are
crack-detection method from noisy concrete surface images. evaluated by a human operator. Thus, qualitative assessment
The proposed method includes two preprocessing steps and of cracks, defects or failures, for diagnosis of concrete struc-
two detection steps. The first preprocessing step is a subtrac- tures, are significantly affected by operator’s experience, skill
tion process using the median filter to remove slight vari- level, and image quality. Hence, it is difficult to assess dete-
ations like shadings from concrete surface images; only an rioration objectively. Therefore, automatic crack detection
original image is used in the preprocessing. In the second is highly desirable for efficient and objective crack assess-
preprocessing step, a multi-scale line filter with the Hessian ment. Furthermore, it leads to quantitative estimation, auto-
matrix is used both to emphasize cracks against blebs or stains matic measurement of length or width of detected cracks,
and to adapt the width variation of cracks. After the pre- or automatic diagnosis of concrete structures, using image
processing, probabilistic relaxation is used to detect cracks processing and image recognition techniques.
coarsely and to prevent noises. It is unnecessary to optimize Because these images include various blebs, shadings and
any parameters in probabilistic relaxation. Finally, using irregularities in crack shape and size, in general, it is very
the results from the relaxation process, a locally adaptive difficult to detect cracks automatically from noisy concrete
thresholding is performed to detect cracks more finely. We surfaces. Thus, it is necessary to overcome these problems
evaluate robustness and accuracy of the proposed method in order to detect cracks automatically from noisy concrete
quantitatively using 60 actual noisy concrete surface images. surfaces with high accuracy.
Recently, many methods for crack detection by means
Keywords Nondestructive test · Concrete · Median filter · of visual inspection have been proposed [1–4], and some
Multi-scale · Relaxation · Thresholding techniques have been evaluated [5,6]. Ito et al. [1] proposed
an inspection system to detect and analyze cracks on the
concrete surfaces. Some image-processing techniques such
1 Introduction
as wavelet transform, shading correction, and binarization
are combined to detect cracks. However, in shading correc-
Automatic crack detection from concrete surface images tion, corrective images are needed for shading correction.
is very effective for nondestructive testing. The crack
Abdel-Qader et al. [5] compared the effectiveness of crack
detection among wavelet transform, Fourier transform,
Y. Fujita (B) · Y. Hamamoto
Sobel filter, and Canny filter. Hutchinson and Chen [2] used
Graduate School of Medicine,
Yamaguchi University, Yamaguchi, Japan Canny filter and wavelet transform for crack detection. How-
e-mail: y-fujita@yamaguchi-u.ac.jp ever, varying crack width in a concrete surface image is not
Y. Hamamoto considered. Yamaguchi et al. [3] proposed a crack-detection
e-mail: hamamoto@yamaguchi-u.ac.jp method based on percolation model. Binarization is included

123
246 Y. Fujita, Y. Hamamoto

after percolation processing. However, in that method, opti-


mization of threshold and procedure for edges, which are not
considered as cracks, are not considered. Kawamura et al. [4]
proposed use of a genetic algorithm for the semiautomatic
optimization of combining of image-processing techniques
or selection of parameters. In the system proposed by Lee
et al. [7], a supervised manipulation was included after auto-
mated process. These conventional methods need to choose
parameters or the combination of processes very carefully by
a human operator. Automatic detection is needed for more Fig. 1 Removal of serious variations like shading
effective and objective visual inspection.
Binarization is an important issue for automatic crack
detection. Many threshold selection techniques have been inal image, to obtain a corrective image. Then, the original
proposed [8–10]. However, global thresholding method may image is subtracted from the corrective image to obtain a
generate pieces of unconnected crack to avoid over-detection, subtracted image as the output of the first preprocessing. The
or it may generate noises, to detect cracks completely. On median filter takes all the local intensity distributions and
the other hand, in locally adaptive thresholding techniques, generates a new image corresponding to the set of median
a threshold is calculated at each pixel, which depends on values. If the size of the filter is several times greater than
some local statistics of pixel neighborhood. This method is the width of crack, cracks are removed from the corrective
not much helpful in assessing whether cracks exist in each image. In contrast, the slight variation remains in the correc-
local area of the image or not. tive image, but it is eliminated in the subtracted image. By
In this paper, we propose a robust automatic crack detec- this, the corrective image is generated from the original image
tion method for noisy concrete surface images. The method by the median filter. The subtracted image is obtained by
includes two preprocessing steps and two detection steps. 
median x j ∈Ri I (x j ) − I (xi )
We evaluate robustness and accuracy of the proposed method Is (xi ) = max , (1)
0
quantitatively using 60 actual noisy concrete surface images.
The remainder of this paper is organized as follows: we where I (x) means the intensity of the pixel x, and Ri means
describe the proposed method in Sect. 2. Experiments using the region around the pixel xi . Figure 2 shows an example of
actual images in Sect. 3 and discussion in Sect. 4 are given. the subtraction preprocessing.

