ARTICLE IN PRESS

NDT&E International 40 (2007) 594601

www.elsevier.com/locate/ndteint

Circular discontinuities detection in welded joints using

Circular Hough Transform
Hazal Yazida,, Haniza Yazidb, Mohd Haruna, Shukri Mohda, A. Aziz Mohameda,
Mohamed Rizonb, Shaharudin Sayutia
a

Industrial Technology Division, Block 34, Malaysian Institute for Nuclear Technology Research (MINT), Bangi, 43000 Kajang, Malaysia
b
Kolej Universiti Kejuruteraan Utara Malaysia (KUKUM), Jalan Kangar-Arau, 02600 Jejawi, Perlis, Malaysia
Received 16 November 2006; received in revised form 26 March 2007; accepted 25 May 2007
Available online 13 June 2007

Abstract
Conventional radiography is one of the common non-destructive testing which employs manual image interpretation. The
interpretation is very subjective and depends much on the inspector experience and working conditions. It is therefore useful to have
pattern recognition system in order to assist human interpreter in evaluating the quality of the radiographed sample, especially
radiographic image of welded joint. This paper describes a system to detect circular discontinuities that are present in the joints. The
system utilizes together two different algorithms, which are separability lter to identify the best object candidate and Circular Hough
Transform to detect the presence of circular shape. The result of the experiment shows a promising output in recognition of circular
discontinuities in a radiographic image. This is based on 81.82100% of radiography lm with successful circular detection by using
template movement of 10 pixels.
r 2007 Elsevier Ltd. All rights reserved.
Keywords: Non-destructive testing; Algorithm; Welded joints; Separability lter; Circular Hough Transform

1. Introduction
As one of the non-destructive testing (NDT), industrial
radiography plays a vital role in detecting volumetric
discontinuities present in the critical area of engineering
component. The most common critical area is welded joint
that forms part of the structures. A variety of NDT
methods are available for the inspection of welded joint [1].
The emerging trend of radiography as one of the NDT
methods is towards automated inspections [26]. This is
feasible as the information of the discontinuities is in the
form of radiographic images. Interpretation and evaluation
of the radiographic images are very important stages for
acceptance of any engineering component. Realization of
this crucial process dealing with images, image analysis and
Corresponding author. Tel.: +6 03 89250510x1194;
fax: +6 03 89250907.
E-mail addresses: hazal@mint.gov.my (H. Yazid),
haniza.yazid@gmail.com (H. Yazid).

0963-8695/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ndteint.2007.05.004

pattern recognition system using a computer have become

the subject of interest in the eld of industrial radiography.
This trend of the radiographic interpretation and evaluation method is aimed towards an automated inspection,
which results in better consistence assessment. Basically, it
requires development of image enhancement and object
detection technique to assist human interpreter in evaluating the quality of welded joints.
The purpose of the radiographic interpretation 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 [7]. There are many types of discontinuities described by
the standard, which includes cracks, lack of weld penetration, lack of weld fusion, porosity, voids and others.
Every discontinuity has its own shape and contributes to
different level of severity to the weldment. According to
ASME Code Section 1 certain types of discontinuities are
classied as major and unacceptable. These unacceptable
discontinuities are called defects and subject to repair and

