NDT&E International 44 (2011) 563–570

Contents lists available at ScienceDirect

NDT&E International
journal homepage: www.elsevier.com/locate/ndteint

Discontinuities detection in welded joints based on inverse

surface thresholding
Haniza Yazid a,b,n, H. Arof a, Hafizal Yazid c, Sahrim Ahmad c, A.A. Mohamed c, F. Ahmad d
a
Electrical Engineering Department, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
b
School of Mechatronic Engineering, Universiti Malaysia Perlis, 02600, Ulu Pauh, Arau, Perlis, Malaysia
c
Industrial Technology Division, Block 34, Malaysian Institute for Nuclear Technology Research (MINT), Bangi, 43000 Kajang, Malaysia
d
Electrical Engineering Department, UTMSPACE, UTM International Campus, 54100 Kuala Lumpur, Malaysia

a r t i c l e i n f o abstract

Article history: Automated detection of welding defects in radiographic images becomes nontrivial when uneven
Received 28 July 2010 illumination, contrast and noise are present. In this paper, a new approach using surface thresholding
Received in revised form method is proposed to detect defects in radiographic images of welding joints. In the first stage, several
26 April 2011
image processing techniques namely fuzzy c means clustering, region filling, mean filtering, edge
Accepted 3 June 2011
Available online 17 June 2011
detection, Otsu thresholding, and morphological operations method are utilized to locate the area
where defects might exist. This is followed by the construction of the inverse thresholding surface and
Keywords: its implementation to locate defects in the identified area. The proposed method was tested on 60
Non-destructive testing radiographic images and it obtained 94.6% sensitivity. Its performance is compared to that of the
Welded joints
watershed segmentation, which obtained 69.6%.
Inverse surface thresholding
& 2011 Elsevier Ltd. All rights reserved.
Fuzzy c means clustering

1. Introduction an algorithm that automatically locates discontinuities that exist

Products made of polymer and metal are subjected to test and
evaluation before they are considered safe to be used. One of the
parts that deserve thorough scrutiny is the welded joints that
hold the metal structure together. Consequently, industrial radio-
graphy is used as one of the non-destructive techniques to
evaluate the quality of the welded joint. In Ditchburn et al. [1],
the latest development on radiographic inspection, ultrasonic,
alternating current potential drop technique, measurement of
residue stress and in production weld inspection had been
reviewed. The advancement in computer technology spurs the
use of image analysis and pattern recognition systems to assist
human interpreter in assessing the radiographed image in an
automated manner. The purpose of the radiographic interpreta-
tion is to reach a decision on the acceptability or rejection of the
component being assessed. This is based on the adopted code and
standard in use, for example ASME Code Section 1 [2]. There are
many types of discontinuities as described by the standard, which
include cracks, lack of weld penetration, lack of weld fusion,
porosity, void and others. The objective of this work is to develop
n
Corresponding author at: Electrical Engineering Department, Faculty of Engi-
neering, University of Malaya, 50603 Kuala Lumpur, Malaysia.
Tel.: þ603 79675205.
E-mail address: haniza.yazid@gmail.com (H. Yazid).
in radiographic images of welded joints.
Algorithms used for detecting discontinuities in weldment
usually comprise a number of steps. These steps may constitute
pre-processing, segmentation and classification. The segmenta-
tion step is crucial as the result from this stage strongly influence
the final quality of interpretation [3]. In this step, locations and
features of the objects of interest like welding defects are
extracted and separated from the background. Usually, welding
defects in radiographic images have lower intensity than non-
defective area that make up the background. It is sufficient to use
this feature to identify defects as the intensity difference is large.
The intensity of an area is essentially based on the gray level of
pixels in it. The process of partitioning images into objects and
background based on pixels0 gray level values is called thresh-
olding. This is a widely used technique that has been studied and
used extensively [4].
In 2005, a comparative study of non-parametric histogram
thresholding for automatic extraction of weld defect was inves-
tigated by Nacereddine et al. [5]. Four methods based on within
class variance minimization, the minimum error of clustering, the
class entropy maximization and the moment preservation were
implemented and the result showed that entropy maximization
gave the best result. Wang and Wong [6] proposed an adaptive
wavelet thresholding and histogram equalization to improve the
quality of the radiographic image. Three level thresholding, which
is based on maximum fuzzy entropy is adopted to classify the

