automated quality control in ceramic industry - … quality control in ceramic industry ... gives...
TRANSCRIPT
Automated Quality Control in Ceramic Industry Rafael Guerreiro Baeta1
[email protected] Mechanical Engineering Department, Instituto Superior Técnico, Lisboa, Portugal
Abstract - This work has the goal to design an
automatic system to detect flaws and edge
irregularities on ceramic plates. Edge
irregularities are detected as well as black spots
on both sides of the plates. The detection of
defects using photography and image processing
techniques allows image interpretation leading
to a decision to accept or reject the plates. A
special system of lightning was designed. This
lightning does not focus directly on the plates,
therefore avoiding undesirable reflects.
To analyse the data, three algorithms were
developed, the first to detect the black dots on
the upper side of the plate, the second with
identical role for the lower side of the plate and
the third to analyse the edge and the geometry.
Binarization and techniques of morphology
were used to detect flaws on the plates.
The tests performed, with the developed
prototype, concluded that the system has a
reliability of 100%. Tests were performed using
643 photographs of plates without flaws and
with the most common flaws. The plates were
carefully rotated, to place the flaws in different
positions towards the camera. The detection
time varied, for the 3 algorithms, between 2,95
and 3,46 seconds for each operation.
To conclude, it is possible to say that a simple
and functional system was projected. This
system has low construction costs and an
elevated reliability, which can contribute to
enhance the quality of the plates and lower its
costs, increasing, therefore, national
competitively.
Index terms - Quality control of ceramic plates,
CAD, Automation Visual Inspection, Image
Acquision System, Ceramic Defects, Binary
Image, Hough Transform
1. Introduction
In order to attain high quality in Portuguese
porcelain, great care is necessary, as well as a good
technology and, specially, a very thorough
inspection to deliver flawless objects and avoid the
sale of defective pieces. Until a few decades ago,
trained workers performed quality control, as
manual labour was cheap and abundant. Lately, this
situation has become unaffordable, due to its costs
[1].
Therefore, automatic systems of quality control
were developed, but due to their cost [2], most
factories chose to continue to employ manual
inspection, using samples from the lot, instead of a
100% inspection of all pieces [3]. This system is far
from perfect and the competition from emerging
countries, such as Asian countries, with very cheap
manual labour, require from the Portuguese
factories a higher quality and no fails in the final
products.
This situation lead to the idea of developing an
automatic system that allows the detection of the
tiniest defect, reducing simultaneously the costs of
manual labour of sample inspection. To this
purpose, a system of automated transport of plates
was projected. This system, allied to a system of
visual inspection and specific software to interpret
the images and select between good or flawed
plates, gives instructions to the machine to separate
the good from the bad.
In this report it’s shown how the developed
algorithms are able to detect any flaw in the upper
and the lower surfaces of the plates used for this
work, which were kindly provided by the company
Matcerâmica. (Figure 1)
Figure 1 - Sample of plate provided for this project
2. Description of the proposed system
The developed system is divided in 5 modules:
1. Centering system;
2. Vision and lighting system;
3. Turning system;
4. Rejection system;
5. Stamping system.
2.1. Centering system
The first problem to arise was how to correct the
random placement of the plates on the feed belt, not
only in what concerns the space between the plates,
but also the correct central placement on the belt.
This perfect alignment with the camera is crucial to
capture the images.
The space between plates was determined taking
under consideration the time frame necessary to
have the plate photographed, in order to have time
for the analysis of the data, and to trigger the
machine that will separate the good plates from the
flawed.
The ideal time is 4 seconds per plate, which means
15 plates per minute or 900 plates per hour.
To solve this problem 2 fed belts were used, one to
receive the plates and the other to take them
sequentially to the centering system (Figure 2).
Figure 2 – Centering system
This way, the first belt has a sensor connected to a
system that controls the start and stop of the belt,
allowing the plate to slide into the second fed belt
at 4 second intervals (Figure 2).
To center the plates, a system installed over the belt
was used. This system comprehends 2 steers
connected to a pneumatic cylinder that allows a
pressure equilibrium that will exert some pressing
on the plate, placing it exactly on the center of the
belt. A low-pressure system was used to avoid any
physical damage to the plate. This belt will feed the
next belt where the vision and lighting system is
installed.
