Sie sind auf Seite 1von 14

Original Article

A detection and classification approach


for underwater dam cracks

Structural Health Monitoring


2016, Vol. 15(5) 541554
The Author(s) 2016
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/1475921716651039
shm.sagepub.com

Pengfei Shi1,2, Xinnan Fan1, Jianjun Ni1 and Gengren Wang1

Abstract
Underwater dam crack detection and classification based on visible images is a challenging task. The underwater environment is very complex with uneven illumination and serious noise problems, which often leads to the distortion of detection. In addition, there are few methods suitable for underwater dam crack classification. To solve these problems, a
novel underwater dam crack detection and classification approach is proposed. Firstly, a dodging algorithm is used to
eliminate the uneven illumination in the underwater visible images. Subsequently, a crack detection approach is proposed, where the local characteristics of image blocks and the global characteristics of connected domains are both used
based on the analysis of the statistical properties of dam crack images. Finally, an improved evidence theory-based crack
classification algorithm is proposed after the crack detection. Experimental results show that the proposed approach is
able to detect underwater dam cracks and classify them accurately and effectively in complex underwater environments.
Keywords
Underwater dam, crack detection, crack classification, image processing, evidence theory

Introduction
Dam safety detection is important work to protect dam
safety and prolong a dams service life. Many factors
threaten the safety of the dam, such as cracks, abrasion,
cavitation, and erosion.1,2
Underwater dam cracks are not only found on the
surface of a dam, but also extend into the interior,
which is a mutation of the process of internal damage
accumulation and can be used to express the degree of
internal damage.3 Hence, underwater dam cracks are
crucial in predicting danger in dams and their detection
has become a research hotspot in the dam safety field.
There are some traditional methods used for dam
crack detection, such as electrical prospecting, elastic
wave testing, tomography, ground penetration, and
vibration-based methodology.47 These traditional
methods have some limitations. For example, most of
them are harmful to the dam in a certain way, or their
cost is high. Some methods are not convenient or sufficiently reliable.8 Recently, underwater dam crack detection based on visible image processing technology has
become one of the most important methods.9 The visible image processing technology is convenient, intuitive, efficient, economical and nondestructive, which
can meet the requirements of the dam crack detection.10 However, there are some problems that need to

be solved in underwater dam crack detection and classification based on visible images. On the one hand,
underwater images are full of serious non-uniform
brightness and fuzzy detail caused by the absorption,
scattering and convolution effects. There is always
uneven distribution of gray and poor contrast ratio in
the underwater visible images,15 especially when the
underwater lighting conditions are bad and an auxiliary
light source is used in a CCD camera.1214 In addition,
in the complex underwater environment, a lot of factors can affect crack detection by using visual images,
including exfoliation, honeycomb, and scratches, as
well as oil paint, cables, water plants, sludge, aquatic
animals, etc.11,16,17 On the other hand, historically
images obtained by underwater equipment always lack
calibration. The location and the numbers of the underwater dam cracks are random and unstructured. As a
result, the crack images are highly uncertain, unstable
and diverse, which makes it difficult to obtain accurate

College of IOT Engineering, Hohai University, China


College of Computer and Information, Hohai University, China

Corresponding author:
Jianjun Ni, College of IOT Engineering, Hohai University, Jingling north
road, Changzhou, 213022, China.
Email: jianjun_ni@hhu.edu.cn

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

542

Structural Health Monitoring 15(5)

information. So, it is difficult to classify the cracks efficiently by comparing sample images with test images,
which both convey fuzzy information.11
Aimed at those problems mentioned above, various
methods based on visible images have been proposed
to deal with underwater dam crack detection. Chen et
al.15 presented a multi-structured and multi-scale elements-based adaptive underwater dam crack edge
detection algorithm. Kabir et al.13 evaluated various
edge-detection algorithms and pointed out that the
statistical-based approach was the most efficient technique for damage assessment from acoustic imagery.
Bernstone and Heyden18 proposed a digital image analysis technique for crack monitoring, using a standard
web-camera to acquire continuous data sets in concrete
dams. Xu and Zhang19 presented an integrated model
by using digital image processing in the numerical representation of concrete structure defects. However,
most of the traditional methods cannot deal with the
crack detection efficiently when the background is
changing or the cracks are tiny. In addition, most of
these methods focus on de-noising and edge detection,
but few of them consider classification of underwater
dam cracks.
To deal with the problems of underwater dam crack
detection and classification based on visible images, a
novel framework is considered in this paper. For detection, local and global features are combined. Firstly,
the image is divided into different blocks, which are
classified with the local features, to remove crackless
blocks. Then, the global features of the crack-connected
domain, different from other suspected defects, are
used based on statistical analysis. For classification, we
propose a new strategy based on fuzzy evidence updating, in which basic belief assignment (BBA) is modeled
according to the fuzzy plausibility measure of a cracks
characteristics. The uncertainty of classification is
decreased by updating the evidence. Finally, some
experiments are carried out to prove the efficiency of
the proposed approach.
The main contributions of this paper are summarized
as follows. (1) A detection and classification approach
based on visible images for underwater dam crack is presented, which is an integration of several methods. (2) In
the proposed approach, local and global features are
both used, with block clustering and statistical analysis.
(3) The crack information obtained from the images,
which contain lots of uncertainty, is mapped to BBA;
then the uncertainty of classification can be decreased by
updating the evidence. (4) A concept of comprehensive
reliability is proposed to reduce the impact of artificial
factors, which allows easy application of the proposed
approach in real-world scenarios.
This paper is organized as follows: the section The
proposed approach presents the proposed underwater

dam crack detection and classification approach based


on visible images. In Experiments and analysis some
experiments are carried out and the results are analyzed. Our final section presents our Conclusions.