0963-8695/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ndteint.2011.06.002
564 H. Yazid et al. / NDT&E International 44 (2011) 563–570
image into three regions namely the black, gray and white. The
optimal threshold is obtained using genetic algorithm. In [7],
histogram concavity threshold approach has been used to seg-
ment the defect. The proposed method was compared with
maximum entropy thresholding and Otsu methods. The proposed
method yields a better result compared to both of the conven-
tional thresholding methods. In 2008, Wang et al. [8] introduced a
method based on multiple threshold and support vector machine
to classify the welding defect. In every block that contains the
defect, Hough Transform is performed to remove the noisy pixels
and obtain an accurate segmentation and location of the defect.
Meanwhile Tang et al. [9] claim that it is possible to obtain the
optimal threshold by maximizing the fuzzy exponential entropy
based on the distribution of the defect and the background classes
in the bound histogram. Valavanis and Kosmopoulos [10] utilized
Sauvola local thresholding and graph based segmentation to
segment the weld defect from the background. Then, the defects
were classified using multi-class support vector machine and
neural network classifiers.
In 2002, Wang and Liao [11] implemented background sub-
traction and histogram thresholding in the segmentation phase of
line weld defect detection. 12 numerical features were extracted
and used to classify the defects using fuzzy k-nearest neighbor
and multi-layer perceptron neural networks. Soo et al. [12]
introduced a defect segmentation using background subtraction
and rank levelling. Comparison is made between the rank level-
ling and polynomial surface fitting for background estimation.
The proposed method indicates a better result to isolate the
defects. In 2006, Alaknanda et al. [13] adopted Canny operator
to detect the defect boundaries and then fixed the boundaries
using morphological image processing. Afterward in 2009,
Alaknanda et al. [14] introduced a new two stage water segmen-
tation to reduce oversegmentation in detecting flaws in weldment
images. In the first stage, the watershed transformation is
adopted and mosaic resulting image is obtained. The Otsu thresh-
olding is applied and converted into a binary image. The mor-
phological and top hat transform is utilized onto the binary image
to separate the overlapping defects. In the second stage, the
watershed segmentation is used to obtain well defined bound-
aries while removing the oversegmented regions.
Lawson [15] combined two networks to locate the defect. In the
first step, a multi-layer perceptron backpropogation network was
utilized to locate the weld region based on the gray level and spatial
structure of the input image. Then, in the second step, the multi-
layer perceptron was trained on a test set of segmented image by a
conventional adaptive threshold method to detect the defect area in
the image. Shafeek et al. [16] utilized a host of algorithms such as
gray level histogram, thresholding, histogram stretching and histo-
gram equalization in addition to the 8 neighborhood boundary chain
code algorithm to identify the contour of defects. Then, information
such as the length, width, area and perimeter of the defect are
calculated. Felisberto et al. [17] adopted genetic algorithm to find a
suitable parameters values such as position, width, length and angle
that are best defined in a window and matched it with the model of
the defect sample. However this method is more suitable for
detecting weld bead stretches that are linear or almost linear rather
that the curved welds. Yazid et al. [18] introduced a system to detect
circular discontinuities that are present in the weld joints. The
system utilizes separability filter to identify the best object candi-
dates and the Circular Hough Transform to detect the presence of
circular objects. The algorithm generates satisfactory result in
recognizing circular discontinuities in the radiographic images.
However, the limitation of this approach is its difficulty in detecting
defects, which are not circular. Vilar et al. [19] utilized several
methods including noise reduction, contrast enhancement, Otsu
thresholding and labeling to assist the recognition of the weld
regions and the detection of the weld defects. On the other hand,
Zapata et al. [20] proposed a set of 12 geometrical features, which
characterise the defect shape and orientation.
Attempts to classify the welding flaws types have been made
using various classifiers and a number of studies on various
different features of welding defects have been conducted and
used for classification [21–24]. Recently Liao [25] introduced an
improvement for feature extraction selection in two versions and
is known as ant colony optimization based algorithm. Four
classification methods are used including nearest mean, k-nearest
neighbor, fuzzy k-nearest neighbor and center based nearest
neighbor. In this paper, a new approach is presented to identify
defects using inverse surface thresholding.
2. Methodology
The flow of steps for the proposed method is illustrated in
Fig. 1. The input image is a grayscale radiographic image of
weldment. It is observed that the area that contains defects
usually lies in the middle of the image and is brighter than the
rest of the area. Thus the first step is to identify and isolate this
area using fuzzy c means clustering. The fuzzy c means clustering
segments the image into 5 regions of which two regions with the
highest average intensity are considered as the area that contain
the welding defects and the remaining three regions are com-
bined as the background and excluded from further processing.
Defects that are present in the high intensity area (or the
region of interest) are low in intensity and thus are usually
excluded from the region of interest (ROI). Therefore these defects
appear as holes in the ROI that need to be filled up, processed and
detected properly. A simple region filling algorithm is performed
on the ROI where each row of the ROI is checked horizontally.
When break (hole) is detected, it is filled with the intensities of
the original pixels back. The ROI is then subjected to edge and
valley detection.
Input image
Fuzzy C Means
Region filling
Clustered image
Mean filter Edge detection
Valley detection Otsu thresholding
Otsu thresholding
Inverse Surface
Thresholding
Detected
defects
Final output
Fig. 1. Flow of steps in the proposed method.
 H. Yazid et al. / NDT&E International 44 (2011) 563–570 565