2 Methods 2.2 Multi-scale line emphasis preprocessing

The proposed method includes two preprocessing steps and The line filter based on the Hessian matrix is used to empha-
two detection steps. The first preprocessing step is a subtrac- size lines such as cracks and to remove noises such as blebs or
tion process using the median filter to remove slight varia- stains. The second derivatives have been typically used for
tions like shadings from concrete surface images. Only an line enhancement filtering. Three-dimensional multi-scale
original image is used in the preprocessing, while the correc- line filter for medical images is proposed by Sato et al. [11].
tive image is also needed in the conventional method [1]. In In this paper, we use the Hessian matrix to measure similar-
the second preprocessing step, a multi-scale line filter with ity of a bright line that describes the second-order structures
the Hessian matrix is used both to emphasize cracks against of local intensity variations around each pixel in two-dimen-
blebs or stains and to adapt the width variation of cracks. sional images. The Hessian matrix has the partial second
Thus, it is possible to detect cracks of varying widths. After derivatives of images and the Hessian matrix of an image
the preprocessing, probabilistic relaxation is used both to I (x) is given by
 
detect cracks coarsely and to prevent noises. It is unnec- I x x (x) I x y (x)
∇ I (x) =
2
, (2)
essary to optimize any parameters in probabilistic relaxa- I yx (x) I yy (x)
tion. Finally, a locally adaptive thresholding is performed to
where partial second derivatives of the image I (x) are repre-
detect cracks more finely, using the result from the relaxation
sented by expressions as I x i y j (x) = ∂ x∂i ∂ y j , (i + j = 2). The
2
process.
combination of two eigenvalues of the matrix can express
2.1 Subtraction preprocessing the difference of the structure of a line, a blob, or a sheet
in the image. Figure 3 illustrates eigenvalues and eigenvec-
Figure 1 illustrates a removal of slight variations like shad- tors of the Hessian matrix in three structures. Let λ1 and
ings. First, we use the median filter, which smooths the orig- λ2 be eigenvalues (λ1 > λ2 ). When |λ2 |  |λ1 | ≈ 0, the

123
A robust automatic crack detection method from noisy concrete surfaces 247

Fig. 2 An example of subtraction preprocessing. (a) original image, (b) smoothed image with median filter (corrective image), (c) subtracted
image, where subtracted image is inverted in intensity level

(a) (b) (c)


Fig. 3 Eigenvalues and eigenvectors of Hessian matrix in a line struc- e2 , respectively. The vectors indicate direction of eigenvectors, and the
ture (a), in a blob structure (b) and in a sheet structure (c). λ1 and length of vector corresponds to the magnitude of the eigenvalue
λ2 mean eigenvalues, and their corresponding eigenvectors are e1 and

structure should be regarded as a line-like structure in the In the implementation, instead of computing directly the
two-dimensional space. When the value of λ1 is negative and second order derivative of the image, the Gaussian convolu-
|λ1 | ≈ |λ2 |  0, the structure should be regarded as blob- tion is combined with the second derivative in 4 in order to
like rather than line-like. When |λ1 | ≈ |λ2 | ≈ 0, it should be turn the filter response to the specific widths of lines as well
regarded as sheet-like structure. Because noise components as to reduce the effect of noise.
such as blebs or stains typically have blob-like shapes, it is 
important that the system possesses the ability to discriminate ∂2
I x i y j (x, σ f ) = G(x; σ f ) ∗ I (x), (i + j = 2), (4)
between lines and blobs. When the value of λ1 is positive, the ∂ xi ∂y j
structure involves concavity in the estimated line direction.
Therefore, we derived a generalized measure of similarity to where G(x; σ f ) is an isotropic Gaussian function with stan-
a line as, dard deviation σ f , and ∗ denotes the convolution. Using

the Hessian matrix ∇ 2 I (x; σ f ) based on the second deriv-
⎪ λ1
⎪ 2|λ | 1 + |λ2 | = |λ2 | + λ1
atives of Gaussian with σ f , λ12 (x; σ f ) as filter responses



