Content uploaded by Mart Tamre
Author content
All content in this area was uploaded by Mart Tamre on Sep 07, 2015
Content may be subject to copyright.
8
th
International DAAAM Baltic Conference
INDUSTRIAL ENGINEERING
19-21 April 2012, Tallinn, ESTONIA
AUTOMATIC PRODUCTS IDENTIFICATION METHOD
Põlder, A.; Juurma, M; Tamre, M.
Abstract: This paper gives an overview of
automatic products identification method
which can be used in some respects for
standalone or complementary tracing of
objects. The method can be applied only in
a case when each traceable object has
unique visual properties (for example
wood). The basis of described work is
traceability of raw material in wood supply
chain, especially traceability of the saw
material in sawmill.
The first development in this kind of
approach originates from EU- 6th
Framework Project Indisputable Key and
is developed further at Tallinn University
of Technology.
Keywords: Identification, traceability,
production automation, machine vision,
saw material.
1. INTRODUCTION
In today’s industry, where the energy
efficiency and reducing carbon footprint
has main importance alongside with quality
of the products, monitoring of products and
materials, tracing, inspection and
respective control of the production
processes are gaining more and more
importance. Those processes are more
adapted at consumer goods production and
for example in metal working industry at
the moment. In renewable materials
industry like wood or polymer or natural
materials industry the principles are only
gaining attention recently.
Considering the wood production sector
the target to implement the
abovementioned principles is to investigate
the quantitative measures of performance
of the wood supply chain, and
quantification of the potential effects on
the performance. Products and materials
automatic or semi-automatic traceability
should support finding tools and methods
for holistic supply chain management,
optimization and trade-off analysis by
combining product quality and process
economy with environmental impact from
life cycle perspective, which could provide
a new dimension to decision support
systems in the industry and would avoid
sub-optimization due to its holistic nature.
Combining the information flow with the
physical flow of material allows to
associate objects with information. In order
to gain information, a data acquisition
infrastructure is needed and the data
acquired has to be sensibly selected
according to the business needs [1].
To achieve material traceability several
technologies can be used, for example
RFID or machine vision systems. When the
objects are traced on the production line
the position specific information can be
used for tracing.
2. SYSTEM OVERVIEW
Wood supply chain consists of several
parts, for example: harvesting, transport,
log sorting etc. One important part is to
ensure the traceability of saw material
inside sawmill.
The marking and reading of the boards
starts from sawmill green sorting position
where the sawn boards are sorted by
quality. Boards are marked and after that
the code is verified by the reading system.
Next reading position is in the position
where the boards are packed for the kilns.
After the kilning process the boards are
sent to the final sorting where they are re-
sorted. In the final sorter the board end
with marked code is removed. To maintain
traceability it is necessary to read the code
before the board is cut. After that the code
is remarked and verified.
The 8x18 Data Matrix ECC200 barcode
was used since it has good error correction,
high redundancy and high information
density. The codes were applied by
industrial printer and read by smart camera
based vision system.
During the tests it occurred that it is very
hard to achieve code readability over 95 %
in one reading position on automated lines
due to different problems regarding to
material surface and marking quality. This
means that in a case when there are for
example 4 reading positions and the
readability in each position is 95 % the
overall readability and therefore
traceability is about 80 %. Therefore it is
necessary to improve the readability.
3. QUALITY PARAMETERS
To ensure code quality and improve
readability the code quality parameters
were introduced [2]. By evaluating the
quality parameters automatically it is
possible locate the source of the problems
for elimination. This evaluation system can
be used in collaboration with reporting
system, which can trigger the alarm when
the quality decreases for some longer
period.
The quality parameters can be used for
estimating if the code will be read in
following positions. If the code has high
quality grade it is not necessary to store
any additional information about the code.
If the code has low quality there is higher
possibility that it is not read in the
following positions and it is necessary to
store some additional information into
database for further matching.
Several dimensions which can be measured
from acquired board end image can be used
for describing board end and code quality.
Board end mean intensity has strong
influence in code readability. When
intensity is low the contrast between the
code and its background is also low.
Another dimension which is used for
quality estimation is board end region of
interest (ROI) histogram which describes
the distribution of pixel intensities. For
quality estimation the lookup histogram is
generated by using certain amount of
images where the board end and image
quality is good. Each processed image
histogram is compared with lookup
histogram and the difference and its
standard deviation gives us the estimation
about how different it is. Histogram is
good for estimating overall board end
brightness and its distribution, indicating if
the contrast between background and code
is sufficient.
To estimate how the pixel intensities are
distributed over the board end area the
board end was distributed into vertical and
horizontal stripes (ROIs). Mean intensity
of each stripe is calculated and differences
between beside intensities were found.
High standard deviation of those
differences indicates that there are lots of
changes in intensities and therefore the
quality is low.