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601

reexamination. Such defects are cracks, lack of weld

penetration and lack of weld fusion. Other than that
including rounded discontinuity is still depending on the
size, distribution and concentration as specied by the
standard.
In this paper, the work is intensied in nding one
type of discontinuity that is circular and appeared as
rounded image on the radiograph. The circular shape is
treated as an object of detection. The approach is to
optimize separability lter technique and Circular Hough
Transform (CHT) in order to obtain the result. The lter
technique is used to recognize the best object candidates
and later identied by CHT. Similar approaches have been
used to detect coconut as circular object in coconut tree
image with prominent success rate by Haniza et al. [8].
The use of CHT extends also to other research elds
such as iris detection for face recognition and ngertips
detection [9,10].
1.1. Industrial radiography
In industrial radiography, typical images are generated
from using two types of radiation source, namely X-ray
and gamma ray. Both differ primarily in energy and
penetration power. X-ray suffers lower penetration power
and energy as compared to gamma ray. The advantage of
X-ray over gamma ray is that it has a wider energy
spectrum while energy for gamma ray is discrete [9]. This
enables X-ray image to acquire higher latitude and better
contrast. In other words, for most of the welded
components, good image quality is attainable through
recommended radiographic practice, thus, making the
conventional radiography interpretation feasible to be
carried out in an automated manner.
2. Methodology
A systematic procedure is established in analyzing the
image information and inferring patterns in such a way
that would yield the targeted object. The innovative
approach in this procedure is to separate good image
quality and directly subject to separability lter and CHT
without undergoing the preprocessing step. The other
image still requires the preprocessing step for image
enhancement. Fig. 1 shows the ow procedure of the
pattern recognition system. Initially the resulted radiographic lm is subjected to lm quality inspection
manually. The inspection covers many criteria that include
lm density, sensitivity, presence of artifacts and identication. All the criteria have to be met as governed by the
adopted standard in order for the radiographic lm to be
accepted for evaluation. Even though the radiographic lm
is accepted for evaluation, it does not mean that the image
appeared on the lm is good in terms of contrast. A lm
would suffer poor radiography image quality or contrast
once the requirement of the criteria is only met on a

595

Radiographic
Image

Good
Image
Digitization

Manual Film
Quality
Inspection
Contrast
Enhancement

Poor
Image

Digitization

Defect
Segmentation
and Noise
Reduction

Separability
Filter

End result

Circular
Hough
Transform

Fig. 1. Flow procedure of the pattern recognition system.

minimal basis. Therefore poor and good radiography

image quality is identied before the lm is digitized.
2.1. Preprocessing
Once poor radiography image quality is identied and
digitized, preprocessing step is initialized. The image is
subjected to contrast enhancement. In this work, 8-bit
image is used as an input image which carries 0255 gray
levels. The image is in the form of PGM (Portable Gray
Map) format and comprises 320  240 pixels. Every single
pixel in an image carries its own intensity. The distribution
of the intensities of the pixels could be viewed statistically
by plotting the intensity histogram. By manipulating this
histogram, the required intensity distribution or contrast is
achieved. It is common to have narrow intensity distribution, which causes poor image contrast, and this requires
rectication to acquire a wider intensity distribution. In
this work, histogram equalization technique has been used
to improve the image contrast. By using this technique, it is
possible to acquire a wider intensity distribution as
compared to the original intensity distribution prole.
One typical example is clearly shown in Fig. 2. Therefore
the image contrast is effectively enhanced. This technique is
particularly suitable for region-based feature extraction.
Typical radiography image which has good contrast and
does not require histogram equalization is depicted in
Fig. 3. Poor radiography image quality that requires
histogram equalization is depicted in Fig. 4. The result
of histogram equalization on poor radiography image
quality is depicted in Fig. 5. The same result of histogram
equalization is depicted in Figs. 6 and 7, respectively.

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601

596

Original Distribution

After HistogramEqualization

99

165

Pixel Counts Vs.Gray Level

1400
1200

Pixel Counts

1000
800
600
400
200

253

242

231

220

209

198

187

176

154

143

132

121

110

88

77

66

55

44

33

22

11

0
Gray Level

Fig. 2. The effect of histogram equalization on pixel counts distribution.

Fig. 3. Typical image, which has good contrast and does not require
histogram equalization.

