Available online at www.sciencedirect.com
Available online at www.sciencedirect.com
ScienceDirect
ScienceDirect
Available
online
atonline
www.sciencedirect.com
Procedia
CIRP
00 (2019)
000–000
Available
at www.sciencedirect.com
ScienceDirect
ScienceDirect
Procedia CIRP 00 (2019) 000–000
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Procedia CIRP
00 (2017)
000–000
Procedia
CIRP 81
(2019) 771–774
52nd CIRP Conference on Manufacturing Systems
52nd CIRP Conference on Manufacturing Systems
www.elsevier.com/locate/procedia
A visual inspection system for KTL coatings
CIRP inspection
Design Conference,
May for
2018,KTL
Nantes,coatings
France
A28th
visual
system
a*
b
Drago Bračun and Igor Lekše
a*
b
a
Dragoof Bračun
and Igor Lekše
methodology
analyze
the functional
and
physical
architecture
University ofto
Ljubljana,
Faculty
Mechanical
Engineering,
Aškerčeva
cesta
6, SI-1000 Ljubljana
A new
of
TPV d.o.o., Kandijska cesta 60, SI-8000 Novo mesto
University
of Ljubljana,
Faculty of Mechanical
Engineering, Aškerčeva
cesta 6,family
SI-1000 Ljubljana
existing products
for
an assembly
oriented
product
identification
* Corresponding author. Tel.: +386-1-4771 747; fax: + TPV
386-1-2518
567. E-mail
address:
drago.bracun@fs.uni-lj.si
d.o.o., Kandijska
cesta
60, SI-8000
Novo mesto
a
b
b
* Corresponding author. Tel.: +386-1-4771 747; fax: + 386-1-2518 567. E-mail address: drago.bracun@fs.uni-lj.si
Abstract
Abstract
Paul Stief *, Jean-Yves Dantan, Alain Etienne, Ali Siadat
École Nationale Supérieure d’Arts et Métiers, Arts et Métiers ParisTech, LCFC EA 4495, 4 Rue Augustin Fresnel, Metz 57078, France
Many metal components in automotive industry are surface-protected using the electrophoretic cathode metal coating (KTL). As
defects
can Tel.:
occur
the components
need to be 100%
inspected,
especially ifcathode
the manufacturers
are (KTL).
committed
*surface
Corresponding
author.
+33in
87
37process,
54 30; E-mail
address:
Many
metal
components
in3 this
automotive
industry
arepaul.stief@ensam.eu
surface-protected
using
the electrophoretic
metal coating
As
to delivering
zero-defect
Due to the
complicated
3D need
curved
andinspected,
black shiny
colour ofifthe
inspection
typically
surface
defects
can occurproducts.
in this process,
components
to shapes
be 100%
especially
thecoating,
manufacturers
are is
committed
carried
out
manually.
To
avoid
labour-intensive
manual
inspection
in
serial
production,
an
automated
visual
inspection
system
is
to delivering zero-defect products. Due to complicated 3D curved shapes and black shiny colour of the coating, inspection is typically
developed.
The
paper
presents
the
implementation
of
a
control
device
for
mass
inspection
of
the
products,
typical
KLT
coating
carried out manually. To avoid labour-intensive manual inspection in serial production, an automated visual inspection system is
Abstract
defects, andThe
the visual
method
for the detection
of coating
defects.
A difference
image
acquisition
method
implemented
developed.
paper inspection
presents the
implementation
of a control
device
for mass
inspection
of the
products,
typicalis KLT
coating
in
order
to
reveal
the
defects
on
the
black
shiny
colored
parts.
An
image
processing
algorithm
for
the
recognition
of
light
patterns
defects,
and
the
visual
inspection
method
for
the
detection
of
coating
defects.
A
difference
image
acquisition
method
is
implemented
Inand
today’s
business environment,
the trend
towardsofmore
product
variety
and customization is unbroken. Due to this development, the need of
a clustering
for the
recognition
are
presented.
in order
to revealapproach
the defects
on systems
the
black
shinydefects
colored
parts.
An image
processing
algorithm
recognition
of light
patterns
agile
and reconfigurable
production
emerged
to
cope with
various
products
and product
families.for
Tothe
design
and optimize
production
and a clustering
recognition
of defectsproduct
are presented.
systems
as well as approach
to choose for
the the
optimal
product matches,
analysis methods are needed. Indeed, most of the known methods aim to
© 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
analyze
product
or one
product family
on the
physical level. Different product families, however, may differ largely in terms of the number and
© 2019aThe
Authors.