For edge detection, Prewitt operator is used to obtain the edge weldment. The selected region is denoted as R(i,j). Further
information in the ROI. Then Otsu thresholding is performed on information on fuzzy clustering can be found in [27].
the resulting edge image to separate pixels of the weak edges
from the strong ones. Then all pixels of the weak edges are set to 2.2. Morphological valley detection
zero and the strong ones are kept for further processing. Mean-
while, a copy of the ROI is subjected to an averaging process, Valley detection algorithm is a type of morphological proces-
which utilizes a 3  3 window. This is followed by valley detection sing. There are several type of structuring element that can be
using morphological operators. Valley detection is performed to applied such as circle, sphere and square. In this paper, half
catch tiny defects that are present in the image that are too small sphere structuring element is applied. Details of the morphologi-
for the inverse thresholding surface to capture. Then, Otsu cal valley detection is available in [28]. Let G denote the set of
thresholding is performed on the valley image to remove pixels integer points on the spherical surface of a half sphere (Fig. 2).
of weak valley by setting them to zero. If the radius of the sphere is r, G is given by:
In order to construct the inverse surface, the original input image
G ¼ fi,j,gði,jÞg r ri r r
and the edge image are used. The inverse surface is then utilized to jpffiffiffiffiffiffiffiffiffiffiffiffik jpffiffiffiffiffiffiffiffiffiffiffiffik
detect defects in the ROI and the result is combined with that  r 2 i2 r j r r 2 i2 ð5Þ
obtained from the thresholded valley image. This is because result of
the valley detection complements that of the inverse surface thresh- for all i and height;
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
olding especially in detecting small defects that have weak edges
gði,jÞ ¼ r r 2 i2 j2 ð6Þ
often overlooked by the inverse thresholding.