from the background. Edge nding and mathematical

morphology are used in this step. Edge-nding technique
provides detection of pixels that belong to the borders of
the objects. As a result, the image contains only a collection
of edge pixels. The collection of edge pixels covers most of
our targeted object and also unwanted information or
noise. Unwanted information is reduced by using mathematical morphology. By doing this we will able to obtain
edge pixel that mostly corresponds to the targeted
object. There are various edge-nding techniques available
and in this work Sobel edge detector has been used.
This technique reveals thick edge as compared to other
technique particularly Canny [10]. Sobel edge detector is
preferred due to its ability to reduce the chances of false
object detection as compared to Canny. Canny reveals too
many details, which in turn cause the CHT to give a wrongtargeted object. Edge information is very important and
comes into play when performing CHT step. Sobel edge
detector is a gradient method, which detects the edges by
looking for the maximum and minimum in the rst
derivative of the image. Gradient is reected by the
difference in intensity from one pixel to the next. Sobel
edge detector uses a pair of 3  3 convolution masks, one
estimating the gradient in the X-direction, Gx and the other
estimating the gradient in the Y-direction, Gy. The Sobel
convolution masks are shown in Fig. 8.
The result of each convolution mask is combined to get
exact magnitude value using the formula

Fig. 4. Typical image, which has poor contrast and requires histogram
equalization.

jGj

2.2. Defect segmentation

This step is crucial to distinguish between the object of
interest and the background. The term segmentation
implies segmenting the object of interest or foreground

p
Gx2 Gy2 .

(1)

The Sobel technique deals with six pixels and thus has a
better average over the neighboring pixels. This approach
with slight blurring can eliminate most noise found in the
image.
Mathematical morphology is considered as region-lling
technique to provide the complete objects. This technique

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601

597

is implemented by using closing and opening operation.

First the closing operation is carried out followed by
opening operation. The closing operation is able to ll in
small holes and removes much of the white pixel noise.
This would yield a fairly clean image as compared to
previous image before the operation. Then opening operation takes place by opening small gaps or spaces between
touching objects in an image. The opening operation can
separate objects that are connected in a binary image. Both
operations generate a certain amount of smoothing on an
object contour given a smooth structuring element.
Fig. 5. Image as in Fig. 4 after histogram equalization.

2.3. Separability filter

After completing segmentation step, the image is
subjected to feature extraction step by using separability
lter technique. This technique is a template-based method
and was proposed by Fukui [11]. This technique employs a
template of Fig. 9 by placing the template at each point
(xi, yi) and then computation on the separability between
the two regions R1 and R2 in the template with size r is
carried out.
Eq. (4) is used to measure the separability between the
two regions. r is treated as a xed parameter based on
maximum acceptable circular defect size specied by the
standard. r is xed at 3, 4 and 5 pixels that correspond
approximately from 0.75 to 1.5 mm defect radius.

Fig. 6. Another image, which has poor contrast and requires histogram
equalization.

N
X

Ixi ; yi  Pm 2 ,

(2)

i1

B n1 P1  Pm 2 n2 P2  Pm 2 ,
Z

B
,
A

(3)
(4)

where nk (k 1,2) is the number of pixels in Rk;

N n1 n2 ; Pk (k 1,2) the average intensity in Rk; Pm
the average intensity in the union of R1 and R2; and Ixi ; yi
the intensity values of pixels xi ; yi in the union of R1
and R2.

Fig. 7. Image as in Fig. 6 after histogram equalization.

Fig. 8. Sobel convolution masks.

Fig. 9. Circle template for object candidate detection.

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601

598

The points (xi,yi) obtained after the opening morphology

in the segmentation process is used as the input to the
separability lter. The range of diameter for the R1 is set to
{2rL,y,2rU} and in this work rL is 3 pixels and rU is 5
pixels. Diameter of R2 region is obtained after considering
the gap of 0.5r. The template of Fig. 9 will move every x
pixels from center to the left and right and y pixels from
center to upwards and downwards as shown in Fig. 10. The
circles that give the local maxima of the separability Z are
selected as object candidates.
In this work; x 10 pixels, y 10 pixels.
2.4. Circular Hough Transform

Fig. 11. Image space.

Circular Hough Transform or CHT is used as pattern

classication to identify the targeted object. This is the last
step need to be carried out before obtaining the end result.
Hough transform is a widespread technique in image
analysis. It was rst introduced by Paul Hough in 1962. It
was rst used to detect straight lines and later extended to
cater to other parametric shapes such as circular, ellipses
and nally being generalized to cater to any parametric
shape. CHT was proposed by Duda et al. [12]. It has been
recognized as a robust technique for curve detection.
CHT is used to transform a set of feature points in the
image space into a set of accumulated votes in a parameter
space. Therefore the difculties in global detection problem
in the image space are reduced to peak nding in a
parameter space. The image space and parameter space are
shown in Figs. 11 and 12, respectively. In this work, the
interesting image points are those that belong to a circle
characterized by its center coordinates (x0, y0) and its
radius r can be represented by
x  x0 2 y  y0 2 r2 .