Published
by Elsevier
Ltd.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
©
2019
The
Authors.
Published
by
Elsevier
Ltd.
This iscomparison
an open(http://creativecommons.org/licenses/by-nc-nd/3.0/)
access
under
the CC BY-NC-ND
license
nature
of
components.
This
fact
impedes
an
efficient
and article
choice
of appropriate
product family
combinations for the production
This
is
an
open
access
article
under
the
CC
BY-NC-ND
license
Peer-review under responsibility of the scientific committee
of the 52nd
CIRP Conference
on Manufacturing
Systems.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
system.
A newunder
methodology
is proposed
to analyze
existing products
in view
their functional
and physicalSystems.
architecture. The aim is to cluster
Peer-review
responsibility
of the scientific
committee
of the 52nd
CIRPof
Conference
on Manufacturing
Peer-review
under
responsibility
of the scientific
committee
of the
52nd CIRPofConference
on Manufacturing Systems.
these
products
in surface
new
assembly
oriented
product
families
for the
optimization
Keywords:
KTL;
defect; inspection;
imaging system;
pattern
recognition;
cluster existing assembly lines and the creation of future reconfigurable
assembly systems. Based on Datum Flow Chain, the physical structure of the products is analyzed. Functional subassemblies are identified, and
surface
inspection;
imagingasystem;
recognition;
cluster architecture graph (HyFPAG) is the output which depicts the
a Keywords:
functionalKTL;
analysis
is defect;
performed.
Moreover,
hybridpattern
functional
and physical
similarity between product families by providing design support to both, production system planners and product designers. An illustrative
coating,caseinspection
typically
carried
outcolumns
manually.
example of a nail-clipper is used to explain the proposed methodology. An industrial
study on twoisproduct
families
of steering
of
1. Introduction
Regardless
of
the
staff
experience,
manual
inspection
is
thyssenkrupp
Presta France is then carried out to give a first industrial evaluation
of
the
proposed
approach.
coating, inspection is typically carried out manually.
© 2017 The Authors. Published by Elsevier B.V.
subjective
and
time-consuming
in
large-series
production.
1. Introduction
Regardless of the staff experience, manual inspection is
Peer-review
under responsibility
the scientific
committee
of the 28th
2018. shows several specific solutions for
Electrophoretic
cathode of
metal
coating (KTL
or E-coat)
[1] isCIRP Design
The Conference
literature review
subjective and time-consuming in large-series production.
surface
defect detection.
system
described
[2] acquires
The literature
review The
shows
several
specificinsolutions
for
multiple
images
under
different
lighting
conditions,
then
surface defect detection. The system described in [2] and
acquires
these images
are processed
separately
and merged
into and
one. then
The
multiple
images
under different
lighting
conditions,
features
extracted
from
this
fused
image
and
the
former
these images are processed separately and merged into one. The
processing
steps
the feature
space for
supervised
of features
the product
rangeestablish
and
manufactured
and/or
extracted
fromcharacteristics
this
fused image
and a the
former
learning
classifier
based
on
artificial
neural
networks.
The
assembled
in this
system.
In this
main
in
processing
steps
establish
thecontext,
feature the
space
forchallenge
a supervised
system
presented
in
[3]
detects
defects
catalogued
as
dings
and
modelling
analysis
is now
only to
copenetworks.
with single
learning and
classifier
based
on not
artificial
neural
The
dents
on
car
body
surfaces.
It
sweeps
the
object
with
the
light
products,
limited product
range defects
or existing
productasfamilies,
system apresented
in [3] detects
catalogued
dings and
patterns,
and
uses
fusion
on
flow
butdents
also to
to image
analyze
and
tobased
compare
products
to to
define
onbe
carable
body
surfaces.
It sweeps
theoptical
object
with
theobtain
light
a
resulting
fused
image
which
holds
the
information
of all
new
productand
families.
It canfusion
be observed
that
classical
patterns,
uses image
based on
optical
flowexisting
to obtain
suffered
by the
product
families
are regrouped
in projected
function
clients
or during
features.
avariations
resulting
fused
image
which
holds of
thepatterns
information
of the
all
sweeping
process,
indicating
the
presence
of
anomalies.