for all j. where byc denotes the greatest integer smaller than or
2.1. Fuzzy C means clustering equal to a real number of G. Given the following definition:
Iðx,yÞð0 r x rm1, 0 ry r n1Þ ð7Þ
Conventional clustering will assign each datum into one and
only one cluster. In contrast, fuzzy clustering assigns each datum I(x,y) is the original image; r is radius of the sphere; byc is the
into several clusters. For each assignment to a cluster, a degree of largest integer smaller than or equal to a; m,n are the length and
membership is specified by a membership grade. A collection of D height of the image.
data (xi; i¼1.2.3yD) to be segmented into C fuzzy groups/classes. max gði,jÞ means the maximum of gði,jÞ overall the points ði,jÞ
There will be a membership matrix u of size C  D. The element uij inside G.
represents the membership of the jth data (xj) to class i. FCM min gði,jÞ means the minimum of gði,jÞ overall the points ði,jÞ
partitions the data into C groups and then find the cluster center inside G.
in each group iteratively while minimizing a cost function of Valley detection algorithm is a closing morphology. The dila-
dissimilarity [26]. Initially, the membership matrix, u is randomly tion is performed first followed by erosion. The overall valley
initialized with values between 0 and 1. detection process is simplified as the following;
X
c
uij ¼ 1, 8j ¼ 1,. . .,d: ð1Þ  Find point (i,j) in half sphere with radius r, then find the
i¼1 corresponding value of g(i,j)
The dissimilarity function, which will be used in FCM as  Perform grayscale dilation onto each pixel in the image and
follow; obtain D(x,y)
D(x,y):¼max{I(x  1,y  1)þg(i,j)}; ( 5o ¼io ¼5)
X
c X
d
 Grayscale erosion is performed onto D(x,y) and stored into
J¼ ðuij Þm d2ij ð2Þ
i¼1j¼1
K(x,y)
K(x,y):¼min{D(x  1,y  1) g(i,j)};
where uij is the in between 0 and 1; ci the cluster center of fuzzy  Perform V(x,y)¼K(x,y) I(x,y) for each pixel (x,y)
group i; dij the Euclidean distance between ith cluster and the jth V(x,y)¼1 if (x,y) is a valley pixel
data point and m the weighting exponent. V(x,y)¼0 otherwise
A minimum of dissimilarity function can be reached using the  If V(x,y) 4 ¼T for a threshold value T, (x,y) is determined to be
following formula a valley.
Sdj¼ 1 um
ij xj
ci ¼ , 8i ¼ 1,    ,C ð3Þ z
Sdj¼ 1 ðuij Þm

1 (0,0,r)
uij ¼ : ð4Þ
SCk ¼ 1 ðdij =dkj Þ2=ðm1Þ
The determination of the cluster0 s center ci and the member-
ship matrix can be illustrated in the following steps:
Step 1: Initialize the membership matrix u with random value
between 0 and 1.
Step 2: Calculate c fuzzy cluster centers ci.
Step 3: Compute the cost function and terminate if u below a y
certain tolerance value. g (i,j)
Step 4: Compute a new u and go to step 2 again.
In our experiment, the input image is segmented into two
regions. The region with the higher average intensity will be x (i,j)
selected for further processing, since the one with the lower
intensity is considered as the background that does not contain Fig. 2. Half sphere structuring element.
566 H. Yazid et al. / NDT&E International 44 (2011) 563–570
2.3. Otsu thresholding
Otsu threshoding is a method that selects a threshold that
minimized the intra class variance. In this work, Otsu thresholding
is utilized to threshold the edge and valley pixels. We expect the edge
values of pixels at the border of the discontinuities to be higher than
other areas in the image. Therefore, we can separate border pixels
from the others by thresholding the edge image E(i,j). Otsu thresh-
olding splits the edge values of E(i,j) into two groups where pixels
from the group with the higher mean (border pixels) are kept while
pixels of the other group are set zero. Otsu method is also used on the
valley image obtained from the morphological operation to eliminate
low valley values from V(i,j) by setting them to zero. After this oper-
ation, only the strong edge and high valley are kept and considered.
Further details on this method can be found in [29]. As a consequence
of this step, we could identify the locations of strong edge and valley
pixels in the region of interest (ROI) now denoted as R(i,j).
2.4. Inverse surface construction
The novelty of our approach lies in the use of an inverse surface
for locating the defects in the R(i,j) region. In constructing the
inverse surface, the locations of edge points (pixels) are given by the
nonzero pixels of Otsu thresholded edge image E(i,j). The intensity of
the inverse surface S(i,j) is derived from the gradients of the original
Fig. 3. Processing directions (a) first pass in left-right-top-bottom(LRTB) direction, (b)second pass in right-left-bottom-top (RLBT) direction, (c) third pass in top-bottom-
right-left (TBRL) direction and (d) fourth pass in bottom-top-left-right (BTLR) direction.
image I(i,j) and the negative image N(i,j) in 4 passes. The difference
between one pass to another is the path of processing the pixels in
the image as shown in Fig. 3. Firstly, for each pass a primary surface
Pm(i,j) (m¼1, 2, 3, 4) is initialized to the original image I(i,j). Then the
intensity of pixels on the primary surface Pm(i,j) is reconstructed
depending on whether the pixels are border pixels or non-border
pixels. Border pixels are those with strong edge (E(i,j)40) and their
gradients are made equal to those of the negative image N(i,j). Non-
border pixels are those with weak edge (E(i,j)¼ 0) and their
gradients are set equal to those of the original image I(i,j). The final
inverse surface is the average of the four primary surfaces.
Of course, the gradients of a pixel (i,j) in an image can be
measured in four directions. For each pixel (i,j), the four gradients
are calculated in horizontal, vertical, diagonal and anti-diagonal dire-
ctions between the pixel (i,j) and four of its nearest neighbors that
have been traversed. The reconstruction of the primary surface P1(i,j)
for the first pass (LRTB) is summarized as follows. The formation of
the other three primary surfaces can be generalized from P1(i,j). Keep
in mind that only locations (pixels) that belong to R(i,j) are processed.
1. Obtain the negative image of I(i,j) by negating it using Eq. (8).
The gradients of each pixel in the negative image N(i,j) should
be the opposite of those in the original image I(i,j)
Nði,jÞ ¼ 2552Iði,jÞ ð8Þ
 H. Yazid et al. / NDT&E International 44 (2011) 563–570 567