The proposed approach


Underwater dam cracks are always located by sonar in
a large scope, followed by detection and classification
by an underwater optical camera in a closed scope. In
this paper, we propose a new detection and classification approach based on optical images. Two main tasks
are considered in this study: underwater dam crack
detection (connected domains in the image are detected
as cracks or not) and underwater dam crack classification (tiny cracks, general cracks and through cracks).
To solve the two tasks, some prior knowledge about
cracks can be used: (1) the image regions are darker
than their surroundings; (2) the connected domain of
the crack region is slender compared with other regions;
(3) when the information entropy becomes larger, the
possibility of existing cracks is higher. In addition,
before the crack detection and classification, the images
should be pre-processed, primarily to eliminate the
effects of light. In this paper, we use a novel dodging
algorithm, based on an improved affine shadow formation model, which can solve the uneven illumination
problem in underwater images, and retain texture
details effectively. The details of this dodging algorithm
can be seen in our previous work.20 After the pre-processing, the image is divided into image blocks. Local
and global features are both used to realize the underwater dam crack detection. Then we propose a crack
classification method based on the fusion of fuzzy evidence. The workflow of the proposed approach is
shown in Figure 1, which will be introduced in detail as
follows.
Remark. In this paper, the crack position on the dam
has been marked and the images are labeled to allow
inspectors to relocate defects on the dam. Since this
paper focuses on the optical images of crack detection
and classification, previously developed methods are
used in this study to obtain, label and tag the images of
the cracks.21,22

Underwater dam crack detection


In recent years, there have been many crack detection
algorithms based on images of pavement or concrete
structures surfaces: the neural network method,23 the
wavelet radon transform based method,24 the data
association algorithm,25 and the self-organizing map
based method.26 These methods can be used to detect
cracks in pavement or concrete structure effectively.
However, they cannot be directly used in the detection

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

543

Test image
database

Image pre-processing

Dodging processing

3-D feature space

2-D feature space

K-means clustering

Detection of dam cracks

Threshold of the
circularity

Statistical
analysis

Global feature threshold


segmentation

BBAs of the circularity

Sample image
database

Classification of dam cracks

Threshold of the
gray intensity ratio

BBAs of the gray


intensity ratio

Evidence updating

Statistical
analysis

Sample image
database

Decreasing the uncertainty of


the classification

Figure 1. The work flow of the proposed approach.

of underwater dam cracks. For most of them are based


on the image pixels or connected domain; the processing of the pixels cannot reflect the relationship between
pixels and the cracks well. The processing of the connected domain cannot discern the cracks and the interferences which have the same features, especially when
the background is changing and the cracks are tiny.
Hence, most of the general algorithms cannot be used
directly for underwater dam crack detection. To deal
with the above problems, a novel dam crack detection
algorithm is proposed which divides the image into
blocks to describe the local features of each pixel. In
order to reflect the relationship between each block
and the overall image, the global features of the connected domain in the image are combined with the
local features. There are two main steps in the proposed crack detection algorithm. The first step is to
remove the image blocks without crack pixels. The second step is to remove the non-crack information based
on the features of the image connected domains.
Clustering analysis of the image blocks. For the crack
images, the image blocks have three simple but effective
local statistical parameters, namely the mean value, the
standard deviation and the information entropy of gray
level values. The crack (crack-like) and background
can be classified by clustering techniques in both the
two-dimensional (2-D) and three-dimensional (3-D)
feature space constructed by the three statistical parameters above.

The original images are divided into nonoverlapping


535 pixel blocks, which are considered for the local
feature detection (see Figure 2(a)). Three characteristic
values are computed for each block, which construct
three matrices for the image: (1) the mean value matrix
(Mm); (2) the standard deviation matrix (STDm); (3)
the information entropy matrix (IFE). A clustering
analysis method is used to distinguish crack (crack-like)
from background with the combination of the 2-D and
3-D space classifiers.
The 2-D feature space is constructed by Mm and
STDm. Each point identifies one image block. Then
the feature set in the 2-D space is defined as
F1 = f(U1 , y1 ) . . . (Un , yn ) : Ui 2 R2 ; yj 2 fc1 , c2 gg

where n is the number of block points for the pattern


vector U, which is constructed by Mm and STDm; yi is
the corresponding class of the ith block. In this study,
there are two classes for the blocks: namely, c1 and c2 ,
which denotes the class of the block without cracks,
and with cracks (crack-like), respectively.
The 3-D feature space is constructed by Mm, STDm
and IFE, which compose a pattern vector V. Then the
feature set in the 3-D space is defined as
F2 = f(V1 , y1 ) . . . (Vn , yn ) : Vi 2 R3 ; yj 2 fc1 , c2 gg

where n is the number of block points for the pattern


vector V.

