Computer Vision System for Weld Bead Analysis
Luciane Baldassari Soares, Atila Astor Weis, Bruna de Vargas Guterres, Ricardo Nagel Rodrigues
and Silvia Silva da Costa Botelho
Science Computer Center - C3, Federal University of Rio Grande - FURG, Brazil
Keywords:
Computer Vision, Discontinuities, Welding.
Abstract:
Welding processes are very important in different industries and requires precision and attention in the steps
that will be performed. This article proposes the use of an autonomous weld bead geometric analysis system
in order to verify the presence of geometric failures that may compromise the weld integrity. Using an vision
system attached to a linear welding robot, images of pre-welded and post-welded metal plates are captured and
compared and metrics are applied for evaluation. The proposed method uses Hidden Markov Model (HMM)
to identify the weld bead edges and calculate several evaluation metrics to detect geometric failures such as
misalignment, lack or excess of fusion, among others.
1 INTRODUCTION
The Welding process consists of a fundamental
practice in several areas such as automotive industry,
metallurgy, civil construction, naval industry and rai-
lway. The welding process is considered laborius due
to the risks the welder is exposed to, such as sparks,
toxic gases, radiation emitted by the arc and electri-
city. For this reason, technological advances in this
area not only provide productivity increase and pro-
cess enhancement, but also assure the welder secu-
rity. Besides the mentioned risks, the welder must
adjust the parameters, guide the torch and control the
weld bead quality (Broering et al., 2005). As a con-
sequence, the weld quality and the productivity may
be compromised since the operator can become over-
worked.
Welding processes represent a very significant
part of the manufacturing costs of a product. The-
refore a weld bead fault not identified during the qua-
lity inspection process can have major consequences
for the equipment manufacturer (Andreucci, 2013).
Due to all above mentioned factors, the use of hu-
man operated welding robots is increasingly common.
Additionally, these robots reduce the rework and pro-
vide a more efficient production line. In the naval in-
dustry, for example, it is indispensable to use more
flexible solutions since it is necessary to weld metal
sheets for different sized ships.
In shipyards, an alternative to automate the wel-
ding process is the Bug-O Matic Weaver robot, shown
Figure 1: Bug-O Matic Weaver Robot - Source: Author.
in Figure 1. The system is remotely operated, decrea-
sing the welder contact with the risk area. The robot,
positioned on a rail attached to any linear surface, car-
ries the torch during the welding process. It is posible
to set different parameters such as weaving regime
(arm oscilatory moviment) and robot and arm linear
velocity.
It is worth noting the predominant welding techni-
que used in shipyards is the FCAW (Flux Cored Arc
Welding). The FCAW has several advantages being
widely used in naval industry which aims to increse
the productivity keeping the quality of the processes
(Machado, 2015).
The automation of the welding process makes it
not only more stable, but also more uniform, thus re-
ducing the risks of discontinuities and avoiding the
weld bead discard (Carvalho et al., 2009). Further-
more, to manually achieve this quality level it would
402
Soares, L., Weis, A., Guterres, B., Rodrigues, R. and Botelho, S.
Computer Vision System for Weld Bead Analysis.
In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 4: VISAPP, pages
402-409
ISBN: 978-989-758-290-5
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
be necessary a welder with years of experience.
Nowadays, the weld bead inspection is performed
not only by Non-Destructive Testing (NDT) inclu-
ding visual methods, penetrating liquid, magnetic par-
ticles, radiography and ultrasound, but also by De-
structive Evaluation such as tensile, impact, folding,
fatigue and hardness tests (Lima et al., 2005). Accor-
ding to (Andreucci, 2013), the solution to assist the
quality control of equipments and materials is related
to NDT techniques. As reported by the Brasilian As-
sociation of Non-Destructive Testing and Inspection
(ABENDI), the NDT techniques are one of the main
quality control tools, they increase the weld reliability
and reduce the process cost (Abendi, 2014).
It is important for the realized tests the identifica-
tion of discontinuities such as incomplete or excessive
fusion, welding bite, porosity, incomplete or exces-
sive penetration, cracks and geometrical weld bead
irregularities that may compromise the weld quality.
Although it is possible to extract some relevant infor-
mation about the weld bead geometry, we still do not
have a tool able to identify all presented discontinui-
ties.
