DEFECTIVE METAL END DETECTION WITH A FUZZY SYSTEM
Perfecto Mari
˜
no Espi
˜
neira, Vicente Pastoriza Santos,
Miguel Santamar
´
ıa S
´
anchez, Emilio Mart
´
ınez Exp
´
osito
Universidad de Vigo, Departamento de Tecnolog
´
ıa Electr
´
onica, E.T.S. Ingenieros Industriales,
Campus Lagoas-Marcosende, 36310 Vigo, ESPA
˜
NA
Keywords:
Image processing, fuzzy control, machine vision system, intelligent fault detection.
Abstract:
The authors have been involved in developing an automated inspection system, based on machine vision, to
improve the repair coating quality control (RCQ control) in can ends of metal containers for fish food. The
RCQ of each end is assesed estimating its average repair coating quality (ARCQ). In this work we present
a fuzzy model building to make the acceptance/rejection decision for each can end from the information
obtained by the vision system. In addition it is interesting to note that such model could be interpreted and
supplemented by process operators. In order to achieve such aims, we use a fuzzy model due to its ability to
favour the interpretability for many applications. Firstly, the easy open can end manufacturing process, and
the current, conventional method for quality control of easy open can end repair coating, are described. Then,
we show the machine vision system operations. After that, the fuzzy modeling, results obtained and their
discussion are presented. Finally, concluding remarks are stated.
1 INTRODUCTION
In the food canning sector, in the easy open can end
manufacturing process, to guarantee the desired prod-
uct lifespan, a manual, nondestructive testing (NDT)
procedure is carried out. Due to the high process-
ing rate, only an small part of each lot is verified.
Therefore, it is important to develop an automated in-
spection system to improve the easy open can end re-
pair coating quality control process (all can ends are
checked, and inline). It is for this reason that we had
been involved in the design and implementation of an
inline, automated machine vision system to evaluate
the repair coating of the can ends, that we have named
end repair coating inspection system (ERCIS). In this
work we explore the use of fuzzy models to make the
acceptance/rejection (A/R) decision for each can end.
A Takagi-Sugeno-Kang fuzzy model is developed us-
ing a neuro-fuzzy modeling. The remainder of this
paper is organized as follows: in the next section we
provide an overview of the easy open can end manu-
facturing process and its repair coating quality control
process. Section 3 shows the machine vision system
operations. Then, Section 4 describes the fuzzy mod-
eling. The results obtained and their discussion are
presented in Section 5. Finally, we state concluding
remarks in Section 6.
2 BACKGROUND
A can consists of can body and can end, which are
made from aluminum or steel. Can ends, from hence-
forth ends, are used for all type of cans and can be
standard or easy open. In this paper, we study a spe-
cific end format named 1/4 Club, with an easy-open
tab in one of its corners.
2.1 Easy open end manufacturing
process
Easy open ends are made from pre-coated metal coils
or sheets. Ends are stamped from coil or sheets in a
press. After stamping, the ends are scored in a pre-
defined geometric shape (scoreline) intended to ease
the end opening. Finally, a tab is attached to form
an easy open end. These steps are performed after
the end piece has been coated and therefore damage
the coating, especially on the scoreline. Repair coat-
ing, which has a fluorescent pigment, is applied after
these steps on the required area to restore the integrity
of the coating.
270
Mariño Espiñeira P., Pastoriza Santos V., Santamaría Sánchez M. and Martínez Expósito E. (2005).
DEFECTIVE METAL END DETECTION WITH A FUZZY SYSTEM.
In Proceedings of the Second Inter national Conference on Informatics in Control, Automation and Robotics, pages 270-274
Copyright
c
SciTePress
2.2 Easy open end repair coating
quality control
Presently a visual inspection of the easy open ends is
carried out (Lin et al., 1998; CFIA, 1998), where the
inspectors assess the repair coating on the scoreline.
This visual inspection is a manual and NDT proce-
dure. To assist in this end inspection the repair coating
has a fluorescent pigment that stands out as a bright
light blue when excited by an ultraviolet black light,
while the background color remains unchanged. This
inspection is based on a statistical sampling. It is be-
cause of this and the high rate of the repair coating
process (100 to 500 ends per minute, depending on
the end format) that only a small part of each lot is
verified. Therefore, defective can ends can be sent to
the canneries.