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

544

Structural Health Monitoring 15(5)


18
16

Standard deviation

14
12
10
8
6
4
2
136

138

140

142

144

146

Mean

(a)

(b)

Information entropy

x 10 (bit / symbol)
1.1859
1.1859
1.1859
15
an

St

146
144

rd

da

10

tio

via

de

142
140

138

an

Me

136

(c)

(d)

Figure 2. The clustering analysis combined with the 2-D and 3-D feature space: (a) nonoverlapping 535 pixel blocks; (b) the two
classes in 2-D feature space; (c) the two classes in 3-D feature space; and (d) the target blocks marked with red square after
clustering.

In this study, the K-means classification approach is


used to do the clustering analysis.22 The class boundary
of the 2-D feature space is shown in Figure 2(b) and
the clustering result in the 3-D feature space is shown
in Figure 2(c). The classes labeled with red circles both
in the 2-D and 3-D feature space are the target class c2 .
The final results of the target blocks are shown in
Figure 2(d), which are marked with red squares.
Statistical analysis of the connected domain. After removing
those blocks that are confirmed to have non-crack
information, the crack detection is done by the global
features of the connected domain on the image blocks
left, to further remove the non-crack information. The
features of the image block describe the local characteristics in the image. It reflects the relationship among
the blocks, but it cannot show the relationship between
the block and the connected domain in the image. So
the global features of the connected domain are used to

further remove the non-crack information. To explain


the crack domain against the background, a negated
binary image is shown in Figure 3(a).
There are several shape characteristics that can
describe the crack, such as the width-length ratio (WL),
the circularity (C), the ratio of the long axis to the short
one of the external ellipse (RELS), and the gray intensity ratio (D).27 And the parameters C and D can be
calculated by
C = 4p3A=P2

D = (f2  f1 )=f2

where A is the total number of pixels of the target; P is


the number of pixels of the target contour; f1 is the
average gray value of the target and f2 is the average
gray value of the targets minimum enclosing rectangle.
The crack statistical results of 20 images for each type
gained from a hydropower project in China are shown

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

545

0. 9
0. 8

Characteristic Values

0. 7
0. 6
0. 5

Widthlength ratio
Circularity
Gray intensity ratio
RELS

0. 4
0. 3
0. 2
0. 1
0

10
Target

(a)

15

20

(b)

Figure 3. Statistical analysis of the connected domain: (a) connected domains in a binary image and (b) characteristic statistical
properties of an underwater dam crack.

in Figure 3(b). Two main pieces of prior knowledge


about cracks can be introduced: (i) crack regions are
darker than their surroundings; (ii) the crack region is
more slender than other regions. The statistical results
of typical characteristic quantity in Figure 3(b) show
that the circularity of the cracks are all below 0.05,
which can reflect the characteristic of slenderness of
cracks well. The gray intensity ratio of the cracks above
0.6 accounts for 0.9, which denotes that the crack
regions are darker than their surroundings. So the two
characteristics are used for detecting cracks by setting
some given threshold values and classifying cracks by
evidence fusion.

Underwater dam crack classification


Crack classification based on images always faces a
problem: the test image feature and the sample image
feature are both full of fuzziness and uncertainty. For
the test image, there are many factors that will lead to
uncertainty in feature detection, such as cracks with
fuzzy randomness and diversity, and the restrictions of
the crack detection algorithm. Furthermore, the feature
of the sample image without calibration and proofreading cannot accurately reflect the relationship between
the image crack and the actual crack. If these imperfections are ignored, and unrealistic assumptions are
made, it will lead to untrustworthy inferences. This
paper proposes a new method of target classification
based on fuzzy evidence combination. In the proposed
method, fuzzy evidence is used to reflect fuzzy information in the images: uncertainty is decreased by combining the evidence. Based on statistical analysis of the
underwater crack images above, the circularity and the
gray intensity ratio are selected to establish the BBA

function (BBAs), for the modeling of fuzzy evidence in


this study. The mapping from distance function to
BBAs is a nonlinear mapping, and their general trends
are opposite: for the circularity, while the circularity
distance becomes smaller, the belief assigned to the evidence is greater; for the gray intensity ratio, while the
difference in gray intensity ratio distance becomes
smaller, the belief assigned to the evidence is greater.
The pattern set of cracks type is Y = fu1 , u2 , u3 g,
where u1 represents tiny cracks, u2 represents general
cracks, and u3 represents through cracks. In general,
these three types of crack correspond to the severity of
the cracks. The through crack means a crack that
breaks throughout the concrete dam, which is the most
severe type. And of course tiny cracks are the most difficult to detect. The corresponding belief of evidence
(BoE) obtained from the circularity (C) is
eO = fY, zO , mO (:)g