⎪ can be turned to a specific width of line. We normalize
⎨ if λ2 ≤ λ
1 ≤ 0 the filter responses of each scale and then select the maxi-
λ12 = |λ2 | 1 − α λ1 = |λ2 | − αλ1 (3)

⎪ |λ2 | mum response among the multiple scales. The multi-scale



⎪ if λ2 < 0 < λ1 < |λα2 | integration of the filter response is defined by

0 otherwise,

where 0 ≤ α ≤ 1. R(x) = max σ 2f λ12 x; σ f . (5)
σf

123
248 Y. Fujita, Y. Hamamoto

Let σ f = s f −1 σ1 ( f = 1, 2, . . . , n) be a discrete sample of where N (R) means the number of pixels in the region R.
σ f , where σ1 is the minimum scale and s is a scale factor Using (8), Q i (λk ) is the average of estimates for label λk in
determining the sampling interval of σ f . Changing the scale neighbors of xi .
of the Gaussian enables examining lines fitting various scales However, in crack detection, if the average of estimates
of the Gaussian filter. in neighbor is used as non-ambiguity, cracks may be elim-
inated unexpectedly because of their liner structure. There-
2.3 Detection processing fore, we divide the neighboring region into four sub-regions
according to direction and use non-ambiguity estimated in
2.3.1 Probabilistic relaxation one sub-region, where four estimates of each pixel are calcu-
lated along four different angular directions 0, 1/4π, 1/2π ,
Probabilistic relaxation is mainly used as the optimization and 3/4π passing through the focal pixel. The maximum
technique for solving labeling problem [12,13]. By probabi- estimate value of four directions is used to update the prob-
listic relaxation, it is possible to label objects to be extracted ability. By this, crack detection works well. We update the
from noisy data. probability of crack Pc (λk ) by
For a focal pixel xi , the probability Pi (λc ) of crack is
Pi (λc )Q id (λc )
assigned to the logarithmic transformation of the output of Pi (λc ) = max (9)
the preprocessing, as initial probabilities. Here, we used the d Pi (λc )Q id (λc ) + Pi (λb )Q id (λb )
following logarithm transformation as, 1 
Q id (λk ) = P j (λl ), (10)
N (Ri )
d
log(R(xi ) + 1) x j ∈Ri
d
Pi (λc ) = , (6)
log(Rmax + 1)
where d(= 0, π/4, π/2, 3/4π ) means the direction and Rid
where R(xi ) means the output of the preprocessing at pixel is the sub-region of the direction d. Sub-regions and exam-
xi , Rmax is the maximum value, and λc means the label for ples of updating are shown in Fig. 4. The updating process
crack. On the other hand, the probability of background is for all pixels is repeated until the labeling converges or sta-
assigned to Pi (λb ) = 1 − Pi (λc ), where λb means the label bilizes. Noises in the image are expected to be removed by
for background. updating the process, whereas cracks remain. Figure 5 shows
Updating probabilities is done by considering the proba- an example the probabilistic relaxation process.
bilities of neighboring pixels. The updating rule proposed by
Peleg [13] is 2.3.2 Improved locally adaptive thresholding process
Pi (λk )Q i (λk )
Pi (λk ) =  . (7) An improved locally adaptive thresholding is used to close
l∈L Pi (λl )Q i (λl )
undesirable unconnected cracks and to extend cracks. First,
The non-ambiguity Q i (λk ) is regarded as the strength of sup- the neighbor pixels of pixels decided as cracks are added to
port that label λk at pixel xi gets from its all neighbor pixels the candidate area. Then it is determined whether the pixels
x j ∈ Ri , for example, it is defined simply as follows: in the candidate area are a piece of crack or not. After this,
 a local thresholding technique is used, in which the thresh-
1
Q i (λk ) = P j (λk ), (8) old value is decided in the local area. The threshold is given
N (Ri ) for each pixel in the candidate area, while the threshold is
x j ∈Ri