A
However,
assembly
oriented
product
familiespatterns
are hardly
to find.the
variations
suffered
by the
projected
during
deflectometry
method
[4]
features
a
high
resolution
camera
and
On the product
family
level, products
differ of
mainly
in two A
sweeping
process,
indicating
the presence
anomalies.
a dedicated
illumination
system of
on displaying
main
characteristics:
(i) the
components
and
(ii) fringe
theand
deflectometry
method
[4] number
features
abased
high
resolution
camera
in a monitor,
the based
detection
irregularities
in
type
of
components
(e.g.allowing
mechanical,
electrical,
electronical).
apatterns
dedicated
illumination
system
on of
displaying
fringe
surfaces.
An
overview
of
three
different
image
acquisition
patterns
a monitor, allowing
the detection
irregularities
Classicalinmethodologies
considering
mainly of
single
products in
Because
of complicated
shapes
and black shiny
color [1].
of the or surfaces.
overview
of three
different
image
acquisition
product
varieties
(high-volume
to low-volume
production)
solitary, An
already
existing
product
families
analyze
the
2212-8271 © 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
To
cope
with
this
augmenting
variety
as
well
as
to
be
able
to
product
structure
on
a
physical
level
(components
level)
which
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
2212-8271
©under
2019responsibility
The optimization
Authors. of
Published
by Elsevier
Ltd.
an
open
access
article under
the CC BY-NC-ND
license an efficient definition and
identify
potentials
in This
existing
causes
regarding
Peer-reviewpossible
the scientific
committee
ofthe
theis52nd
CIRP
Conference
on difficulties
Manufacturing
Systems.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
production
system, it is important to have a precise knowledge
comparison of different product families. Addressing this
an Electrophoretic
important technological
process
for painting
protecting
cathode
metal
coating
(KTL orand
E-coat)
[1] is
Keywords:
Assembly; Design
method;
Family
identification
the
parts
from
environmental
impacts.
The
quality
of
coating is
an important technological process for painting and protecting
achieved
by managing
the coating
process
and byofinspecting
the parts from
environmental
impacts.
The quality
coating is
parts
after
coating.
There
are
a
number
of
different
coating
achieved by managing the coating process and
by inspecting
defects,
including:
craters
(bowl-shaped
depressions),
1. Introduction
parts after coating. There are a number of different coating
roughness including:
(patches on acraters
cured film
that exhibits an
alternately
defects,
(bowl-shaped
depressions),
non-uniform
and
smooth
appearance),
redissolution
(part
ofofthe
Due
to (patches
the faston development
in exhibits
the domain
roughness
a cured film that
an alternately
coat
film
washes
off
or
dissolves),
dirt
(three
sources:
process,
communication
andsmooth
an ongoing
trend redissolution
of digitization
non-uniform and
appearance),
(partand
of the
environmental
and
oven),
streaking
(duefacing
to pretreatment,
digitalization,
manufacturing
enterprises
important
coat
film washes
off or
dissolves),
dirt are
(three
sources:
process,
rinsing and
pinholing
(a pattern
relatively
small,
challenges
in racking),
today’s
market
environments:
a continuing
environmental
and oven),
streaking
(dueof to
pretreatment,
random
volcano-like
holes),
air
entrapment,
gloss
variations,
tendency
reduction
of product
development
times and
rinsingtowards
and racking),
pinholing
(a pattern
of relatively
small,
color variations,
thinholes),
coating,
corrosion
and
peel.
shortened
product
lifecycles.
In addition,
there is
an orange
increasing
random
volcano-like
air entrapment,
gloss
variations,
Example
of
defects
are
shown
in
Fig.
1.
Coating
defects
on
demand
customization,
being at corrosion
the same time
a global
color ofvariations,
thin coating,
and in
orange
peel.
products
should
be
identified
immediately
after
the
coating
competition
with
competitors
all over
the 1.world.
Thisdefects
trend, on
Example of
defects
are shown
in Fig.
Coating
process.
For
that be
purpose,
coated
parts
aremacro
visually
inspected.
which
is inducing
the identified
development
from
micro
products
should
immediately
after to
the
coating
Because
of complicated
shapes
black
shiny
color
of the
process.