where zO is the core which contains circularity subset of


Y.
In this paper, the BBAs are calculated by the likelihood measure t(Ck (x), Ckq (x)), where Ck (x) refers to the
circularity of cracks and the likelihood measure
t(Ck (x), Ckq (x)) refers to the matching degree of Ck (x)
belonging to the linguistic terms Ck1 (x), Ck2 (x), and
Ck3 (x). Based on the circularity, we can distinguish three
classes of crack, but this discrimination is fuzzy. The
membership functions are chosen by experiment statistics of a characteristic sample set. The characteristic
crack data of the three classes from the sample are used
for calculating the mean value and the standard deviation value, after which the Gauss membership function
is constructed. The results are shown in Figure 4(a).

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

546

Structural Health Monitoring 15(5)

q = 1, 2, 3; k = 1, . . . , n1

where n1 is the number of the evidence. The matching


degree is the maximum value of the intersect point ordinate between the test membership degree curve and the
sample membership degree curve. While the intersecting
degree is greater, the matching degree is larger, which in
turn provides more evidence. The evidence reflects the
extent of support which is directly used as a BBA function after normalization. Thus, the BBA values can be
calculated as follows
t(Ck , Ckq )
mYO (C)k = n
P
t(Ck , Ckq )

6
Likelihood measure value

t(Ck , Ckq ) = supx minfCk (x), Ckq (x)g,

x=0.0218
y=0.88

0.8

0.6

0.4
C1(x)

0.2

C2(x)
C3(x)
C(x)

0
0

0.01

0.02
0.03
Amplitude

0.04

0.05

(a)
1

obviously,

mYO (C)k = 1.

k=1

Similarly, BBAs of the other connected area can be


calculated. However, the gray intensity ratio (D) cannot
discern u1 and u3 . So the Gauss membership functions
of the tiny crack and the through crack are similar.
Here, set T1 = fu1 , u3 g, T2 = fu2 g. The results are shown
in Figure 4(b). Then the corresponding BoE for the
gray intensity ratio is
eT = fY, zT , mT (:)g

x=0.658
y=0.7471

0.6

0.4

D1(x)

0.2

D(x)

0
0.5

0.6

0.7
0.8
Amplitude

0.9

(b)

Figure 4. Statistical properties pattern of a sample


characteristics set and test characteristics: (a) circularity (C) and
(b) gray intensity ratio (D).

JLR, CUR is used in this paper. The CUR can be calculated by30

k=1

The likelihood measure t(Dk , Dkq ) can be calculated


in the same way as t(Ck , Ckq ), where
t(Dk , Dkq ) = supx minfDk (x), Dkq (x)g,
q = 1, 2; k = 1, . . . , n1

0.8

D (x)

where zT is the core which contains gray intensity ratio


subset of Y. In the same way, BBAs for the gray intensity ratio are obtained as follows
t(Dk , Dkq )
mYT (D)k = P
n
t(Dk , Dkq )

Likelihood measure value

k=1

n
P

mYO (C)k + 1 = akmYO (C)k,


X
b(D)kmYT (CjD)k
+
DYT

10

and the result is shown in Figure 4(b).


The mass function established by the matching
degree of fuzzy feature distance is the BBA of uncertainty. To reduce the uncertainty of classification and
improve the robustness of decision-making, evidence
from both the circularity and the gray intensity ratio
should be combined. There are several classic rules to
combine evidence, such as the Dempster Combination
Rule (DCR), the Jeffery-like Rule (JLR)28 and the
Conditional Update Rule (CUR). Because it is difficult
to fuse the conflicting BoEs by DCR29 and unreasonable to obtain results only related to new evidence from

11

where

eO k + 1[eO k / eT k,

8k  0,

ak +

P
DYT

b(D)k = 1, 8k  0, and b(D) = 0,


mYT (CjD)k can be calculated by31
P
mYT (CjD) =

E:EC

m(E)
P

8D 62 =YT .

Pl(C) 
m(X )
X :X 2l(C)
X
m(EjD)


12

E:EC

where l(C) = fX 2 Y : X = F [ E, [ 6 F  D, [ 6 E 
C  Dg.