In many cases the marked code has
deformed due to board end vibrations
during the code marking. To estimate if the
vibration is present the code edge
coordinates in several rows and their
standard deviation is found. If the deviation
is high then the code has vibrated and
therefore the quality is low.
For deciding if the board end has
acceptable quality it is necessary to
calculate value which summarizes all the
parameters. The total quality estimation
can be calculated by summing up all the
quality parameters giving each one specific
scale. The scales were found
experimentally by measuring all the
parameters of large scale test reading, the
base of decision was the positive reading
result. All the described quality parameters
do not have strong correlation with
readability and therefore they are used as
estimators. [2]
4. HISTOGRAM BASED MATCHING
Quality parameters give the estimation if
the code is readable in following positions
and approximate estimation why the code
is not read. They do not give additional
information which could be used for
identifying board in following positions.
For that some additional information is
needed. It is not reasonable in first
approach to measure specific features or
defects but to try to evaluate the whole
region of interest at once and try to
implement something more common and
simpler like histogram of board end.
Histogram shape depends on the lighting
conditions, size of board end (size of ROI)
and changes inside ROI (Fig. 1. and Fig.
2.). If the conditions in between reading
positions differ then the histograms are
different. Therefore it is necessary to keep
the lighting conditions and all the other
parameters as similar as possible in each
reading position.
Fig. 1. Marked code on the board end
Board end histogram
0
1000
2000
3000
4000
5000
6000
1 21 41 61 81 101 121 141 161 181 201 221 241
Intensity
Frequency
Fig. 2. Board end histogram (peaks starting
form left – image background, marked
code, board end area)
Each board end histogram is quite unique,
especially when there is a marked code
which gives a peak on certain range of
intensity values.
When using histogram as unique identifier
it is necessary to pre-process it due to
probable changes in between positions and
changes on the board end when it is
moving from one position to another. All
the histograms have to be made
comparable, for example if the board end is
darker or lighter than original or the
distance from the camera is changed the
histograms are not comparable anymore.
To compare histograms it is necessary to
smooth, rescale and resample the
histogram. Rescaling of the re-sampled and
shifted histogram ensures that the
maximum peak frequency is the same on
all the resulting histograms. Re-sampling
of histogram compensates board end
distance change and rescaling compensates
brightness change.
From the test it occurred that histogram is
quite sensitive to changes in between
different positions: brightness changes,
board end size, defocusing and rotation. It
appeared that lots of visually quite
different board ends had similar histograms
and that resulted in mismatches. One
downside of histogram based traceability
enhancement is that the information gained
has no position specific information (for
example code location on the board end).
That means that two totally different board
ends can have similar histogram. That
indicates that it is not possible to use
histogram alone for improving traceability.
[3]
5. LINEAR AVERAGES
Another way to improve the matching of
the boards is to use linear averages based
matching.
Linear averages on x and y axis direction
are the arrays of mean values of each pixel
line in the image region of interest (ROI).
The positions of different objects on the
board end like code or branches are
reflected on the directional averages
making this method more unique than
histogram only [1].
For testing the linear averages method the
algorithms for measuring and comparison
were created. The measuring algorithm
loads image from file, finds board end
region of interest and rotates image to
ensure that board end wider edge is
horizontal. After that the linear averages in
x and y direction is calculated (Fig. 3.).
Fig. 3. Linear averages (white plots)
From the averages plots it is possible to
aim the location of the code resulting the
lower average values and therefore lower
part of the plot. After that the results are
saved for further processing. The
measuring application measures each
image and no comparison takes place in
this phase.
In the comparison algorithm the simple
linear correlation is used for finding the
correlation between two comparable data
sets. Correlation coefficient close to 1
indicates good match.
Since linear correlation is independent of
both origin and scale of the samples it is
not necessary to rescale the measured
values. That means a small unified change
in lighting does not have effect on the
comparison results.
The highest correlation coefficient of the
results is probably the match if the
conditions are acceptable and images of the
same board end are similar.
Board end and therefore region of interest
has to be rectangular, when the board end
is not positioned correctly in front of the
vision system it may appear not completely
rectangular on the image. When measuring
linear averages on those images some parts
of board end may be out of ROI or some
background can be included into ROI and
results are strongly affected.
To evaluate if it is possible to match the
images in between the different reading
positions it is first necessary to estimate in
what range the correlations coefficients
between not matching images are.
In the first test 500 images of marked
board ends, acquired from saw line were
measured and correlation coefficients
between all the measurement results were
found (total 124750 comparisons were
made). The results show how the
correlation coefficients are distributed.
From x axis linear averages correlation
coefficients histogram it appeared that
most of them are distributed around 0.5
and highest correlation coeffcient is 0.94
(Fig. 4.) Note that chart y axis (frequency)
indicates the amount of correlation
coefficient which falls in defined value
range (bin).