3 MACHINE VISION SYSTEM
OPERATIONS
The vision algorithm running on ERCIS take care of
inspects the repair coating quality (RCQ) on each end.
The vision algorithm has two parts: one offline and
other inline. The flowchart of ERCIS is shown in Fig.
1.
Figure 1: Flowchart of ERCIS
3.1 Offline
Before the ERCIS begins the continuous or online in-
spection of ends is necessary to configure or reconfig-
ure a series of parameters that will be used later in the
inline processing.
3.1.1 Time delay configuration
An offline adjustment can be necessary to set the de-
lay, input delay time (IDT), between the sensor at the
ERCIS input detecting the end and the end reaching
the camera, and the delay, output delay time (ODT),
between the ERCIS output sensor detecting the end
and the end arriving next to the rejection system.
3.1.2 Region of interest definition
The scoreline is enclosed in the region of interest
(ROI). Each quadrant of this ROI, and by simmetry,
the whole ROI is geometrically modeled by the para-
meters b and r (Fig. 2). Besides these parameters it is
necessary to add an e width to the ROI (Fig. 2).
ROI
HOR
= f (b, r, e) (1)
Figure 2: End ROI quadrant geometrical model
Parameters will be adjusted before the line continu-
ous working, and they depend directly on the distance
between and and camera (working distance).
3.1.3 Parameter configuration to decide the end
rejection
In order to take the A/R decision for each end dur-
ing the inline processing is necessary to offline con-
figure the Minimum Average Repair Coating Quality
(MARCQ) of the end to not reject it. The object of
this parameter can be seen on Subsection 3.2.2.
3.2 Inline
The end continuous inspection can only start after the
offline parameter configuration. This inline process
has for each end the following sequence of steps:
3.2.1 Acquisition
This process undertakes the detection of the ends and
acquisition of images of them. It is divided into the
subprocess: end detection before the ERCIS, and end
acquisition.
3.2.2 Evaluation
If the decision queue is not empty then an image has
already been acquired and can be processed. This
process has the following steps:
End center location and end orientation.
ROI rectification. The ROI is converted into a
straight line strip to ease its analysis. This strip is
a Look-up-Table (LUT) whose size is n × e pixels.
The length n, in which is divided the ROI perime-
ter, depends on the selected resolution. The recti-
fication method employed selects for each one of
n × e pixels the nearest 4-neighbour (Gonz
´
alez and
Woods, 2002).
Repair coating quality: The RCQ is assessed ana-
lyzing one-by-one all the n positions of the LUT
by means of the model obtained in Section 4.
End A/R decision: The average repair coating qual-
ity (ARCQ) is computed from the RCQ values at
the n positions in which have been divided the ROI.
Then, an end have to be rejected when the condi-
tion ARCQ<MARCQ is given (see the meaning of
MARCQ in Subsection 3.1.3).
Update the decision queue: The decision queue
must be updated after making the A/R decision of
each end.
3.2.3 Expulsion
This process is subdivided in: end detection after the
ERCIS, and end expulsion.
3.2.4 Stop
If stop signal is activated then the process goes to the
flowchart end with independence of the current state
of the process.
4 MODELING
In order to evaluate the RCQ on each one of n posi-
tions of the LUT, a set of a few attributes that contain
most of the relevant information on each one of the n
positions is studied. The 9 attributes computed from
each e-pixel group of n
th
LUT position are: Maxi-
mum pixel intensity (Max), Minimum pixel intensity,
Mean pixel intensity is a measure of central tendency
(location), Median pixel intensity is a measure of cen-
tral tendency (location), Pixel intensity standard devi-
ation ( Std) is a measure of dispersion, Pixel intensity
skewness is a measure of the asymmetry, Pixel inten-
sity center of mass (CoM), Pixel intensity moment of
inertia about an axis passing through the CoM, and
Pixel intensity bisector.