Fig. 4 Explanation of four sub-regions and updating in relaxation. a pixel and the central circle means the focal pixel. In (b) and (c), the
(a) Division of four sub-regions. (b) Up case of an updating probability. selected sub-region in updating for the focal pixel is enclosed by a dot
(c) Down case of an updated probability, where each circle indicates line

123
A robust automatic crack detection method from noisy concrete surfaces 249

Fig. 5 Probabilistic relaxation process of the proposed method. (a) Original image. (b) Result of pre-processing from (a) (t = 0). (c) Result of
relaxation from image (b) with t = 2. (d) Result of relaxation from image (b) with t = 10, where t denotes the iteration number of updating

Fig. 6 An example of the improved locally adaptive thresholding of image (b) with t = 10. (d) Result of improved locally adaptive threshold-
the proposed method. (a) Original image. (b) Result of relaxation from ing from image (b) with t = 67, where t denotes the iteration number.
image (a). (c) Result of improved locally adaptive thresholding from The Updating process is converged at t = 67

not given for pixels not included in the candidate area. So, crete surface images. We obtained correct focus on a subject
this approach is considered as an improved locally adaptive by using autofocus. All algorithms were implemented in Java
thresholding technique to prevent noises. In updating the can- and tested on a computer with Intel Core 2 2.4 GHz CPU and
didate area, neighbor pixels of the pixel labeled as crack are 2 GB RAM.
added to candidate area, while the pixels that were labeled as
non-crack ones are excluded from the candidate area. Updat-
ing process is performed until convergence. 3.1 Two preprocessing steps
Figure 6 shows an example of the improved locally adap-
tive thresholding. We used the receiver operating characteristic (ROC) anal-
ysis [14,15] to evaluate the performance of the two pre-
processing steps. In this analysis, the different approaches
to be considered are compared in terms of the sensi-
3 Experimental results tivity and the specificity. The sensitivity or true posi-
tive fraction (TPF) means the probability that cracks are
The proposed method was evaluated using 60 actual concrete detected correctly at pixel level. On the other hand, the
surface images including various blebs, brains, and shadings. specificity or false negative fraction (TNF) means the
These digitized images have 640×480 pixels, approximately probability that backgrounds are correctly identified at
0.1 mm pixel size, and 8-bit gray levels. Note that our tar- pixel level. The choice of the threshold value influences
get cracks are over 0.2 mm in width, and cracks in used both the sensitivity and the specificity. The ROC curve
images have over two pixels width. In our experiment, we is the graphic representation of this reciprocal relation-
used SONY Cyber-shot DSC-F828 camera to obtain con- ship between the sensitivity and the specificity, calcu-

123
250 Y. Fujita, Y. Hamamoto

Table 1 Comparison between the proposed method and other three approaches in terms of AUC
Preprocessing Proposed Only subtraction Only line emphasis No
method preprocessing preprocessing preprocessing
Average and 95% 0.98 0.92 0.97 0.89
Confidence interval 0.98, 0.99 0.90, 0.93 0.97, 0.98 0.87, 0.91
t test 0.07±0.02** 0.01±0.01* 0.09±0.03**
Paired samples t test 0.07±0.02** 0.01±0.00** 0.09±0.03**
* p < 0.05, ** p < 0.001