2.2. Vision and lighting system
This is the true innovation of our project, as all the
other parts, in spite of being important for the
overall functioning, are nothing more than a
mechanical control of simple robotics.
The plate will enter a special dome, in the shape of
a half cylinder (Figure 3), illuminated by leds, as
suggested by Ahmed Patel [4]. This shape ensures
a diffuse illumination where there are no light
reflexions on the plates whatsoever. In spite of
being known that the lighting system usually
employed in this kind of work is a dome, in our
project this system was used, with better results.
Figure 3 – Vision and lightning system
This system has the capacity to receive plates until
450 mm of diameter and is composed by a lateral
wood base and a dome in metallic plate of 1,5 mm,
painted in white matte to avoid any reflections. A
rail with all the lighting system was placed inside,
so that that light would not shine directly on the
plates, but indirectly through the reflection on the
dome.
The illumination was installed through a led
system, with a ribbon with 5m and 150 leds/m type
SMD5630 (cold white 6000k), cut and distributed
in an even way through the rails. The power of the
ribbon (40w) requires an adequate power source
and the choice was one from the brand MeanWell,
model 100F12, and giving 8,5A at 12 VDC.
The plate stops 1 second under the camera to be
photographed and here the choice was a DSLR
digital single lens reflex camera with a sensor type
CMOS (complementary metal-oxide-
semiconductor). In spite of not being an industrial
equipment, it fills perfectly our needs and it’s
cheaper, as it has a CMOS sensor and not a CCD
(charge coupled device), which would allow for an
image of better quality, but this is not necessary
thanks to the developed software.
During the rehearsals the camera used was the
Canon 600D with 18 megapixels and a Canon lens
EF-S 15-85mmR/3,5.5,6 IS USN, which, of course,
kept the same configurations during all the tests.
The configurations used were as follows:
• Focal distance 15 cm;
• Index of sensibility (ISO) of 100;
• Opening shutter speed 1/100;
• Shutter opening 5,6.
2.3. Turning system
This module will allow an 180º rotation of the
plate, in order for it to go through a second module
of vision and lightning to detect flaws in the other
side of the plate.
This system is completely mechanical and is made
of two belts (a feeding belt and a leaving belt) and
the turning mechanism itself. Using V-belts, the
first belt will take the plate to the turning
mechanism. The belt has an opening that allows
three fingers to pass through it. These capitation
fingers will take the plate from the first belt and
lead it to the second belt in identical conditions, in
order to go into the second vision lightning system.
There are two supports that keep the plate always
centered (for the second inspection), preventing the
plate from moving (Figure 4).
Figure 4 – Turning system and detail of the griping system
2.4. Rejection system
This module is rather simple, composed by a
pneumatic system using a fixed cylindrical
backstop which, when activated, pushes the plate
out of the belt.
The developed software, when detecting a flawed
plate, triggers an electric valve firing the cylinder,
pushing automatically the plate outside the belt
(Figure 5).
Special attention must be given to the belt, which
cannot be made of traditional covering, because
this will cause friction with the moving plate, so, in
this project, a motorized roller belt was used to
transport the plates.
When exiting this module, the flawless plates
remain on the main belt and proceed for stamping.
The flawed plates are steered to a parallel line,
where they are inspected manually and can be
recycled if the flaws are deemed acceptable.
Figure 5 – Rejection system
2.5. Stamping system
This is a simple mechanical system composed by a
belt and stamping equipment.
The plate, after being inspected, enters this belt and
stops under the stamping equipment for one
second, which is the time allotted to this function
and the logo and the batch are placed unto the plate
(Figure 6). This takes us to the end of the line and
now the plates only need to be packed and prepared
for expedition.
Figure 6 – Stamping system
The assembled complete system is shown in Figure
7.
Figure 7 - Complete system top view and side view
3. Implemented methods
3.1. Default types of plates
For our work 3 types of plates were used: white
plates with worked rim, white plates with simple
rim and plates with decal and simple rim.
From the lot of plates received from Matcerâmica,
some had the kind of flaws that are currently found
on their production line. The flaws found on those
plates are black dots, clear and dark granules,
bubbles, decal and broken edges.