2. Initialize the primary surface P1(i,j) to the original image I(i,j)
P1 ði,jÞ ¼ Iði,jÞ
3. For every pixel(i,j), adjust the intensity of P1(i,j) according to
the following cases
(i) Case 1: If the pixel(i,j) is a border pixel (E(i,j)40) we
equate the gradients of P1(i,j) to those of the negative
image N(i,j).
Find the gradients of the pixel(i,j) in four directions relative
to four of its nearest neighbors that have been traversed in
the negative image. In LRTB pass, the neighboring pixels are
(N(i,j 1), N(i 1,j), N(i 1,j-1), N(i 1,jþ1)) and the gra-
dients in horizontal, vertical, diagonal and anti-diagonal
directions are
D1 ¼ Nði,jÞ2Nði,j1Þ D2 ¼ Nði,jÞ2Nði1,jÞ
D3 ¼ Nði,jÞ2Nði1,j1Þ D4 ¼ Nði,jÞ2Nði1,j þ1Þ
Then using each gradient, we estimate the value of P1(i,j)
using its corresponding neighbors as follows
e1 ði,jÞ ¼ P1 ði1,j1Þ þ D1 e2 ði,jÞ ¼ P1 ði,j1Þ þ D2
e3 ði,jÞ ¼ P1 ði1,jÞ þ D3 e4 ði,jÞ ¼ P1 ði1,j þ 1Þ þ D4
The four estimates are then averaged to give us the final
value of P1(i,j) for the pass
P1 ði,jÞ ¼ ðe1 ði,jÞ þ e2 ði,jÞ þ e3 ði,jÞ þe4 ði,jÞÞ=4:0
(ii) Case 2: If the pixel(i,j) is a non-border pixel (E(i,j) ¼0) the
gradients of P1(i,j) are the same as those of the original
image I(i,j).
The neighboring pixels are I(i,j 1), I(i 1,j), I(i 1,j  1),
I(i  1,j þ1) and the gradients in horizontal, vertical, diag-
onal and anti-diagonal directions are
(c) For BTLR pass, N(i,j 1), N(iþ1,j), N(i 1,j 1), N(iþ
1,j1) and I(i,j1), I(iþ1,j), I(i1,j 1), I(iþ1,j 1) are
the respective neighbors used for border and non-border
pixels in reconstructing P4(i,j).
4. The final inverse surface S(i,j) is the average of the four
primary surfaces
Sði,jÞ ¼ ðP1 ði,jÞ þP2 ði,jÞ þ P3 ði,jÞ þ P4 ði,jÞÞ=4:0 ð11Þ
2.5. Surface thresholding
In surface thresholding, the main task is to adjust the position of
the inverse surface S(i,j) so that only the areas containing defects
will overlap with the original image I(i,j). Fig. 4(a) shows the
intensities of pixels that fall on a line drawn on I(i,j) while
Fig. 4(b) shows the same pixels on S(i,j). The positions of the
edge pixels are marked with ‘‘ þ’’. The main idea is to treat S(i,j) as
a thresholding surface and to position the two surfaces as in Fig. 5
such that the edge pixels coincide. This is done by adjusting the
intensities of S(i,j) by adding/subtracting a constant to it, in order
to move the inverse surface up or down. Since edge pixels do not
have the same intensity, the strategy is to minimize the absolute
difference between the edge pixels on the input image I(i,j) and
the adjusted inverse surface S(i,j)þk, where k is the added
constant. The value of k that achieves this minimum sum of
absolute difference is denoted as ko.
Once the right position of the inverse image (or constant ko) is
found, we can extract the defects as the area where I(i,j) is greater
ð9Þ than S(i,j)þko. In Fig. 5 the defects are designated by the area
where the bold line is higher than the thin line. The whole process
of the surface thresholding can be summarized as follows;
Step 1: Construct the inverse image S(i,j) as directed in
previous section
Step 2:Find the value of a constant ko that minimizes the sum
of the absolute difference of I(i,j) S(i,j)þk for the edge pixel at