For
that
coated
parts
are
inspected.
markets,
results
in purpose,
diminished
lot and
sizes
due visually
to augmenting
Peer-review under responsibility of the scientific committee of the 52nd CIRP Conference on Manufacturing Systems.
2212-8271 © 2019 The Authors. Published by Elsevier Ltd.
This is an©open
article Published
under theby
CC
BY-NC-ND
2212-8271
2017access
The Authors.
Elsevier
B.V. license (http://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review
under
responsibility
of scientific
the scientific
committee
theCIRP
52ndDesign
CIRPConference
Conference2018.
on Manufacturing Systems.
Peer-review
under
responsibility
of the
committee
of the of
28th
10.1016/j.procir.2019.03.192
772
2
Drago Bračun et al. / Procedia CIRP 81 (2019) 771–774
Drago Bračun, Igor Lekše / Procedia CIRP 00 (2019) 000–000
approaches for automatic visual inspection of metallic surfaces
is presented in [5]. The first method is gray-level intensity
imaging with bright and dark field lighting techniques.
Subsequently, two range imaging techniques are introduced
which may succeed in contrast to intensity imaging if the
reflection property across the intact surface changes. However,
range imaging for surface inspection is restricted to surface
defects with three-dimensional characteristics, e.g. cavities.
A surface defect detection system for mobile phone screen
glass [6] corrects the image misalignment and then the
combination of subtraction and projection is used to identify
defects. To segment the defects, an improved fuzzy c-means
cluster algorithm is developed.
The vision system for on-line surface inspection in an
aluminum casting process [7] implements similarity-based
algorithms as well as texture algorithms for defect detection in
a high-textured surface where, for most defects, segmentation
cannot be achieved solely by means of threshold techniques.
The system was applied to continuous cast aluminum
inspection, where up to fifteen different kinds of defects must
be detected and classified.
Fig. 2. Automated visual inspection system [8].
Fig. 3. Part transportation system with AGV-s.
3. Image processing
Fig. 1. Examples of coating defects: (a) coagulated paint; (b) air entrapment;
(c) roughness; (d) corrosion; (e) redissolution; (f) dirt.
2. Automated visual inspection system
The automated visual inspection system presented in Fig. 2
is integrated into the part transportation system from the KTL
process to the automated warehouse. The part transportation
system is carried out with AGV’s and industrial robots, as is
demonstrated in Fig. 3. The inspection is carried out employing
16 cameras and carefully designed illumination for each
camera. The inspection device has four nests, of which the first
and the last are used for robotic part handling, and the middle
two are intended for visual inspection. Part transportation
between the nests is carried out with two servo-drive based part
transportation systems. The control of the device is carried out
with the PLC and i7 based industrial PC for image processing.
The challenges which were encountered and are presented in
the continuation of this paper are detection of defects on
complicated 3D curved shapes with black shiny coating (see
Fig. 4) and the distinction between critical defects and visual
flaws.
Coating defects are usually visible as blobs of different
image brightness, as is demonstrated in Fig. 4. They are
typically brighter than their surroundings, varying in size from
a few pixels to larger parts of the image. They can also take the
form of a cluster of small blobs. In the image processing stage,
these light blobs are initially extracted from the image and then
classified either as actual defects or as visual flaws. Their
extraction is carried out first by employing threshold values to
the image to obtain a mask in form of binary image, with pixel
values 0 or 1. The value is 1 if the pixel brightness in the image
of interest is brighter than the threshold value, and 0 if it is
below. Blobs are identified as areas where the value of adjacent
pixels is 1. After being identified, their area i.e the number of
adjacent pixels with value 1, is calculated. The initial criterion
for the defect identification is the area of the blob as well as
other geometric properties, e.g. location and shape.
Fig. 4. Image A. A detail of a typical KTL-coated part with 3D form and black
shiny color. Arrows point to the coating damage, dust and fingerprint.
Drago Bračun et al. / Procedia CIRP 81 (2019) 771–774
Drago Bračun, Igor Lekše / Procedia CIRP 00 (2019) 000–000
However, blob extraction is not so straightforward since the
coated objects are curved in all directions. Feature recognition
in imaging systems typically starts with the illumination of the
object, which should be adjusted so as to reveal the features of
interest, i.e. bright or dark field illumination. At curved objects,
it is not possible to adjust the illumination to all object features,
and from some parts of the object a bit of light reflection will
occur directly in the camera while at others the light will reflect
away. These reflections resemble the defects in image, which
interferes with image processing.