The developed software allowed the detection of
any irregularities on the edge of the plates as well
as any dots on the white plates.
3.2. Pre-processing
Now that the types of flaws that have to be detected
are identified, it is possible to proceed to the
description of the algorithms implemented in this
project. These were developed using the software
Matlab version R2009b 64bit Mac OSx.
Three algorithms were developed:
• One algorithm to detect black dots on the
upper side of the plate;
• One algorithm to detect black dots on the
lower side of the plate;
• One algorithm to analyse the edge and
geometry of the plate.
In the pre-processing stage there will always be the
following steps (Figure 8):
Image opening - the image, obtained from the
vision and lightning system, is read by the
algorithm with its original name. The algorithm is
able to read the image where it is saved, without
having to be introduced in the program by an
operator.
Image resizing – the original size of the image is
reduced.
Binarization – a value of threshold is defined and
the initial image is transformed into a binary image.
New image resizing – the binary image is again
reduced.
Application of the Hough transform – with the
reduced image, the algorithm of the Hough
transform is introduced, to calculate the outline of
the plate. The outputs obtained are the radius of the
figure and the coordinates of its center.
Crop – this allows us to cut the image and leave
only the plate with a small contour. This cut is only
possible thanks to the coordinates of the center of
the plate, as well as the radius obtained earlier.
Figure 8 – Pre-processing stages
3.3. Algorithm for the detection of flaws on
the upper side of the plate
Binarization and morphology techniques were used
to calculate this algorithm. This is a simpler
Image opening Image resizing Binarization
New image resizing
Application of the Hough transform
Crop
approach with simple implementation and fast
processing (Figure 9).
In this process the critical point will be the
threshold value to be used. The camera used in this
project, in spite of not being considered a high-
speed camera, is considered a high-resolution
camera. This fact leads to image detection of
details of the plate that are practically invisible to
the human eye. This way, when proceeding to
binarization, if the threshold value is too high,
beyond the flaws of the plate, there will be the
detection of undesirable details, when the objective
is to have detection only of the black dots
considered as flaws. So it was decided to use a
manual threshold, with this value perfected for the
slightest flaw (approximately 1 mm).
To guaranty only the presence of flaws in the
image, the algorithm was strengthened,
implementing erosion, eliminating undesirable
details.
After choosing a threshold and applying erosion,
the algorithm proceeds to the inversion of the
colour of all the pixels and counts the number of
defects (set of white dots) existing in the image.
Figure 9 - Algorithm for the detection of flaws on the upper side of the plate
3.4. Algorithm for the detection of flaws on
the lower side of the plate
Concerning the lower part of the plate, the
algorithm used is practically the same (Figure 10).
However, it is necessary to eliminate the image of
the edge of support of the plate, since there may be
some irregularities on the support that would distort
the results.
Figure 10 - Algorithm for the detection of flaws on the lower side of the plate
3.5. Algorithm to analyse the edge and
geometry of the plate
After the pre-processing phase, the perimeter of all
the components of the image is detected. Therefore
an image with edge of the plate is showed, as well
as the edges of the flaws, if they exist.
The limit lines of the image are analysed. After the
detection of the first and the last lines, the exact
coordinates are obtained in the spacing between
them. The same process is applied for columns and
the exact center of the plate is obtained.
Subsequently, all internal and external pixels are
removed and a survey for all the pixels of the edge
is conducted.
The (x,y) position of all the pixels allows, using the
center, to know the distance of each to the center.
After obtaining all the distances, the mean
distances and the standard deviation are calculated.
The detection of a flaw on the edge is made using
the minimal distance detected on the table of
distances of the several pixels to the center. If this
value is inferior to the mean distance calculated,
minus the standard deviation, a flaw is signalled on
that spot.
To analyse a bulge on the edge, the process of
detection is identical, and the maximal distance
value found is used as reference.
Threshold value (pre-processing)
Color inversion Defects count
Threshold value (pre-processing)
Lower rim elimination Color inversion
Defects count
4. Results
4.1. Illumination
The use of a diffuse illumination system proved to
be a good choice as, allied to a black base, it
eliminates parasite glows and allows the camera to
visualize all characteristics of the plate, including
rim and decal, as well as all the flaws.