The computer vision may be the solution to
quickly and precisely identify some of the above dis-
continuities. In conclusion, computational techniques
are studied and applied in order to stablish an auto-
matic and ideal solution for the inspection and de-
tection of discontinuities on the weld seam (Felis-
berto, 2010).
2 LITERATURE REVIEW
In the works (Cook et al., 1995; Hou and Liu, 2012;
Kumar et al., 2012; Kumar et al., 2014) it is pre-
sented methods offline for welding quality inspection
and evaluation using digital cameras and lighting sy-
stem. More classic approaches are carried out in
(Cook et al., 1995) and (Hou and Liu, 2012). It is
assumed that the weld width (which can be identi-
fied through the horizontal integration) and the spa-
cing between the undulations (which can be obtained
through the peak-to-peak distance) can not oscillate
for good weld quality. Moreover Hou et al. makes
use of the Canny algorithm for identifying the edges
of the weld bead. After the edges are identified, they
are plotted in 2D histogram, in order to compare them
with histograms of images of good quality weld be-
ads. Through a pre-established threshold the weld is
classified good or not.
In the article of Kumar et al. (Kumar et al.,
2012)(Kumar et al., 2014) an evaluation system for
MIG (Metal Inert Gas) welding in “V” groove is pro-
posed. Four frames are captured in sequence and then
the regions of interest of the image are segmented and
some characteristics based on the average intensity
of the pixels are extracted. In (Kumar et al., 2012),
the authors use a back-propagation neural network to
classify the weld in four categories: good, excessive,
insufficient and without weld. The system was tested
with 80 samples and shows 95% accuracy in the clas-
sification. Yet, in (Kumar et al., 2014) the same clas-
sifier was used but with improved feature extraction,
based on a chi-square test to verify compliance with a
Gaussian distribution of pixels. This increased preci-
sion to 96.25%.
In the articles on (Schreiber et al., 2009) and (Sun
et al., 2005) the visual inspection methods are applied
online, that is, they are performed during the welding
process. In these methods the camera is usually pla-
ced next to the torch, obtaining an image of the re-
cently welded bead. The biggest obstacle to online
processing is the acquisition of images due to condi-
tions in the welding environment such as arc bright-
ness, smoke, sparks, welding spatters and reflections.
The works of Schreiber et al. and Sun et al. use
training techniques and pattern recognition, so that
the algorithm learns the good and bad characteristics
of a weld bead. Schreiber et al. presents a system
that consists of two distinct phases: training and ve-
rification. In the training phase the system learns the
quality criterion required from the training performed
through images of weld beads considered to be well
welded, examined by a welding engineer. This step
should be performed only once before using the ro-
bot and consists of extracting measurements from the
images of the training weld beads. These measures
are combined in order to obtain reference values and
tolerance ranges.
The verification step occurs by comparing each
frame obtained from the image of the welding being
performed, with the reference frames obtained from
the weld beads of training. In this process the evalu-
ated characteristics are the position and the width of
the cord, as well as the distribution of light around the
arc.
Sun et al. proposes a method of visual inspection
using a Fuzzy Pattern Recognition algorithm. The
method can detect defects such as porosity, lack of
penetration and lack of material in the weld bead. If
the defect exceeds the established minimum standard,
the system will automatically send an alarm and mark
the position where defects appear to help workers re-
pair them. The system is efficient since the speed of
detection can reach 3-4 frames per second. It was re-
ported that the method identified 65 of the 66 defects
included in the test set.
Computer Vision System for Weld Bead Analysis
403
As described in the works above, the acquisition
and processing of images of the online weld bead be-
comes much more difficult due to environmental con-
ditions. Because of this, we often end up missing
important information from the weld bead image, so
in this research we opted for offline analysis. Bet-
ween the techniques studied for weld bead discon-
tinuity analysis, the (Schreiber et al., 2009) method
and (Sun et al., 2005) method had a great contribu-
tion to our research since we followed the same trai-
ning and verification behavior as we can see in the
next section.
3 METHODOLOGY
The edge analysis is based on the the empty groove
characteristics. Through the proposed method it is
possible to detect the discontinuities shown on Figure
2 wich are misalignments along the weld bead, angu-
lar misalignments, excessive and insufficient material
deposition.