lated for all possible threshold values. The vertical axis


of the graph shows the sensitivity or TPF. The hori-
zontal axis represents the false positive fraction (FPF =
1 − specificity). Each operating point on the ROC curve
represents the combination of the sensitivity and the spec-
ificity at a given threshold value. When the ROC curve is
close to the upper left corner, the preprocessing works well.
We used the area under the curve (AUC) to estimate the
performance of the preprocessing in crack detection. The
major advantage of the use of the AUC is that this is done
independently of decision criteria, thus eliminating influence
of the threshold value on the sensitivity and the specificity
values.
Fig. 7 Comparison of the ROC curves among four approaches. (a) Pro-
In evaluation, we compared the proposed method with the posed method including two preprocessing (AUC = 0.9903). (b) Only
approach with no preprocessing, the approach with the sub- line emphasis preprocessing (AUC = 0.9509). (c) Only subtraction pre-
traction preprocessing, and the approach with the line empha- processing (AUC = 0.9145). (d) No preprocessing (AUC = 0.8662)
sis preprocessing. Here, we determined the size of the median
filter on the subtraction preprocessing and the parameters
where Ntd is the number of detected crack pixels
of the line emphasis processing according to preliminary
and Nc is the total number of crack pixels. The false
experiment. The size of the median filter was 21 × 21 pix-
negative fraction Fn are calculated as follows:
els. For the multi-scale line filter, we carried out preliminary
experiment for parameter optimization, fixing the parame-
Nfd
ter α =√0.25. From
√ the preliminary
√ experiment,
√ parameters Fn = , (12)
σ f = 2, 2, 2 2, 4(σ1 = 2, s = 2, andn = 4) were Nb
determined.
where N f d is the number of the detected pixels not
The average and 95% confidence interval of the AUC of
included in the crack regions and Nb is the total num-
each approach for 60 noisy concrete images were calculated
ber of background pixels on the truth image. Draw
as follows:
Step 1 Trace the cracks manually on the original image to the ROC curve from 256 relationships between the
obtain the gold standard by a human operator. It is sensitivity and the specificity, and calculate the AUC.
used for comparison. Step 5 Repeat Steps from 1 to 4 for each of 60 original
Step 2 Apply a pre-processing to be evaluated to an original images.
concrete image. Step 6 Compute the average and 95% confidence interval
Step 3 Threshold to separate cracks from background. of the AUC.
Threshold values are changed from 0 to 255, to obtain Table 1 shows the average and 95% confidence interval of
256 binary images as results of crack detection. the AUC of each approach. The significance of the difference
Step 4 Compare 256 results detected automatically with the between the proposed method and only line emphasis prepro-
traced image, respectively. The true positive fraction cessing is provided by the t test ( p < 0.05), and the signifi-
T p is defined by cance of the difference between the proposed method and the
other two approaches is provided by the t test ( p < 0.001).
Ntd Figure 7 shows the comparison of ROC curves among four
Tp = , (11) approaches. It is clear that the ROC curve of the proposed
Nc

123
A robust automatic crack detection method from noisy concrete surfaces 251

Fig. 8 Comparison of preprocessings. (a) Original image. (b) Pro- preprocessing (AUC = 0.985). (e) No preprocessing (AUC = 0.904).
posed method including two preprocessing steps (AUC = 0.994). (f) Wavelet transform (level = 2, AUC = 0.964)
(c) Subtracted preprocessing (AUC = 0.954). (d) Line emphasis

method is located higher than the others. Figure 8 shows the Table 2 The performance of the wavelet transform method with the
results of preprocessings. These results show that combina- hysteresis thresholding, compared with that of the proposed prepro-
cessing with single thresholding
tion of the subtraction preprocessing and the line emphasis
preprocessing is effective to detect cracks from noisy con- Level Wavelet transform Proposed
crete surface images. with hysteresis preprocessing with
Next, we compared the proposed method with the wavelet thresholding single thresholding
transform method described by Hutchinson et al. [2], using
ROC analysis. In the wavelet transform method, a hysteresis 1 0.87 0.98
thresholding is used, considering high and low thresholds for 0.84, 0.89 0.98, 0.99
the hysteresis, T1 and T2 (T2 = 0.7 × T1 ) as described in [2]. 2 0.93
Four levels of wavelet decomposition are considered. The 0.92, 0.94
performance of the wavelet transform method with the hys- 3 0.90
teresis thresholding is shown Table 2. As shown in Table 2, 0.89, 0.92
the proposed preprocessing is superior to the wavelet trans- 4 0.82
form method for crack detection. Figure 8f shows the result 0.80, 0.85
of the wavelet transform. Upper: average, lower: 95% confidence interval

3.2 Automatic detection


steps. In the adaptive thresholding of the proposed method,
For comparison, we evaluated the proposed method and the we used Otsu’s method [16] to determine the threshold for
global thresholding, which is considered as a comparative each local area. The size of the local area was 21 × 21.
method, using images obtained by the two preprocessing Both approaches are evaluated quantitatively and visually.