D1 ¼ Iði,jÞ2Iði,j1Þ D2 ¼ Iði,jÞ2Iði1,jÞ
D3 ¼ Iði,jÞ2Iði1,j1Þ D4 ¼ Iði,jÞ2Iði1,j þ1Þ

The estimates for P1(i,j) using its corresponding neighbors

as follows

e1 ði,jÞ ¼ P1 ði1,j1Þ þ D1 e2 ði,jÞ ¼ P1 ði,j1Þ þ D2

e3 ði,jÞ ¼ P1 ði1,jÞ þ D3 e4 ði,jÞ ¼ P1 ði1,j þ 1Þ þ D4

They are then averaged to give us the final value of P1(i,j)

for the pass
P1 ði,jÞ ¼ ðe1 ði,jÞ þ e2 ði,jÞ þ e3 ði,jÞ þe4 ði,jÞÞ=4:0 ð10Þ Fig. 4. (a) A cross section of I(i,j) and (b) the same cross section on S(i,j).

For the other passes, steps (1) and (2) are identical and for
step (3) the neighbors to consider for horizontal, vertical,
diagonal and anti-diagonal directions are
(a) For RLBT pass, N(i,jþ 1), N(iþ1,j), N(iþ 1,jþ1),
N(iþ1,j 1) and I(i,jþ 1), I(iþ1,j), I(iþ1,jþ1), I(iþ1,j-1)
are the respective neighbors used for border and non-
border pixels in reconstructing P2(i,j).
(b) For TBRL pass, N(i,jþ 1), N(i 1,j), N(iþ1,jþ1), N(i 1,jþ
1) and I(i,jþ1), I(i1,j), I(iþ1,jþ1), I(i 1,jþ1) are the
respective neighbors used for border and non-border
pixels in reconstructing P3(i,j). Fig. 5. Combination of image surface (bold line) and inverse surface (thin line).
568 H. Yazid et al. / NDT&E International 44 (2011) 563–570

Fig. 6. (a) Test image, (b) fuzzy clustered image, (c) region filled image, (d) thresholded edge image, (e) thresholded valley image, (f) inverse surface image and
(g) final thresholded image.

the border S(i,j) where k is an integer that varies from  255 to 3. Results and discussions
255.
2 The algorithm is developed using C programming in LINUX
argmin X operating system with Intel Core 2 Duo processor. The system
k0 ¼ 4 9Iði,jÞSði,jÞ þ k9,
k was tested on 60 radiography images. It is observed that, the new
border pixels ði,jÞ
proposed method is able to detect defects of varying shapes and
sizes. Fig. 6(a) shows the original radiographic image while
kA f255,254,. . .0. . .,254,255g ð12Þ
Fig. 6(b) is the clustered image of the FCM operation wherein
the holes can be clearly observed. The result of hole filling on the
Step 3: Segment R(i,j) into defect and background.
clustered image is shown in Fig. 6(c). Edge detection is performed
( on the image in Fig. 6(c) followed by the Otsu thresholding that
0ðdefectsÞ If Iði,jÞ 4 Sði,jÞ þ ko
resultði,jÞ ¼ ð13Þ removes weak edges. The result is shown in Fig. 6(d). Fig. 6(e) is
255ðbackgroundÞ If Iði,jÞ o Sði,jÞ þ ko
the outcome of morphological valley detection on Fig. 6(c) where
 H. Yazid et al. / NDT&E International 44 (2011) 563–570 569