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

547

In the general fusion process introduced above, the


parameter values of a and b are set artificially. There
are some limitations of this artificial assignment
method. The main reason is that it is difficult to find a
uniform measurement between two pieces of evidence,
which is used to measure the value of the credibility for
the heterogeneous information. Furthermore, the artificial assignment method lacks rigorous reasoning. The
parameter values of a and b can be obtained from statistical values which are obtained from historical dam
crack images. To improve the adaptability of the
method, a concept of comprehensive reliability is proposed in this paper where the credibility of evidence is
not only related to its own credibility in the evidence
fusion process, but also to the support of other evidence. The credibility of the evidence is denoted by
Crdk , which is calculated by
1=dk
Crdk = N
P  
1= dj

13

j=1

where dk refers to the sum of distance between one evidence and others, it can be obtained by
dk =

N
X

( arg max (Zk (x))  arg max (Zj (x))),

j=1

14

j 6 k
Zk (x), Zj (x) 2 2Y

In this paper, the parameter values of ak and bk


are defined as
fak, bkg = fCrdk , 1  Crdk g

15

If the distance between one piece of evidence and


another
/querydistance between one evidence and other evidence cant really have distance between one evidences: ? difference in meaning between one type of
evidence and another. (Again, evidence is a mass noun,
so you dont usually have evidences)
is smaller, then the mutual support among them is
higher. In the fusion process of heterogeneous features,
the evidence which has been updated is combined in the
following order
eY k + 1 = (eO k + 1 / eO k) / (eT k + 1 / eT k)

16

The workflow of the crack classification in the proposed approach is summarized as follows.
Step 1. The values of the crack features C and D are calculated by (3) and (4) for the crack sample images of
different types of cracks.

Step 2. Gauss membership functions of C and D are


established by the values of each type of cracks.
Step 3. The values of the crack features C and D are calculated for the test image.
Step 4. The matching degrees are calculated for C and
D by (6) and (10) respectively.
Step 5. The BBA values are calculated for C and D by
(7) and (9) respectively. In the same way, the BBA values are calculated for the test images of the same target
crack with different perspectives.
Step 6. The BBA values of C and D from the test
images are combined by (11) to (16).
Step 7. The test image with the maximum BBA values
is classified to the corresponding type of crack.

Experiments and analysis


The proposed algorithm in this paper is used to analyze
four typical cracks of underwater dam surface images
captured from a dam hydropower project: the first is
that the dam surface image contains only one normal
crack with the background under even illumination
(see (a)i in Figure 5); the second is that the dam surface
image contains only one normal crack with the background under uneven illumination (see (a)ii in Figure
5); the third is that the dam surface image contains two
huge cracks with a complex background under uneven
illumination (see (a)iii in Figure 5); the fourth is that
the dam surface image contains only one tiny crack
with a complex background under even illumination
(see (a)iv in Figure 5). In this study, the images are
obtained by a UWC-325/P camera, with sensitivity
0.01Lux@F1.4 and resolution 480TV lines standard.
And the distance from the camera to the defects varies
between 500 mm and 1200 mm. For performance evaluation, we collect a set of 60 dam crack images with
various kinds of cracks (each type crack with 20
images), which are used as the sample database and the
test database. All these images have a size of 768 3
494. Some experiments are carried out that are coded
in Matlab 2011 and on a PC with 2.6 GHz CPU and
4G RAM. The details of the experimental process are
introduced as follows.

Underwater dam crack detection experiments


First of all, the four original images are processed to
eliminate the negative impact of light by using the proposed method of texture dodging, and the results are
shown in Figure 5(b). The results show that the uneven
illumination of the images has been eliminated, and the
crack textures in the images are well preserved.
After that, the non-crack information of the images
that have been dodged is removed partly according to
the local features in the image blocks. The results are

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

548

Structural Health Monitoring 15(5)

Figure 5. Underwater dam crack detection experiments with typical cracks. (a) Four typical cracks of underwater dam surface
original images. (b) The uneven illumination of images has been eliminated after texture dodging maintaining. (c) Clustering analysis
of blocks local features both in 2-D and 3-D. (d) Statistical analysis of global features of the connected domain.

shown in Figure 5(c). And then, the non-crack information is further removed, based on the global features
of the connected domain. The results are shown in
Figure 5(d).
To test the performance of the proposed approach,
it is compared with the wasp colony algorithm,32 the
crack tree method33 and the crack IT method,22 which
are common methods in crack detection. The wasp colony algorithm (WASP) is an optimization algorithm
based on intelligent behaviour in a honey wasp swarm.
The crack tree (Crack-T) method builds a crack probability map using tensor voting. And the crack IT

(Crack-I) method uses clustering analysis based on 2-D


feature space. The comparison of the four methods is
shown in Figure 6. Some indices34 such as mr (missing
rate), fd (false detecting rate), pr (precision), re (recall),
pc (performance criterion) are used to evaluate the proposed method and the other three methods, which are
defined as
fd =

N1
N4

17

mr =

N2
N4

18

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

549

Figure 6. Image detection results comparing the proposed method and other classical methods: (a) Original image, (b) Wasp
algorithm, (c) Crack tree, (d) Crack IT, and (e) The proposed method.

pr =

N3
N4

19

re =

N3
N5

20

pc =

23pr3re
pr + re

21

where N1 refers to the number of connected domains


wrongly detected as cracks, N2 refers to the number of
crack regions undetected as cracks, N3 refers to the

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

550

Structural Health Monitoring 15(5)

Table 1. The evaluation of the image crack detection results.