One solution to the problem is to create a background image
B (see Fig. 5) that is subtracted from the image of interest A
(see Fig. 4). The background image represents the average of a
large number of images under the same lighting conditions and
the same image acquisition setup. These images are acquired
by inspecting the same type of product in many products
without defects (>20).
773
3
second image. There are three parameters (Tx , Ty) for
translation and angle  for rotation. The estimation of all
motion parameters can be time-consuming in case of many
large images. A simplification based on the knowledge of the
product clamping inside the inspection device can be carried
out. It was found that variability is most significant in
positioning the part inside the inspection device and that the
main motion is rotation, while the translation can be neglected.
In addition, since the point of rotation is known, it is only
necessary to identify the value of rotation. The search algorithm
is straightforward: (i) rotate the image A with respect to the
image B in sequence of small steps, (ii) in each step calculate
the difference image D and (iii) calculate the sum off all pixels
on the image D. The rotation angle with the minimum sum is
the best alignment angle.
Variability in dimensions among parts still remains and is
most noticeable in the image D at the edges of the product
where there is no overlap between the parts due to difference
in shape (see Fig. 6). These bright edges are removed
employing an additional mask that specifies the region of
interest. Further image processing continues as already
described at the beginning of this chapter, by thresholding, blob
search, calculation of blob properties and defect identification.
4. Defect identification
Fig. 5. The background image B. It is the average of a large number of images
(>20) under the same lighting conditions and the same image acquisition setup.
In many cases, defects do not take the form of big blobs but
rather appear as concentrated groups of small blobs. Fig.7
shows such an example. The left white circle marks a group of
large area blobs that are actually a surface defect, while the
right circle marks the group of small blobs that show dust on
the surface and are not a defect. Defect identification that
searches for a big blob does not recognize any of these blobs as
a defect. To solve this problem, the blobs are grouped into
clusters, and for each cluster the parameter quantifying the
likelihood of a defect is calculated.
Fig. 6. The difference image D is the image A with removed background B.
By subtracting the background image B from the image of
interest A and calculating the absolute value of the differences,
we eliminate brightness variations caused by the object form.
If we assume that images A, B and D are arrays of same size
and data type, with number of rows U and columns V, where
and indices i,j index pixel values from i = 0 to U, and j = 0 to
V. The difference image is calculated pixel wise as Di,j=abs(Ai,j
- Bi,j), where abs stands for absolute value (see Fig. 6). If there
are visible defects in image A, they are also present in
difference image D, which can be observed by comparing Fig.
4 and Fig. 6.
The weak point of this method is in the alignment of images.
The object of interest must be at the same position on both
images. This cannotbe perfectly achieved, because there are
two main sources of misalignment or variability: the first is
variability among parts, i.e. how repeatable their dimensions
are, and the second is in positioning of products inside the
inspection device. Both can be minimized to a certain extent,
but never completely eliminated. To deal with them, a software
alignment of images is carried out based on Euclidean motion,
where the first image is rotated and shifted with respect to the
Fig. 7. Defects are visible as groups of small blobs.
For grouping the blobs, a k-means clustering algorithm is
employed. The basic idea behind this clustering is to define
clusters such that the total intra-cluster variation is minimized.
The standard k-means clustering algorithm defines the total
within-cluster variation as the sum of squared Euclidean
distances (total within-cluster sum of squares TSS) between
items (blob centers) and the corresponding cluster centroid.
Let us assume that we arrive at k clusters. For each cluster
from k = 1 to K a total area of all blobs can be calculated as
N
Ak =  al ,
l =1
(1)
Drago Bračun et al. / Procedia CIRP 81 (2019) 771–774
Drago Bračun, Igor Lekše / Procedia CIRP 00 (2019) 000–000
774
4
where 𝑎𝑎𝑙𝑙 is the blob area and N is the number of blobs in
cluster k. Similarly, a cluster mean radius Rk is calculated as
N
Rk =
 (r  a )
l =1
l
Al
l
(2)
,
where rl is the distance from the cluster centroid to the center
of l-th blob. The parameter E which we use to quantify the
likelihood of a defect can be now defined as
Ek =
Ak
.