123
252 Y. Fujita, Y. Hamamoto

Fig. 9 Effect of comparison


of relaxation and adaptive
thresholding. (a), (b) Original
images. (c), (d) Results by the
proposed method. (e), (f)
Results by Otsu’s thresholding
method including the proposed
preprocessing

Fig. 10 Effect of noise elimination. (a) Original image. (b) Result of crack detection by the proposed method. (c) Result of noise elimination from
image (b) with T = 200

In quantitative assessment, two approaches are assessed in Step 3 Determine a threshold for the global thresholding
terms of both sensitivity and specificity. Evaluation was per- approach, which is considered as the comparative
formed as follows: method, against the image obtained by the prepro-
cessing, so that sensitivity (or specificity) becomes
Step 1 Detect cracks by the proposed method. the same as that of the proposed method.
Step 2 Evaluate the result of the proposed method in terms Step 4 Evaluate the result of the global method in terms of
of sensitivity and specificity. specificity (or sensitivity).

123
A robust automatic crack detection method from noisy concrete surfaces 253

Table 3 Sensitivity and


specificity for varying area size Area size Without noise 100 200 300 400
of noise elimination T elimination

Sensitivity 0.80 0.79 0.78 0.76 0.75


0.77, 0.84 0.75, 0.79 0.74, 0.82 0.72, 0.80 0.71, 0.79
Specificity 0.992 0.992 0.993 0.993 0.993
Upper: average, lower: 95% 0.991, 0.993 0.992, 0.993 0.992, 0.994 0.993, 0.994 0.993, 0.994
confidence interval

Step 5 Compare the specificity (or the sensitivity) between


the proposed method and the global thresholding
method.

The average and 95% confidence interval of the sensitivity of


the proposed method for 60 images are 0.80 and (0.77, 0.83),
respectively, and those of the specificity are 0.992 and (0.990,
0.994), respectively. When the global thresholding method
performs the same as the sensitivity of the proposed method,
those of the specificity of the global method are 0.991 and
(0.990, 0.992), respectively. On the other hand, when it per-
forms the same as the specificity of the proposed method,
those of the sensitivity of the comparison method are 0.79
and (0.76, 0.83), respectively. The significance of the differ-
ence between the proposed method and the global method is
provided by the paired t test ( p < 0.001). From the results,
the performance of the proposed method is found to be supe-
rior to the global thresholding in terms of both sensitivity and
specificity.
In visual assessment, we compared the proposed method
with Otsu’s method [16] to determine a threshold based
on integration of the gray-scale histogram. Figure 9 shows
results of the proposed method and the Otsu’s method using
the image obtained by the preprocessing. Application of
Otsu’s method results in the loss of the connectivity of
detected cracks and noise generation. These results show the
proposed method can detect cracks with higher accuracy than
Otsu’s method.
Next, we show the results of noise elimination for vari-
ous area sizes. By calculating pixels in the area extracted by
labeling, areas measuring less than T pixels were removed
as noise. The result is shown in Fig. 10. It is possible to elim-
inate noise with a proper area size. For varying the area size
T , the sensitivity and the specificity are shown in Table 3.
When the size parameter T is too large, noises are eliminated
sufficiently, while cracks are also removed. Thus, it is neces-
sary to optimize the size parameter for each image. This will
be the subject of further work.
Finally, cracks detected by the proposed method may be
classified according to the width of cracks using the result of
the multi-scale line emphasis preprocessing. Figure 11 shows
the coloring of detected cracks according to crack width by Fig. 11 Coloring according to crack width using result of the multi-
the proposed method. In practical use, this technique is use- scale line emphasis preprocessing