4.2. Database
Three sets of plates received from Matcerâmica
were tested: white plates with worked rim, white
plates with simple rim and plates with decal and
simple rim. In those sets of plates we had all the
commonly found flaws.
Due to the limitation of the number of samples and
to increase the credibility of the test, each plate was
photographed 5 to 6 times, rotating the plate in
order for the flaws would not be located always on
the same spot, towards the camera.
All the results were registered in Excel and the
flowchart of the database is given in Figure 11.
Figure 11 – Database Flowchart
Independently of the position of the plates, the
camera always identified the flaw in the same
manner.
4.3. Results of Implemented methods
For the upper side algorithm we used 98 plates that
include flawless plates and plates with flaws. For
the lower side, we used 67 plates with the same
conditions. All these plates are included in the 643
images database. The main results for all 3
algorithms are presented on Table 1.
Table 1 – Results for the implemented 3 algorithms
Plates correctly classified
Precision of the number of defaults
Average processing
time Upper side algorithm
100% 98,97% 3,46s
Lower side algorithm
100% 74,63% 3,04s
Edge and geography algorithm
100% - 4,8s
It is safe to say the first 2 algorithms have the same
ability of correctly classifying all kinds of plates
(flawless and with flaws). However, the first
algorithm has a higher percentage of accuracy.
Both algorithms have good processing times, witch
means they can be implemented almost directly
into the automated inspection system. On Figure
12 and Figure 13 we can see defects detected on
the upper side of the plate and the lower side,
respectively.
Figure 12 - Plate with default on the upper side
Figure 13 - Plate with default on the lower side
Using the third algorithm, all 11 places with
defective rims were signalled as having flaws
(Figure 14). This algorithm has a slightly bigger
processing time. This algorithm was used to check
the geometry of 98 plates and no size defects were
detected. All plates were deemed acceptable. This
algorithm has also a good processing time.
Figure 14 – Plates with defective rims
5. Conclusions
With the development of this project, the main goal
is to show that it is possible to use a low cost and
low energy expenditure equipment to install an
automatic system to select plates produced by a
ceramic factory.
This system not only optimizes labour but also
allows the inspection of plates using a reliable and
quality system that replaces the inspection by
samples of a batch [5].
On the other side, the advantage of not falling in
the routine of fatigue that befalls the manual
inspectors leads to an elimination of human errors.
Several authors describe this and some defend that
the inspection of samples from a lot is superior to
the inspections of 100% of items, to avoid fatigue
and human error.
This project focus in the automatic inspection
system of the flawed plates, but the assembly line
can easily be completed with an automatic feeding
system, with a cleaning system by air jet to clean
the plates, considering that the materials used in
ceramics cause respiratory problems [6] and,
finally, with an automatic packing system.
The described equipment is 100% reliable,
economically profitable, and can make a
contribution for the development of Portuguese
industry, giving it conditions to compete in the
global world market.
6. References
[1] Timothy S. Newman and Anil K. Jay, "A survey of Automated Visual Inspection," Computer Vision and Image Understanding, vol. 61, no. 2, pp. 231-262, March 1995
[2] P. Wambacq. and A. Oosterlinck L. Van Gool, "Intelligent Robotic Visions Systems," Inteligent Robotic Systems, pp. 457-507, 1991
[3] José Queirós, Cerâmica Portuguesa e Outros Estudos, 2nd ed., Editorial Presença, Ed. Lisboa, Portugal, 2002.
[4] Ahmed Patel, Leila Yazdi, Anton Satria Prabuwono Ehsan Golkar, "Ceramic Tile Border Defect Detection Algorithms in Automated Visual Inspection System," Journal Of American Science, vol. 7, pp. 542-550, 2011
[5] Timothy S. Newman and Anil K. Jay, "A survey of Automated Visual Inspection," Computer Vision and Image Understanding, vol. 61, no. 2, pp. 231-262, March 1995
[6] "Guia de Boas Práticas para a Redução da Exposição à Sílica Cristalina Respirável na Indústria Cerâmica," Centro Tecnológico da Cerâmica e do Vidro, Coimbra, Relatório de trabalho nº: 333.18146-4/10 2012