Method name

WASP
Crack-T
Crack-I
Proposed method

The evaluation index


mr

fd

pr

re

pc

tm

96.1%
49.0%
37.3%
7.8%

19.6%
58.8%
11.8%
9.8%

16.7%
39.4%
84.2%
88.7%

3.9%
51.0%
62.7%
92.2%

6.3%
44.5%
71.9%
90.4%

89.8(s)
946.2(s)
1.3(s)
1.9(s)

number of connected domains correctly detected for


cracks, N4 refers to the total number of crack regions
of real ground, and N5 refers to the total number of
crack regions detected. In addition, to compare the
time cost of the four methods in performing the processes, an index of average time (tm) in crack detection
is used. The results of the evaluation for the four methods are listed in Table 1.
The results in Table 1 show that the wasp colony
algorithm is unable to deal with crack detection when
the background is complex and changing. The reason is
that the interference is always detected as a target by
the wasp colony algorithm (see Figure 6(b)). The crack
tree method can detect pavement cracks with a simple
background and less crack-like interference because it
is based on tensor voting. But when it is used to deal
with images in an underwater environment, the crack
probability-map obtained from tensor voting promotes
false detection. Moreover the result obtained by the
crack tree method is a probability but not a real one
(see Figure 6(c)), which cannot reflect real characteristics, such as circularity and the gray intensity ratio. The
crack IT method is able to detect underwater dam
cracks with high precision, but it is a difficult task for
the crack IT method to deal with tiny cracks (the missing rate is high). The main reason is that the crack IT
method is based on 2-D feature space and cannot distinguish a crack from its complex background.
However, the proposed method uses local and global
characteristics and can deal with the crack detection
problem effectively (see Figure 6(e)).
With regards to the time cost of performing the
crack detection process, the proposed method and the
crack IT are almost the same (see Table 1). However,
the wasp algorithm needs 89.8(s). The main reason is
that there are multiple iterations in the wasp algorithm.
Because the voting process in the crack tree method
contains ball voting and stick voting, it requires
946.2(s) useless in real-world applications. The results
show that the proposed method can deal with the
underwater dam crack detection problem effectively
and quickly.

Crack classification decision experiments


The crack detection experiments show that the proposed approach can handle images which are full of
uneven illumination, even when the crack is tiny and
other factors interfere. After that, some experiments
are carried out to test the performance of the proposed
crack classification approach based on fuzzy evidence
combination, when the test images and the sample
images are both fuzzy. In these experiments, different
types of dam crack are used as the targets in a complex
environment. There are three possible target types,
identified as ui , i = 1, 2, 3. A CCD is rotated around the
target to obtain different images. The test images are
taken from five angles respectively, which are shown in
Figure 7.
At first, underwater dam cracks are detected by the
proposed algorithm from images with five different
perspectives. Then, the BBA values of the circularity
(eO (a)) of the dam crack can be obtained by (7). In the
same way, the BBA values of the gray intensity ratio
(eT (a)) are obtained by (9).
To reduce the uncertainty of classification, evidence
from both the circularity and the gray intensity ratio
from different perspectives are combined. The BBA
values are updated by the proposed evidence fusion
method (see Section Underwater dam crack classification). To show the performance of the proposed fusion
method (PRO), it is compared with DCR, JUR and
LUR. The final experimental results for the three different types of crack, based on the four methods, are
shown in Figure 8 and the percentage of correct classification by the four methods is shown in Table 2. To
introduce the evidence update process, some calculating
results are given out for the tiny crack, which is the
most difficult type in the crack classification (see
Tables 3 and 4).
The results in Figure 8 and Table 4 show that the
fusion results by the proposed method conform to the
evidence consistently, when the support degree of evidence update changes slightly. The support degree of
the update results by the proposed approach is superior
to the other three methods (see the corresponding BBA

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

551

Figure 7. Underwater dam crack images of five different perspectives for three different classes of crack.

Table 2. The percentage of correct classification for different types of crack.


Type of crack

Classification method

Tiny crack
Normal crack
Through crack

DCR

JUR

LUR

PRO

20%
60%
90%

40%
65%
95%

50%
75%
85%

80%
90%
95%

Table 3. The BBA values of C and D for tiny crack images.


The angle a

a1
a2
a3
a4
a5

eO (a)fu1 , u2 , u3 , YO g

eT (a)fT1 , T2 , YT g

m(u1 )

m(u2 )

m(u3 )

m(YO )

m(T1 )

m(T2 )

m(YT )

0.1698
0.1853
0.1761
0.1674
0.1696

0.1806
0.1763
0.1122
0.1929
0.1317

0.3892
0.3515
0.3734
0.3092
0.4032

0.2604
0.2869
0.3383
0.3305
0.2955

0.5123
0.5458
0.4905
0.4093
0.5755

0.2665
0.1796
0.2298
0.2612
0.2178

0.2212
0.2746
0.2797
0.3295
0.2067

values of the five objectives in Figure 8). When the support degree of the evidence changes dramatically, the
proposed method in this paper is the most sensitive to
change. Furthermore, after re-updating the new evidence supports the original status: which is to say, the

proposed method can recover the BBA values back into


those values before the support degree changed. This
means that the proposed approach can conquer the
influence of abnormal evidence on the update results to
improve its accuracy, which is very important for crack