123
254 Y. Fujita, Y. Hamamoto

ful because the width of cracks is important for maintenance 2. Hutchinson, T.C., Chen, Z.: Improved image analysis for evaluat-
of concrete structures. ing concrete damage. J. Comput. Civ. Eng. 20(3), 210–216 (2006)
3. Yamaguchi, T., Nakamura, S., Saegusa, R., Hashimoto, S.: Image-
based crack detection for real concrete surfaces. Trans. Electr. Elec-
tron. Eng. 3, 128–135 (2008)
4 Conclusion 4. Kawamura, K., Miyamoto, A., Nakamura, H., Sato, R.: Proposal
of a crack pattern extraction method from digital images using an
interactive genetic algorithm. Proc. Japan Soc. Civ. Eng. 742, 115–
In this paper, we proposed a new method for robust auto- 131 (2003)
matic crack detection from noisy concrete surface images. 5. Abdel-Qader, I., Abudayyeh, O., Kelly, M.E.: Analysis of edge
Two preprocessing steps are used to reduce influence of vary- detection techniques for crack identification in bridges. J. Comput.
ing blebs, stains, and shadings. In the proposed method, the Civ. Eng. 17(3), 255–263 (2003)
6. Kabir, S., Rivard, P., He, D., Thivierge, P.: Damage assessment for
corrective image was generated from the original image by
concrete structure using image processing techniques on acoustic
the median filter, and we used the multi-scale line filter to borehole imagery. Constr. Build. Mater. 23(10), 3166–3174 (2009)
overcome the variation in the width of cracks. Moreover, for 7. Lee, J.H., Lee, J.M., Kim, H.J., Moon, Y.S.: Machine vision sys-
automatic detection, we applied probabilistic relaxation to tem for automatic inspection of bridges. Congress Image Signal
Process. 3, 363–366 (2008)
prevent noise generation. Finally, the improved locally adap-
8. Weszka, J.S.: SURVEY: a survey of threshold selection tech-
tive thresholding was performed to detect cracks exactly. An niques. Comput. Graph. Image Process. 7, 259 (1978)
advantage of relaxation over conventional methods is that it is 9. Lee, S.U., Chung, S.Y., Park, R.H.: A comparative study of several
does not need optimization of the threshold for binarization. global thresholding techniques for segmentation. Comput. Graph.
Image Process. 52, 171–190 (1990)
In the experiments, two preprocessing steps are evaluated 10. Sezgin, M., Sankur, B.: Survey over image thresholding tech-
by ROC analysis. Combination of two preprocessing steps niques and quantitative performance evaluation. J. Electr. Imag-
improves the performance of crack detection from noisy con- ing 13(1), 146–168 (2004)
crete surface images. Moreover, combination of the proba- 11. Sato, Y., Nakajima, S., Shiraga, N., Atsumi, H., Yoshida, S.,
Koller, T., Gerig, G., Kikins, R.: Three-dimensional multi-scale
bilistic relaxation and adaptive thresholding obtains higher line filter for segmentation and visualization of curvatures in med-
sensitivity or specificity than the global thresholding. Using ical images. Med. Image Anal. 2(2), 143–168 (1999)
the result from the multi-scale line emphasis preprocessing, 12. Rosenfeld, A., Hummel, R.A., Zucker, S.W.: Scene labeling using
detected cracks may be classified according to crack width. relaxation operation. IEEE Trans. Syst. Man Cybern. vol. SMC-6,
no. 6, pp. 420–433 (1976)
The proposed method presents high accuracy and the result 13. Peleg, S.: A new probabilistic relaxation scheme. IEEE Trans. Pat-
should be easy to use for further applications. tern Anal. Mach. Intell. vol. PAMI-2, no. 4, pp. 362–369 (1980)
14. Erkel, A., Pattynama, P.: Receiver operating characteristic (ROC)
analysis: basic principles and applications in radiology. Eur. J.
Radiol. 27(2), 88–94 (1998)
References 15. Fawcett, T.: An introduction to ROC analysis. Pattern Recognit.
Lett. 27(8), 861–874 (2006)
1. Ito, A., Aoki, Y., Hashimoto, S.: Accurate extraction and mea- 16. Otsu, N.: Threshold selection method from gray-level histograms.
surement of fine cracks from concrete block surface image. Proc. IEEE Trans. Syst. Man Cybern. 9(1), 62–66 (1979)
IECON 3, 2202–2207 (2002)

123

You might also like