(5)

This equation indicates that a point (x, y) of the image

space can be mapped to a cone surface in the (x0, y0, r)
parameter space, which represents all possible circles
passing through (x, y). In this work, the edge orientation

Fig. 10. Template movement.

Fig. 12. Parameter space.

information is used to reduce the ranges of parameters and

the complexity of the voting operation. The center of a
circle must be on the straight line passing through the edge
point (x, y) in a direction given by the gradient value at the
edge point. The use of edge orientation information is
proposed by Kimme et al. [13] and this limits the possible
positions of the center for each edge point. This method
applies the use of an arc plotted perpendicular to the edge
orientation at a distance r from the edge point.
Each edge point contributes a circle of radius r to an
output accumulator or parameter space. The output
accumulator space has a peak where these contributed
circles overlap at the center of the original circle. The peak
point is projected back to the image space as (x0, y0, r)
which reveal the nal detected circle object. This circle
consists of many edge points at its perimeter.
Hough transform and its modied version offer several
advantages. First the possibility of parallel processing since
each image point is treated independently and the
algorithm is able to work even when the shape is occluded.
Secondly, the method is robust even in the presence of
noise. Noisy image points are very improbable to
contribute to a peak in the parameter space. Thirdly,
the method is able to detect different instances of the
desired shape at the same time, depending on the number
of peaks, which are considered in the parameter space.

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601

599

The advantages over other techniques make it an attractive

method for circle detection in many research elds.
3. Results and discussions
The system was tested by evaluating 11 radiography
images. The radius of the object as the input parameter was
set to 3, 4 and 5 pixels, which correspond to approximately
1.53.0 mm defect diameter size. Table 1 shows the results
of the experiment.
First and second detections successfully detect the
present of circular discontinuities for all the radiography
images. Typical successful detection is depicted in Fig. 17.
The same image before the detection is depicted in
Fig. 16. Third and fourth detections also give higher
percentage of successful detection with 90.9% and 81.82%,
respectively.
The reduction in successful or correct detection is due to
the present of ambiguous object detected by the system as
in Fig. 19. Image before detection is depicted in Fig. 18. It
is not surprising for CHT to give false detection since the
presence of ambiguous object does exist after applying
Sobel edge detector as in Fig. 14. Thus the performance
of the proposed system is far better if the ambiguous
object outside the region of interest is eliminated after
applying the edge detector. Fifth detection gives 54.55% of
successful detection. Almost half out of the total lm is
able to give correct detection. Besides the factor of
ambiguous object, this is also an indicator that all the
circular discontinuities that are present in the area of
interest have been detected by the system and the rest of the
detection become false detection in order for a computer
program to complete.
In this work, Sobel edge detector has been selected.
This type of edge detector would reduce the amount of
ambiguous object and reveal cleaner image as compared to
Canny edge detector. Both of the edge detectors are
depicted in Figs. 13 and 14 for comparison. Some of the
image does not require histogram equalization due to the
fact that it has already a good image quality and contrast.
This typical image is depicted in Fig. 3. On the contrary,
poor image quality and contrast image are required
to undergo histogram equalization. This typical image
is depicted in Fig. 4. The same image after applying
histogram equalization is depicted in Fig. 5. The improved

Fig. 13. Typical image after applying Canny edge detector for image in
Fig. 3.

Fig. 14. Typical image after applying Sobel edge detector for image in
Fig. 3.

Fig. 15. Effect of mathematical morphology. Image (a) indicates image

before being subjected to mathematical morphology. Image (b) indicates
image after being subjected to mathematical morphology.