A fuzzy inference system (FIS) (Kosko, 1992;
Yager and Zadeh, 1994; Klir and Yuan, 1995), whose
inputs are the selected attributes, will be used to eval-
uate the RCQ on each of the n positions. As an exces-
sive number of inputs prevents the interpretability of
the underlying model and increases the computational
burden, we look for a model with a trade off between
high accuracy and reduced number of inputs. We got
a modeling problem with 9 candidate inputs and we
want to find the 3 most influential inputs as the in-
puts of the model. We so can build 84 fuzzy mod-
els, each one with a different combination of 3 inputs.
The proposed FIS model is a Takagi-Sugeno-Kang
(TSK) inference system (Takagi and Sugeno, 1985;
Sugeno and Kang, 1988). These models are suited for
modeling non-linear systems by interpolating multi-
ple linear models. The TSK model is designed with
zero order (singleton values for each consequent), 3
of 9 attributes as inputs and RCQ as output. The TSK
model is developed using the adaptive neuro-fuzzy in-
ference system (ANFIS) algorithm (Jang, 1993; Jang
and Sun, 1995).
We use a quick and straightforward way of neuro-
fuzzy modeling input selection using ANFIS to im-
prove the interpretability (Jang, 1996). This input
selection method is based on the hypothesis that
the ANFIS model with smallest RMSE (root mean
squared error) after one epoch of training has a greater
potential of achieving a lower RMSE when given
more epochs of training.
Representative input-output data set of the system
should be selected to tune a model. We have only
worked with a specific end format named 1/4 Club,
with an easy-open tab in one of its corners. We have
selected a collection of 11 ends that agglutinate all
possible end repair coating defects. The obtained
LUT for each end has a length n of 702 positions, with
a width e of 19 pixels. After removing instances with
outlier values, the data set was reduced to 6669 en-
tries. This data set is divided into training and testing
sets of size 3335× 19 and 3334× 19 respectively. The
testing set is used to determine when training should
be terminated to prevent overfitting.
It has been selected grid partitioning as the AN-
FIS partition method. The best model after one epoch
of training selects as input attributes the maximum
(Max), the standard deviation (Std), and the center
of mass (CoM). The problem is that this partitioning
leads to a high number of rules, 2
3
= 8 rules for each
model.
Figure 3: Fuzzy model membership functions
In order to reduce the model complexity we use
subtractive clustering (Chiu, 1994) for the 3 inputs
previously selected. We have selected the model with
range of influence 0.5 that has 3 rules that gives a
RMSE of 0.3106 for training and 0.3091 for testing.
Table 1: Fuzzy Models Rules
Rules Repair Coating
Quality
If Max is Max1 & Std is Std1 S1[5]
& CoM is CoM2 Defective
If Max is Max2 & Std is Std2 S1[5]
& CoM is CoM3 Defective
If Max is Max3 & Std is Std3 S2[10]
& CoM is CoM1 Acceptable
S[m] = singleton (m=mean)
We can further refine said model performance ap-
plying extended ANFIS training. The final model ob-
tained use Max, Std, and CoM as model inputs, RCQ
as output, and 3 rules (Table 1) to define relationships
among inputs and output. The membership functions
for each input feature are shown in Fig. 3 and single-
ton values for each consequent in Table 1.
This model gives a RMSE of 0.0102 for training
and 0.0101 for testing, more similar values which in-
dicate that there is no overfitting. Regarding the inter-
pretability of the model and from its rules, is deduced
that the RCQ at n
th
LUT position is acceptable if and
only if at said position, see Fig. 3 and Table 1, the
maximum pixel intensity is higher than 150 and the
standard deviation is higher than 35 and the center of
mass is close to 9.5, which is the e-pixel group center.
The interpretation of this is the following:
The higher the maximum pixel intensity, the higher
the lacquer quantity.
As a defect region has little or no lacquer and is
more uniformly distributed than an acceptable re-
gion, then the pixel intensity is less scattered (less
Std) in defect regions.
As the ROI is positioned in the way that the score-
line is at its center zone, and as an acceptable end
has the highest lacquer level at scoreline, then at
each one of n LUT positions the nearer of the e-
pixel group center is the CoM, the better the RCQ.