Firstly, images were taken from several weld seam
which qualities vary between acceptable or not. They
were taken using a camera with an passive illumina-
tion system (green leds) which is perpendicularly at-
tached to the proof body. The distance between them
is 15 cm providing a top view of the interest area.
In conclusion, a computer visual system is propo-
sed in order not only to highlight the weld bead edges,
but also to analyze its geometry and detect some dis-
continuities. Furthermore, through a combination of
existent techniques such as PCA (Principal Compo-
nent Analysis) to assist the data reduction, Gaussian
mixture as a probabilistic method and HMM (Hidden
Markov Model) as a statistical model, a edge detection
system is composed. The method execution is illus-
trated on Figure 3. In addition, the MATLAB (Matrix
Laboratory) software was used to develop the system
due to its facility for parameters testing.
The method can be divided on two stages: trai-
ning and testing. In the training stage several weld
bead images are presented in order to make the sy-
stem learn to identify not only good quality edges, but
also discontinuities. The testing stage consists on the
pos-weld verification on wich the weld bead edges are
highlighted and the metrics, present in the section 4,
are applied. Through this stage it is possible to assist
not only the welder, but also the inspector to identify
discontinuities that may cause the weld bead invalida-
tion.
3.1 Training
This stage encompasses the image capture of different
weld processes performed by the Bug-O Matic Wea-
ver robot with not only acceptable quality but also dis-
continuities.The obtained images are used as ground-
thruth and its edges are manually marked.
At a first moment, the edges of three weld bead
images of 3288 pixels height by 4608 pixels wide
were highlighted. Thus we have a total of 19728
annotated lines since each image has a 3288 pixels
height and two annotated edges. Therefore, each
image provides 6576 samples and 19728 edge points
samples, called edges profile, since each annotated
point is located on a single line.
For each point along the edge an edge profile is
extracted consisting on β pixels on the right and β
pixels on the left of the annoted edge. The edge pro-
file represents each pixel intensity and its concept is
exemplified on Figure 4. After extraction of all edge
profiles, there is a set of vectors that will be used to
learn the edge model. Thus, β = 80 is used.
Applying the PCA technique in the set of edge
profiles possible overlaps are reduced thus obtaining
a reduction in the data without significant loss of in-
formation. It is used 85 % of the total variance of
the analyzed components resulting on a PCA size of
nineteen.
The PCA space edge profiles projections are used
to train a gaussian mixture. The equation 1 represents
the probability of a given projection being an edge.
P(x|θ) =
k
i=1
w
i
g(x|µ
i
, Σ
i
), (1)
Where w
i
= 1, 2,...,k are the k components wheig-
hts, µ
i
represents an D-dimensional mean vector and
Σ
i
is a covariance matrix.
The Gaussian mixture parameters are learned by a
Maximum Likelihood Algorithm (Expectation Max-
imization - EM) which is a parameter estimation
technique that allow to deal with missing data. Furt-
hermore, the number of Gaussians k was set to seven
since the BIC (Bayesian Information Criterion) pre-
sents a graphical stable behavior at this value when a
1-30 range is analyzed.
It is emphasized that these images are used only in
the training, the test stage makes use of new images
of other weld processes performed by the same robot.
3.2 Tests
After training the proposed method, the testing stage
is performed. To evaluate the proposed method, the
detection of the weld beads edges are oerformed in
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
404
Figure 2: Weld beads geometry.
Application PCA
Application HMM
Extraction Edge ProfilesPictures
Images with the
highlighted edges
Gaussian Mixture
Figure 3: The proposed method block diagram.
β=80 β=80
Figure 4: Right edge image profile graphically represented.
different images that were not used to train the pro-
posed method. For this purpose, the HMM techni-
que is used since it is a statistical model on which the
modeling system is a Markov process with unknow
parameters. Furthermore, it provides the hidden pa-
rameters knowledge based on the observable ones.
The model extracted parameters can be used to rea-
lize new analysis such as pattern recognition (Anzai,
2012) which is the main purpose of its implementa-
tion.
For each given image, a region of interest (ROI)
is determined cosidering 250 pixels for each side of
the empty groove position. It assists to calculate the
probability of a pixel being an edge or not.