Table 1
Success in circular detection
No. of success in circular
detection

Percentage of lm according to
success no. of detection (%)

1
2
3
4
5

100
100
90.9
81.82
54.55

image contrast after histogram equalization eases into the

following process ow of object detection. Additionally, in
this work, the effect of mathematical morphology eases the
burden of CHT by removing pixels or unwanted information and thus leaves only potential pixels for CHT to do the

ARTICLE IN PRESS
600

H. Yazid et al. / NDT&E International 40 (2007) 594601

analysis. This is depicted in Figs. 1519. Every single step

in the process ow is essential especially before the
application of separability lter and CHT technique.
4. Conclusions

Fig. 16. Image before detection.

Fig. 17. Image as in Fig. 16, after detection.

The work could be considered as an evidence of the

applicability of the separability lter and CHT in the
detection of circular object in the radiography image.
A 100% of lm was found to have rst and second
correct circular detections by the proposed system.
Third and fourth detections also give higher lm percentage of successful detection with 90.9% and 81.82%,
respectively. This indicates that the system is suitable
for the use of circular detection in the welding area.
By having a proper image quality and contrast coupled
with Sobel edge detector, the circular object is easily
detected by separability lter and CHT technique. Additionally, this work shows good image contrast or
quality could avoid image enhancement in the preprocessing step of the pattern recognition ow procedure. Even
though the proposed system is efcient in circular
detection, care has to be taken in the edge detector stage
to avoid ambiguous object that could lead to false
detections. Further research will be focused on nding a
way to avoid the presence of ambiguous object and nding
a technique to accurately determine the size of the detected
circular object.
Acknowledgments
The authors wish to thank MINTS NDT group and
Khairiah Yazid for providing the image data and
Sheriffah Noor Khamseah for proofreading the paper.
The excellent support from MINT and KUKUM throughout the research is highly appreciated.
References

Fig. 18. Image before detection.

Fig. 19. Image as in Fig. 18, after detection.

[1] Edward GR. Inspection of welded joints. ASM handbook, welding,

brazing and soldering, vol. 6. Materials Park, OH: ASM International; 1993. p. 10818.
[2] Boerner H, Strecker H. Automated X-ray inspection of aluminium
castings. IEEE Trans Pattern Anal Mach Intell 1988;PAMI-10(1):
7991.
[3] Demandt K, Hansen LK. Real-time X-ray system with fully
automated defect detection and quality classication. In: Halmshaw
R, editor. X-ray real time radiography and image processing,
Proceedings of symposium, Newbury, Berkshire; 1988. p. 96119.
[4] Gayer A, Sayer A, Shiloh A. Automatic recognition of welding
defects in real time radiography. NDT Int 1990;23(3):1316.
[5] Builtjes JH, Rose P, Daum W. Automatic evaluation of weld
radiographs by digital image processing. In: Halmshaw R, editor.
X-ray real time radiography and image processing, Proceedings of
symposium, Newbury, Berkshire; 1988. p. 6372.
[6] Munro JJ, et al. Weld inspection by real time radioscopy. Mater Eval
1987;45(11):13039.
[7] American Society of Mechanical Engineers (ASME), Section 1; 1998,
p. 204.

ARTICLE IN PRESS
H. Yazid et al. / NDT&E International 40 (2007) 594601
[8] Yazid H, et al. Object detection using circular Hough transform. Am
J Appl Sci 2005;2(12):16069.
[9] DOrazio T, Guaragnella C, Leo M, Distante A. A new algorithm for
ball recognition using circle Hough transform and neural classier.
Pattern Recognition 2004;37:393408.
[10] Kawaguchi T, Hidaka D, Rizon M. In: Proceedings of IEEE ICRP,
vol. 1; 2000, p. 4952.

601

[11] Fukui K, Yamaguchi O. Facial feature points extraction method

based on combination of shape extraction and pattern matching.
Trans IEICE Jpn J80-D-II 1997;8:21707.
[12] Duda RO, Hart PE. Use of the Hough transformation to detect lines
and curves in picture. Commun ACM 1972;115.
[13] Kimme C, Ballard D, Sklansky J. Finding circles by an array of
accumulators. Commun ACM 1975;18(2):1202.