Rk
(3)
The likelihood of a defect is high when the total area Ak of
blobs in cluster is high (big blobs), and when they are
concentrated within small radius Rk. This is demonstrated on
the left cluster in Fig.7, where E amounts to 1.12. On the right
cluster, with dust on the surface, this parameter amounts to only
0.27. Testing on many other images revealed that this approach
reliably separates actual defects, e.g. scratches and coagulated
paint, from dust and other small irregularities that are spread
over larger surfaces.
An important issue with the k-means algorithm is that it
requires the user to specify the number of clusters k to be
generated. A simple and popular solution is to inspect the
dendrogram produced using hierarchical clustering to see if it
suggests a particular number of clusters. For example, the
Elbow method looks at the total TSS as a function of the
number of clusters and proposes to choose the number of
clusters so that adding another cluster does not improve the
total TSS considerably. In our test we did not observe TSS but
the maximum Ek as the function of the number of clusters k.
Table 1. shows the results of such an analysis for a single
image.
could be a problem without controlled environment. An
automated inspection device is therefore carefully designed to
meet all these requirements. In many cases surface defects are
visible as concentrated groups of small blobs. For that purpose,
a special parameter E, which defines the likelihood of a defect
in a cluster, is developed. Serial production testing has shown
that apart from the parameter E, the maximum blob size should
still be considered for selecting defective parts. The presented
inspection device eliminates defective parts from further
processes. Furthermore, we are interested in providing
feedback on the KTL coating process, by providing the type,
location and frequency of defects. The most demanding task
here is to identify the type of defect, because various defects
look similar in the image. We did several initial tests with
neural networks, but very ambiguous results were achieved
because of similarities. Our future work is therefore aimed at
the improvement of the measuring systems, as well at the
identification of additional parameters that will contribute to a
clearer definition of the defect type.
Acknowledgements
This work was supported by the Ministry of Higher Education,
Science and Technology of the Republic of Slovenia, research
program P2-0270, and by the TPV Group. The authors would
also like to thank the staff of the TPV for their assistance and
support in this research work.
References
[1]
[2]
Table 1. The parameter E with respect to the number of clusters k.
k
max(Ek)
2
0.338
3
0.579
4
0.753
5
0.757
6
0.850
7
0.853
[3]
[4]
The cluster with the maximum Ek was always in the same
location when k > 2. If k is 4 and 5, Ek is almost the same,
indicating that the elbow is around 4. By repeating this test on
other images, similar results are achieved. For this reason, we
use k = 4 at k-means clustering.
[5]
5. Conclusion
[7]
The difference image acquisition and processing method
reliably reveals surface defects on black shiny KTL-coated
parts. For its successful operation, a reference image of the
background is required. It must be aligned with the examined
image and acquired within same lightening conditions. This
[8]
[6]
“E-Coat (KTL).” [Online]. Available: http://www.elameta.lt/e-coatktl.html. [Accessed: 10-Feb-2019].
S. Satorres Martínez, C. Ortega Vázquez, J. Gámez García, and J.
Gómez Ortega, “Quality inspection of machined metal parts using
an image fusion technique,” Measurement, vol. 111, pp. 374–383,
2017.
L. Arnal, J. E. Solanes, J. Molina, and J. Tornero, “Detecting dings
and dents on specular car body surfaces based on optical flow,” J.
Manuf. Syst., vol. 45, pp. 306–321, 2017.
A. Isasi-Andrieu, E. Garrote-Contreras, P. Iriondo-Bengoa, D.
Aldama-Gant, and A. Galdran, “Deflectometry setup definition for
automatic chrome surface inspection,” in IEEE International
Conference on Emerging Technologies and Factory Automation,
ETFA, 2018.
F. Pernkopf and P. O’Leary, “Image acquisition techniques for
automatic visual inspection of metallic surfaces,” NDT and E
International. 2003.
C. Jian, J. Gao, and Y. Ao, “Automatic surface defect detection for
mobile phone screen glass based on machine vision,” Appl. Soft
Comput., vol. 52, pp. 348–358, 2017.
C. Fernandez, C. Platero, P. Campoy, and R. Aracil, “Vision system
for on-line surface inspection in aluminum casting process,” in
Proceedings of IECON ’93 - 19th Annual Conference of IEEE
Industrial Electronics, 1993, pp. 1854–1859 vol.3.
“TPV Group.” [Online]. Available: https://www.tpv.si/en/.
[Accessed: 10-Feb-2019].