Fig. 7. (a) Original image, (b) valley detection and (c) final thresholded image.

Table 1
Comparison between human grader and the proposed method.
Original image Human grader Inverse surface Watershed segmentation
thresholding

the weak valley has been removed by the Otsu thresholding.
Meanwhile Fig. 6(f) shows the inverse image after inverse surface
thresholding. Lastly, the final result is indicated in Fig. 6(g).
In order to demonstrate the importance of inverse threshold-
ing; the test image, the valley detected image V(i,j) and the final
thresholded image are shown in Fig. 7(a), (b) and (c), respectively.
The circle on the images shows the defect missed by valley
detection but picked up by inverse thresholding.
To show the contribution of the proposed method, the sensi-
tivity and specificity with and without inverse surface thresholding
is calculated. Sensitivity measures the ratio of the actual positive
which is correctly identified meanwhile specificity measures the
ratio of the negative which are correctly identified. In this paper,
sensitivity measures the proportion of the defect pixels that is
correctly identified as defects while specificity measures the
proportion of the normal pixels that is correctly identified as
normal. The sensitivity of the inverse surface threholding is 94.6%
while its specificity is 99%. On the other hand, the valley detection
alone obtains a lower sensitivity of 75% and the same specificity of
99%. For further comparison, result of the method is compared to
570 H. Yazid et al. / NDT&E International 44 (2011) 563–570
that of the watershed segmentation. Table 1 shows several
examples of detected defects from human grader, inverse surface
thresholding and watershed segmentation.
In watershed segmentation, several methods had been
adopted namely thresholding, extended maxima, minima imposi-
tion and contrast adjustment. One major drawback of this method
is the manual selection of the threshold value to convert the
grayscale into binary image before watershed segmentation is
performed. In watershed segmentation method, the sensitivity
achieved is 69.6% while the specificity is 98.8%.
4. Conclusions
The novelty of this work is in the application of inverse surface
thresholding for defect detection. The method is simple and yet
powerful. The inverse thresholding method is introduced to
construct a new thresholding surface to extract the defects. Then
the least square distance measure is used to adjust the position of
the inverse image for optimum thresholding. The main advantage
of this proposed method is it can be applied to databases with
uneven illumination and contrast. This method does not depend
on selected parameters that need to be changed for different
databases. The proposed method demonstrates 94.6% sensitivity
and 99% specificity in detecting defects in weldment images. In a
few images, some of the defects are either not detected or overly
detected. However, most of the time the missing defects are
either too small or blurred.
Acknowledgement
The authors wish to thank Materials Technology Group for
providing the image data.
References
[1] Ditchburn RJ, Burke SK, et al. NDT of welds: state of the art. NDT&E
International 1996;29(2):111–7.
[2] American Society of Mechanical Engineers (ASME), Section 1; 1998, pp. 204.
[3] Soler L, Malandrin G, Delinguette H. Segmentation automatique: application
aux angioscanner 3D. Revue de Traitement de Signal 1998;15.
[4] Sezgin M, Sankur B. Survey over image thresholding techniques and quanti-
tave performance evaluation. Journal of Electronic imaging 2004;13(1):
146–65.
[5] Nacereddine N, Hamami L, Tridi M, Oucief N. Non-parametric histogram
based thresholding methods for weld defect detection in radiography. World
Academy of Science, Engineering and Technology 2005;9.
[6] Wang X, Wong BS. Radiographic image segmentation for weld inspection
using a robust algorithm. Research in Nondestructive Evaluation 2005;16(3):