Furthermore, the edge position along the image li-
nes follows a Markov model. It means that given the
Computer Vision System for Weld Bead Analysis
405
edge position at line t 1 the probability of its posi-
tion at line t, p(y
t
|y
t1
) is obtained. The transition
models are defined by the equations:
p(y
t
= y
t1
) = 0, 9 (2)
p(y
t
= y
t1
1) = 0, 05 (3)
p(y
t
= y
t1
+ 1) = 0, 05 (4)
In conclusion, the appearance model p(x|y) con-
sists in estimating the probability of the variable y
being an edge x. Thus, based on the observable data
(edge profile of the weld bead images defined as x) the
hidden variables (edge position along the weld bead)
are estimated. It means the HMM is responsable for
inferring the edge profile location on the new weld
bead image.
By means of Viterbi algorithm, edge position se-
quence estimation is realized. Moreover, it is neces-
sary to identify the next states on a excellent way jus-
tifying the implementation of the Viterbi algorithm
since it is a optimal search algorithm as can be seen
on Equation 5.
y
= arg
y
maxp(y|x) (5)
After this stage, tests are performed on new ima-
ges from the testing set. The region of interest is se-
lected based on the empty groove position. It allows
to analyze smaller areas of the image enhancing the
time aspect of its processing. Finally, the probability
of each pixel within the ROI being an edge is analy-
zed.
4 EVALUATION METRICS
The proposed metrics aim to identify some discon-
tinuities shown on Figure 2 such as lack or excess
of weld deposition and misalignments along the weld
bead. The applied metrics are: minimum, maximum
and average distance between the edges and the empty
groove position, weld bead standard deviation along
the edge using the sliding window method and an-
gle between the edge and the empty groove using the
same method. It is woth mentioning the sliding win-
dow method consists on analyzing adjustable sized
portions of the ROI (window) along its height.
Through the method developed by (Steffens et al.,
2016) the empty groove lines are captured for later
comparison with the filled one. The mentioned aut-
hor uses the LSD (Line Segment Detector)algorithm
to correctly find the empty groove lines since it shows
major performance to other methods in detecting li-
nes in metal sheets. Furthermore, the proposed al-
gorithm realizes the lines filtering only keeping the
vertical ones. The weld parameters, which encompas-
ses electric current and voltage, wire feed rate, robot
linear velocity and oscillation type and velocity, are
automatically determined after extracting the empty
groove lines. Hereafter, the FCAW welding process is
initiated by the Bug-O Matic Weaver robot requiring
four steps for the groove complete filling.
After the welding process is finished, the weld
bead is imaged and, finally, the proposed method pro-
vides the weld bead edges detection. Soon after, geo-
metric evaluation visual metrics are implemented for
discontinuities verification.
4.1 Minimum Value
The first implemented metric, as can be seen on Equa-
tion 6, consists on the minimum distance between the
weld bead edge and the empty groove line for both
sides of the weld bead. Through this metric it is pos-
sible to identify regions with lack of material deposi-
tion.
M
1
= min |Y
i
C
i
| (6)
Where i = 1...n represents the analyzed image
line, Y
i
refers to the edge position on the image and
C
i
is the groove position.
4.2 Maximum Value
The considered metric consists on the maximum dis-
tance between the weld bead edge and the empty
groove line for both sides of the weld bead as shown
in Equation 7. This metric allows the identification of
regions with excessive material deposition.
M
2
= max|Y
i
C
i
| (7)
Where i = 1...n represents the analyzed image
line, Y
i
refers to the edge position on the image and
C
i
is the groove position.
4.3 Average Value
This metric consists on the average distance between
the weld bead edge and the empty groove line for both
sides of the weld bead as shown in Equation 8. This
metric assists the analysis of the weld bead regularity
since a pattern without considerable oscillations is de-
sired.
M
3
=
n
i=1
|Y
i
C
i
|
n
(8)
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
406
Where i = 1...n represents the analyzed image
line, Y
i
refers to the edge position on the image and
C
i
is the groove position.
4.4 Standard Deviation using sliding
window
This metric consists on calculationg the standard de-
viation using the sliding window method with size w
as shown in equation 9. Its application assists the
identification of misalignments points along the weld
bead edge.
M
4
(k) =
v
u
u
t
k+w1
i=k
(Y
i
C
i
k+w1
j=k
|Y
j
C
j
|
t
)
2
w
(9)
Where n represents the image lines quantity, w is
the window size, k = {1...n w}, Y
i
refers to the edge
position on the image and C
i
is the groove position.