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

552

Structural Health Monitoring 15(5)


0.5

0.35

DCR
JUR
LUR
PRO

0.45
0.3

0.4
0.35

BPA Values

BPA Values

0.25
0.2
0.15

0.2
0.15

0.1
DCR
JUR
LUR
PRO

0.05
0

0.3
0.25

0.1
0.05
0
6

Target

Target

(b)

(a)
0.4
DCR
JUR
LUR
PRO

0.35

BPA Values

0.3
0.25
0.2
0.15
0.1
0.05

Target

(c)
Figure 8. The BBA values of three different classes of cracks: (a) The tiny crack, (b) The general crack, and (c) The through crack.

Table 4. The BBA values of the evidence update process for tiny crack images.
The angle a

Original
First update
Second update
Third update
Fourth update

eO (a)fu1 , u2 , u3 , YO g

eT (a)fT1 , T2 , YT g

m(u1 )

m(u2 )

m(u3 )

m(YO )

m(T1 )

m(T2 )

m(YT )

0.1698
0.2434
0.2361
0.2158
0.2244

0.1806
0.2283
0.1540
0.2846
0.1584

0.3892
0.3709
0.4282
0.3246
0.4630

0.2604
0.1574
0.1817
0.1749
0.1542

0.5123
0.5608
0.4965
0.4218
0.5892

0.2665
0.1981
0.2585
0.2876
0.2307

0.2212
0.2411
0.2449
0.2906
0.1800

classification in the complex underwater environment.


The results in Table 4 show that all four methods can
classify the through crack easily. However, when dealing with normal and tiny cracks, the accuracy of the
proposed method is higher than the other three.

Conclusions
In this study we have considered the task of underwater
dam crack detection and classification, and proposed a
novel approach. In the proposed approach, the local
features of the image blocks and the global features of
the connected domains are combined and used based

on the analysis of the statistical properties of dam crack


images. Both 2-D and 3-D features are used for crack
clustering analysis. In the proposed approach, circularity and the gray intensity ratio are used to be mapped
into BBAs, which can reduce the uncertainty of classification. Furthermore, a concept of comprehensive reliability is proposed to reduce the impact of artificial
factors. The proposed approach has some good performance. The crack detection algorithm we propose can
handle images which are full of uneven illumination
even when the crack is tiny and interfered with by other
factors. The proposed crack classification algorithm
can solve the classification problem well when the test

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Shi et al.

553

dam crack images and the sample images are both


fuzzy. The experiments show that the proposed method
can detect and classify underwater dam cracks efficiently when the environment is complex and the cracks
are tiny.

10.

11.

Declaration of Conflicting Interests


The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.

12.

Funding

13.

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this
article: This work was supported by the National Natural
Science Foundation of China (grant numbers 61203365,
61273170, 61573128), the Jiangsu Province Natural Science
Foundation (grant number BK2012149), the Fundamental
Research Funds for the Central Universities (grant number
2015B20114), the Specialized Research Fund for the Doctoral
Program
of
Higher
Education
(grant
number
20120094120023), and the colleges and universities in Jiangsu
Province Plans to Graduate Research and Innovation (grant
number CXZZ120255).

14.

15.

16.

17.

References
1. Glisic B and Inaudi D. Development of method for inservice crack detection based on distributed fiber optic
sensors. Struct Health Monit 2012; 11(2): 161171.
2. Wu Z, Li J, Gu C, et al. Review on hidden trouble detection and health diagnosis of hydraulic concrete structures.
Sci China Ser E 2007; 50(1): 3450.
3. Wang S and Ren Q. Dynamic response of gravity dam
model with crack and damage detection. Sci China Technol Sc 2011; 54(3): 541546.
4. Zhou Z, Zhang B, Xia K, et al. Smart film for crack monitoring of concrete bridges. Struct Health Monit 2011;
10(3): 275289.
5. Su H, Li J, Hu J, et al. Analysis and back-analysis for
temperature field of concrete arch dam during construction period based on temperature data measured by dts.
IEEE Sens J 2013; 13(5): 14031412.
6. Bouchette G, Church P, Mcfee J, et al. Imaging of compact objects buried in underwater sediments using electrical impedance tomography. IEEE Trans Geosci Remote
Sens 2014; 52(2): 14071417.
7. Schallhorn C and Rahmatalla S. Crack detection and
health monitoring of highway steel-girder bridges. Struct
Health Monit 2015; 14(3): 281299.
8. Trucco A, Palmese M and Repetto S. Devising an affordable sonar system for underwater 3-d vision. IEEE Trans
Instrum Meas 2008; 57(10): 23482354.
9. Stephan T, Fruehberger P, Werling S, et al. Model-based
image restoration for underwater images. In: SPIE Optical Metrology (international society for Optics and

18.

19.

20.

21.

22.

23.

24.