131–42.
[7] Saravanan TBS, John Philip, Jayakumar T, Baldev Raj. Segmentation of defects
from radiography images by the histogram concavity threshold method.
Insight—Non-Destructive Testing and Condition Monitoring 2007;49(10):
578–84.
[8] Wang Y, Sun Y, et al. Detection of line weld defects based on multiple thresholds
and support vector machine. NDT&E International 2008;41(7):517–24.
[9] Tang Y, Zhang X, et al. Application of a new image segmentation method to
detection of defects in castings. The International Journal of Advanced
Manufacturing Technology 2009;43(5):431–9.
[10] Valavanis I, Kosmopoulos D. Multiclass defect detection and classification in
weld radiographic images using geometric and texture features. Expert
Systems with Applications 2010;37(12):7606–14.
[11] Wang G, Liao TW. Automatic identification of different types of welding
defects in radiographic images. NDT&E International 2002;35(8):519–28.
[12] Soo SL, Ratnam MM, Samad Z, Khalid MA. Segmentation of weld defects in
radiographs using rank-levelling and background subtraction. Insight: Non-
Destructive Testing and Condition Monitoring 2008;50(4):188–94.
[13] Alaknanda, Anand RS, et al. Flaw detection in radiographic weld images
using morphological approach. NDT&E International 2006;39(1):29–33.
[14] Alaknanda Anand RS, Kumar Pradeep. Flaw detection in radiographic weld-
ment images using morphological watershed segmentation technique.
NDT&E International 2009;42:2–8.
[15] Lawson SE, Parker GA. Intelligent Segmentation of industrial radiographic
images using neural networks, in: Proceedings of SPIE Architectures and
Systems Integration III 2347, (1994), pp. 245–55.
[16] Shafeek HI, Gadelmawla ES, et al. Assessment of welding defects for
gas pipeline radiographs using computer vision. NDT&E International
2004;37(4):291–9.
[17] Felisberto MK, Lopes HS, et al. An object detection and recognition system
for weld bead extraction from digital radiographs. Computer Vision and
Image Understanding 2006;102(3):238–49.
[18] Yazid H, et al. Circular discontinuities detection in welded joints using
circular hough transform. NDT&E International 2007;40:594–601.
[19] Vilar R, Zapata J, et al. An automatic system of classification of weld defects
in radiographic images. NDT&E International 2009;42(5):467–76.
[20] Zapata J, Vilar R, et al. An adaptive-network-based fuzzy inference system for
classification of welding defects. NDT&E International 2010;43(3):191–9.
[21] Liao TW, Ni J. An automated radiographic NDT system for weld inspection:
part I—weld extraction. NDT&E International 1996;29(3):157–62.
[22] Liao TW, Li Y. An automated radiographic NDT system for weld inspection:
part II—Flaw detection. NDT&E International 1998;31(3):183–92.
[23] Liao TW LD, Li YM. Detection of welding flaws from radiographic images with
fuzzy clustering methods. Fuzzy Sets System 1999;108(2):145–58.
[24] Liao TW, Li D, et al. Extraction of welds from radiographic images using fuzzy
classifiers. Information Sciences 2000;126(1–4):21–40.
[25] Liao TW. Improving the accuracy of computer-aided radiographic weld
inspection by feature selection. NDT&E International 2009;42(4):229–39.
[26] Chuai-Aree SS, Lursinsap C, Sophatsathit P, Siripant S, Fuzzy C-Mean: A
statistical feature classification of text and image segmentation method, in:
Proceedings of International Conference on Intelligent Technology 2000,
December 13–15, Assumption University Bangkok, Thailand, 2000,
pp. 279–84.
[27] Sergios Theodoridis, Koutroumbas Konstantinos. Pattern Recognition.3rd
Edition USA: Academic Press; 2006.
[28] Fauzan Ahmad, Face Localization for Facial Features Extraction using Sym-
metrical Filter and Linear Hough transform, Dissertation (M.Eng.Sc.), Elec-
trical Engineering Department, Faculty of Engineering, University of Malaya,
2007.
[29] Otsu. A threshold selection method from gray-level histogram. IEEE Transac-
tions on Systems Man and Cybernetics 1979;9(1):62–6.