4.5 Angle between Empty Groove and
Weld Bead Edge using sliding
window
Aiming to obtain the angle between the empty groove
and the edge for both sides the weld bead, it is used
a linear regression method. This way, both the linear
and angular coefficients of the groove and the edge
approximate lines are obtained.Thus, the Equation 10
allows to calculate the desired angle. Through this
metric it is possible to identify if the weld bead exhi-
bits angular misalignments with respect to the empty
groove.
M
5
= arctan|
a
c
a
y
1 + a
c
.a
y
| (10)
Where a
c
is the angular coefficient of the empty
groove approximated line and a
y
is the angular coef-
ficient of the weld bead approximated line both with
respect to the analized window.
5 RESULTS
This section is divided in two parts. Firstly, the weld
bead edge identification method results are presented.
Secondly, the results of the evaluation metrics appli-
cation, as described on section 4, are discussed.
5.1 Proposed Method
In this section some images with highlighted edges
are presented. It is worth mentioning the highlighted
edges were identified by the proposed method. Thus,
at the left side of Figure 5, the original weld bead
image is shown. At its center, the manually highlig-
hted edges and, finally, at the right side of the image,
the results of the proposed metod are shown.
a) b) c)
Figure 5: a) Weld bead Image. b) Manually highlighted
edges. c) Proposed method highlighted edges.
Based on Figure 5 the mean error between the ma-
nually highlighted edges and the edges highlighted by
the method is calculated. Thereby, the resultant ab-
solute mean error is 2,9675 pixels for the right edge
and 6,0727 pixels for the left one, as can be seen in
the graphs shown in the Figures 6esquerda. Consi-
dering, based on the camera calibration, that 1 milli-
meter (mm) equals 46 pixels the resulting errors are
0,0645 mm and 0,1320 mm for the right and the left
edges respectively. It is notable the diference between
the ground truth and the proposed method result is mi-
nimal.
Figure 6: Graph of the difference between the manually an-
notated right edge and those highlighted by the method.
Computer Vision System for Weld Bead Analysis
407
Figure 7: Graph of the difference between the manually an-
notated left edge and those highlighted by the method.
Lastly, the table 1 demonstrates the resultant ab-
solute mean error in pixels (px) for the right and left
sides of all tested grooves.
Table 1: Absolute average error between the manually an-
notated border and the border found by the method.
Welding Seams Right Edge Left Edge
(px) (px)
1 2,9675 6,0727
2 5,1442 6,0073
3 6,4234 5,9605
4 4,4349 6,7628
5 3,5703 6,2743
6 5,8406 6,5693
7 8,9605 10,9635
8 4,5493 6,1439
9 8,2597 7,0973
5.2 Evaluation Metrics Results
In this section some evaluation metrics results are de-
monstrated. The results presented below were imple-
mented using the image shown on Figure 5.
The first two metrics presented, minimum and
maximum values, are able to identify points on which
insufficient or excessive fusion occurs. According
to American National Standards (ANS) (SOCIETY,
2001), there is a major worry about lack of fusion,
since if the groove is not fully filled, the weld can be
invalidated or may need another weld pass.
On Figure 8 it is possible to visualize the right
weld bead edge on which there is a possible lack of
fusion. It indicates the weld bead edge (highlighted in
red) is too near of the empty groove line (represented
by a blue line). It should be noted the parameters of
minimum and maximum values may be modified by
the weld inspector since the mentioned norm only
stablishes the groove needs to be fully filled not spe-
cifying any minimum or maximum value.
Figure 8: Results of the Minimum value metric.
Through the metric that uses the standard devia-
tion with sliding window it is possible to graphically
identify, through the peaks plotted on Figure 9, where
a weld bead discontinuity may happen.
Figure 9: Graphic of the left and right weld bead edges with
the standard deviation metric using sliding window.
Using the metric of the angle between the empty
groove and the weld bead edge using sliding win-
dow, it is possible to identify if the weld bead exhibits
an angular misalignment with respect to the empty
groove. Furthermore, it is possible to visually iden-
tify the regions on which there is a greater disparity
in angle. On Figure 10 it is possible to see not only
these regions, but also the graph of the mentioned an-
gle.
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
408
Figure 10: Graphic of the angle between the empty groove
and the weld bead edge using sliding window for the left
and right sides of the weld bead.