5 RESULTS
The RCQ of each end is assessed estimating its
ARCQ. This average quality is computed from the
RQ of the n positions in which has been divided the
ROI, and where the RCQ at each n position is ana-
lyzed by means of the FIS obtained.
Figure 4: ARCQ classification of the ends
Fig. 4 shows the ARCQ classification of the 11
ends previously used to tune the fuzzy model. The
ends were sorted in descending order of ARCQ.
As each end have to be rejected when
ARCQ<MARCQ is given (see the meaning of
MARCQ in Subsection 3.1.3) then what ends are
rejected will depend on the MARCQ value selected.
For example, if MARCQ is 80 then all 11 ends are
rejected. This flexibility to be able to modify the
rejection threshold is an important property of the
ERCIS.
But the most important result is that, with indepen-
dence of the MARCQ selected, the ARCQ classifi-
cation agrees with the one made by an expert human
inspector.
6 CONCLUDING REMARKS
We have been involved in the implementation of a
machine vision system to improve the repair coating
quality control of the easy open can end manufactur-
ing process. The system has the following properties:
End classification in agreement with the one made
by an expert human inspector.
Flexibility to be able to modify the rejection thresh-
old.
Interpretability supplied to the operators in order to
find out the failure causes and reduce mean time to
repair (MTTR) during failures.
Total inspection of 100% end production.
In spite of the fact that the end repair coating process
of only one end format (1/4 Club) has been studied, as
this process is common to all formats, it is reasonable
to think that fuzzy models like the found model can
be obtained to make the A/R decision for another end
format.
All this leads to the conclusion that is possible
to design an inline, automated machine vision sys-
tem, which only extracting the ARCQ from each end,
makes a right A/R decision. ANFIS, the neuro-fuzzy
modeling technique used to optimize the fuzzy model,
provided excellent prediction accuracy.
In the future we will study the existence of mod-
els that estimate the easy open can end repair coat-
ing process failure causes. Furthermore, we will re-
search the application of coevolutionary genetic fuzzy
modeling techniques that improve the interpretability
without a significant loss of accuracy.
REFERENCES
CFIA (1998). Metal can defects; identification and clas-
sification manual. Technical report, Canadian Food
Inspection Agency (CFIA).
Chiu, S. (1994). Fuzzy model identification based on clus-
ter estimation. J. of Intelligent & Fuzzy Systems,
2(3):267–278.
Gonz
´
alez, R. and Woods, R. (2002). Digital Image Process-
ing. Prentice Hall, Upper Saddle River, NJ, 2nd edi-
tion.
Jang, J. (1993). Anfis: Adaptive network based fuzzy in-
ference system. IEEE, Transactions on Systems, Man,
and Cybernetics, 23(3):665–685.
Jang, J. (1996). Input selection for anfis learning. pages
1493–1499. Proceedings of the IEEE International
Conference on Fuzzy Systems.
Jang, J. and Sun, C. (1995). Neuro-fuzzy modeling and con-
trol. volume 83(3), pages 378–406. The Proceedings
of the IEEE.
Klir, G. and Yuan, B. (1995). Fuzzy Sets and Fuzzy Logic:
Theory and Applications. Prentice Hall, Upper Saddle
River, NJ.
Kosko, B. (1992). Neural Networks and Fuzzy Systems:
A Dynamical Systems Approach to Machine Intelli-
gence. Prentice Hall, Englewood Cliffs, NJ.
Lin, R., King, P., and Johnston, M. (1998). Examination
of metal containers for integrity. In Merker, R., edi-
tor, FDAs Bacteriological Analytical Manual Online.
Center for Food Safety and Applied Nutrition (CF-
SAN), U.S. Food & Drug Administration (FDA).
Sugeno, M. and Kang, G. (1988). Structure identification of
fuzzy model. Fuzzy Sets and Systems, 28(1):15–33.
Takagi, T. and Sugeno, M. (1985). Fuzzy identification
of systems and its applications to modeling and con-
trol. Transactions on Systems, Man, and Cybernetics,
15(1):116–132.
Yager, R. and Zadeh, L., editors (1994). Fuzzy Sets, Neural
Networks, and Soft Computing. Van Nostrand Rein-
hold, New York.