Photonics), Bellingham, WA, 1416 May 2013,


pp.87911F87911F. Munich, Germany: SPIE.
Jarry J. Sar, nautile, saga, elit four new vehicles for
underwater work and exploration: The ifremer approach.
IEEE J Oceanic Eng 1986; 11(3): 413417.
Chen C, Wang J, Zou L, et al. A novel crack detection
algorithm of underwater dam image. In: Systems and
informatics (ICSAI), 2012 international conference on,
Computer Society Piscataway, NJ, 1920 May 2012,
pp.18251828. Yantai, China: IEEE.
Chiang J and Chen YC. Underwater image enhancement
by wavelength compensation and dehazing. IEEE Trans
Image Process 2012; 21(4): 17561769.
Kabir S, Rivard P, He D, et al. Damage assessment for
concrete structure using image processing techniques on
acoustic borehole imagery. Constr Build Mater 2009;
23(10): 31663174.
Schulein R and Javidi B. Underwater multi-view
three-dimensional imaging. J Disp Technol 2008; 4(4):
351353.
Chen C, Wang J, Zou L, et al. Underwater dam image
crack segmentation based on mathematical morpholog.
Appl Mech Mater 2012; 220: 13151319.
Williams SB, Pizarro O, Johnson-Roberson M, et al.
Auv-assisted surveying of relic reef sites. In: OCEANS,
Computer Society Piscataway, NJ, 1518 September
2008, pp.17. Quebec, QC, Canada: IEEE.
Murphy RR, Steimle E, Hall M, et al. Robot-assisted
bridge inspection after hurricane Ike. In: Safety,
security rescue robotics (SSRR), 2009 IEEE international
workshop on, Computer Society Piscataway, NJ, 36
November 2009, pp.15. Denver, CO: IEEE.
Bernstone C and Heyden A. Image analysis for monitoring of crack growth in hydropower concrete structures.
Measurement 2009; 42(6): 878893.
Xu X and Zhang X. Crack detection of reinforced concrete bridge using video image. J Cent South Univ 2013;
20(9): 26052613.
Fan X, Wu P, Ni J, et al. An illumination balance algorithm based on improved affine shadow formation model
for underwater image. In: Image and signal processing
(CISP), 6th international congress on , Hangzhou, China,
1618 December 2013, pp.910916. New York, NY:
IEEE.
Chang M, Chou YC, Lin PT, et al. Fast and highresolution optical inspection system for in-line detection
and labeling of surface defects. Comput Mater Con 2014;
42(2): 111126.
Oliveira H and Correia P. Automatic road crack detection and characterization. IEEE Trans Intell Transp Syst
2013; 14(1): 155168.
Li L, Sun L, Ning G, et al. Automatic pavement crack
recognition based on bp neural network. PROMETZAGREB 2014; 26(1): 1122.
Nejad FM and Zakeri H. An optimum feature extraction
method based on wavelet-radon transform and dynamic
neural network for pavement distress classification.
Expert Syst Appl 2011; 38(8): 94429460.

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

554

Structural Health Monitoring 15(5)

25. Yin Z, Wu C and Chen G. Concrete crack detection


through full-field displacement and curvature measurements by visual mark tracking: A proof-of-concept study.
Struct Health Monit 2014; 13(2): 205218.
26. Mathavan S, Rahman M and Kamal K. Use of a selforganizing map for crack detection in highly textured
pavement images. J Infrastruct Syst 2015; 21(3), 11 pages.
27. Andalo F, Miranda P and Torres R. Shape feature
extraction and description based on tensor scale. Pattern
Recognit 2010; 43(1): 2636.
28. Ichihashi H and Tanaka H. Jeffrey-like rules of conditioning for the dempster-shafer theory of evidence. Int J
Approximate Reasoning 1989; 3(2): 143156.
29. Liu W. Analyzing the degree of conflict among belief
functions. Artif Intell 2006; 170(11): 909924.
30. Premaratne K, Murthi M, Zhang J, et al. A dempstershafer theoretic conditional approach to evidence

31.

32.

33.

34.

updating for fusion of hard and soft data. In: Information


fusion, 2009. FUSION09. 12th international conference on, Seattle, WA, 0609 July 2009, pp.21222129.
New York, NY: IEEE.
Kulasekere E, Premaratne K, Dewasurendra D, et al.
Conditioning and updating evidence. Int J Approximate
Reasoning 2004; 36(1): 75108.
Zhang Y, Tian X and Ren P. An adaptive bilateral filter
based framework for image denoising. Neurocoputing
2014; 140: 299316.
Zou Q, Cao Y, Li Q, et al. Cracktree: Automatic crack
detection from pavement images. Pattern Recognit Lett
2012; 33(3): 227238.
Anwar S and Abdullah M. Micro-crack detection of multicrystalline solar cells featuring an improved anisotropic
diffusion filter and image segmentation technique. EURASIP J Image Vide 2014; (1): 117.

Downloaded from shm.sagepub.com at UNIV AUTO DEL ESTADO DE MEXICO on December 8, 2016

Das könnte Ihnen auch gefallen