6 CONCLUSIONS
The gains provided by the welding process automa-
tion in the industry are plenty since it is a insalubri-
ous activity for the welder who stays in contact with
sparks, smoke and eletricity. Therefore, it is increa-
singly common to use robots in this type of task since
it guarantees the final process quality, increases pro-
ductivity and reduces the human intervention.
Through a computer visual system attached to the
linear welding robot, the presented work proposes
a weld bead edge identification method and metrics
to identify some discontinuities based on the weld
bead geometry. The proposed method successfully
identified the weld bead edges with maximum errors
of 10.96 pixels. Furthermore, through the perfor-
med tests it is possible to conclude the proposed me-
trics contribute to identify geometric failures that may
compromise the welding process.
REFERENCES
Abendi (2014). Guia end de inspec¸
˜
oes. http://
www.abendi.org.br/abendi/.
Andreucci, R. (2013). Aspectos industriais: Protec¸
˜
ao ra-
diol
´
ogica. Associac¸
˜
ao Brasileira de Ensaios N
˜
ao De-
strutivos (ABENDI).
Anzai, Y. (2012). Pattern recognition and machine lear-
ning. Elsevier.
Broering, C. E. et al. (2005). Desenvolvimento de sistemas
para automac¸
˜
ao da soldagem e corte t
´
ermico.
Carvalho, R. S. et al. (2009). Rob
ˆ
o cnc para a automac¸
˜
ao
da soldagem mig/mag em posic¸
˜
oes e situac¸
˜
oes de ex-
trema dificuldade.
Cook, G., Barnett, R., Andersen, K., Springfield, J., and
Strauss, A. (1995). Automated visual inspection and
interpretation system for weld quality evaluation. In-
dustry Applications Conference, 1995. Thirtieth IAS
Annual Meeting, IAS’95., Conference Record of the
1995 IEEE, 2:1809–1816.
Felisberto, M. K. (2010). T
´
ecnicas autom
´
aticas para
detecc¸
˜
ao de cord
˜
oes de solda e defeitos de soldagem
em imagens radiogr
´
aficas industriais.
Hou, X. and Liu, H. (2012). Welding image edge detection
and identification research based on canny operator.
In Computer Science & Service System (CSSS), 2012
International Conference on, pages 250–253. IEEE.
Kumar, G. S., Natarajan, U., and Ananthan, S. (2012). Vi-
sion inspection system for the identification and clas-
sification of defects in mig welding joints. The Inter-
national Journal of Advanced Manufacturing Techno-
logy, 61(9-12):923–933.
Kumar, G. S., Natarajan, U., Veerarajan, T., and Ananthan,
S. (2014). Quality level assessment for imperfections
in gmaw. Welding Journal, 93(3):85–97.
Lima, E., Castro, C. A. C., Bracarense, A. Q., and Campo,
M. (2005). Determinac¸
˜
ao da relac¸
˜
ao entre par
ˆ
ametros
de soldagem, largura da poc¸a e aspecto do cord
˜
ao de
solda utilizando c
ˆ
amera de alta velocidade. soldagem
Inspection-Out/Dez, pages 182–189.
Machado, E. (2015). Influ
ˆ
encia dos ventos sobre a quali-
dade das soldas realizadas em estaleiros pelo processo
arame tubular.
Schreiber, D., Cambrini, L., Biber, J., and Sardy, B.
(2009). Online visual quality inspection for weld se-
ams. The International Journal of Advanced Manu-
facturing Technology, 42(5-6):497–504.
SOCIETY, A. W. (2001). American welding society. aws
3.0: Standard welding terms and definition.
Steffens, C. R., Leonardo, B. Q., d. S. Filho, S. C., H
¨
utner,
V., d. Rosa, V. S., and d. C. Botelho, S. S. (2016). Wel-
ding groove mapping implementation and evalua-
tion of image processing algorithms on shiny surfaces.
In VISAPP 2016 11th Joint Conference on Computer
Vision, Imaging and Computer Graphics Theory and
Applications, volume 3, pages 265–272.
Sun, Y., Bai, P., Sun, H.-y., and Zhou, P. (2005). Real-time
automatic detection of weld defects in steel pipe. NDT
& E International, 38(7):522–528.
Computer Vision System for Weld Bead Analysis
409