0
2000
4000
6000
8000
10000
-1 -0.5 0 0.5 1
Correlation coefficient bin
Frequency
Fig. 4. Histogram of x axis linear averages
correlation coefficients
In the y axis case (Fig. 5.) the coefficients
are distributed around 0.6 and 355
coefficients (about 0.3 %) are in range
from 0.95 to 1.
0
2000
4000
6000
8000
10000
-1 -0.5 0 0.5 1
Correlation coefficient bin
Frequency
Fig. 5. Histogram of y axis linear averages
correlation coefficients
The reason why the peaks is shifted to right
is smaller amount of data since board end
is rectangular.
In many cases the high correlation
coefficient in one direction linear averages
does not mean high correlation coefficient
in another direction. That means that
indicating the correlation between both, x
and y direction should give better result.
To do that the x and y direction
correlations coefficients are multiplied. In
this case histogram peak is around 0 to 0.5
and no close matches exist, no multiplied
correlation coefficients fall in the range
from 0.9 to 1 (highest is 0.8740). Therefore
this parameter seems to be very good
candidate for the comparison (Fig. 6.).
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
-1 -0.5 0 0.5 1
Correlation coefficient bin
Frequency
Fig. 6. Histogram of multiplied linear
averages correlation coefficients
To evaluate the effect of probable changes
in between reading positions several test
images were generated (changed image
size, changed image brightness, half
darkened and added dark dot). Similar test
were run as described before only in
between modified images and original. It
appeared that in a case when the image size
was changed the multiplied x and y
correlation was below 0.5.
To improve the response of changes on
board end size change the data was
interpolated. The smaller array is
interpolated so that it has the same number
of elements as major array and the first and
last element value remains the same.
The previous test were run again and now
the correlations between each modified
image which size or brightness was
changed had correlation coefficient (x*y)
over 0.9.
Finally the test with set of 500 images and
modified images were run.
In this case the matches between the
original and its modifications are more
widely distributed; all values above 0.86
belong to randomly selected image and its
modifications. The highest correlation
coefficient between other images is 0.85
(Fig. 7. and Fig. 8.).
0
50
100
150
200
0.6 0.7 0.8 0.9 1
Correlation coefficient bin
Frequency
Fig. 7. Histogram of multiplied linear
averages correlation coefficients without
modified images
0
50
100
150
200
0.6 0.7 0.8 0.9 1
Correlation coefficient bin
Frequency
Fig. 8. Histogram of multiplied linear
averages correlation coefficients with
modified images
According to that it is possible to match the
images of board end modifications if the
number of comparable objects is small. In
real application that can be much more
complicated since the number of images is
larger and the differences in between
comparable positions can vary more. In
described tests the comparable images
were modified manually and the actual
differences in between the positions can be
more complex. [1]
6. PROBLEMS
When using histogram and linear averages
for achieving a traceability of the objects
additional comparison algorithm is needed.
When reading the code in first position and
the same code in second position it is quite
easy to match those two readings – the
code is exactly the same and that means
that only one database query is needed. In a
case of histogram and linear averages
reading systems are similar, codes are read
and all the parameters are measured and
data is saved into database. To match the
readings the matching algorithm is needed.
Basic idea would be to compare data of
each image with all the other data entries
and to find correlation between them. The
highest correlation is probably the correct
match. Depending on the physical distance
between two reading positions the amount
of processing data can be huge since each
data entry must be compared with all the
other entries.
One way to reduce the necessary entry
comparisons for matching is to classify the
data. Choosing the right classifiers and
amount of classes is important. Using
classes is useless if most of the entries are
in same class or if there are lots of classes.
Most important thing is to avoid situation
when the entries about the same board end
fall into the different classes, in this case it
is impossible to match them.
One possibility to create classes is to divide
all entries into groups simply by some
parameter, for example board end
brightness. The interval in between
maximum and minimum can be divided
into several smaller intervals. That raises
another problem with the cases when the
entry is close to the edge. Due to
differences in between positions (lighting
position etc) and possible changes on the
board end all the parameters can vary in
between positions. Therefore it is possible
that in one position the entry is placed into
one class and in another in another class.
One simple way would be to add entries
close to the edges to the both classes. That
increases amount of time for comparison
but reduces the risk of lost match.
Another problem arises when there is no
good, outstanding match – one solution
would be to consider that entry to be lost.
Another approach would be to give several
possible matches as a result. Later some of
the close matches would be eliminated by
other matches and therefore it is possible
that finally only the correct match remains.
7. CONCLUSIONS
The main scope of this paper is to give an
overview of automatic products
identification method based on board end
traceability in sawmill. The described
method is not tested in real application; it
is tested in the laboratory using previously
acquired data.
Overview of board end quality parameters
helps us to understand the main sources of
problems and is a base for decision in a
case when the marked codes are used in
parallel with object uniqueness based
traceability system. The board end
histogram can be used as second step for
estimation since histogram itself is not
unique enough for traceability purposes.
Final match of objects can be made using
linear averages.
The described approach needs further
development. Board end finding algorithm
must be improved since in many cases it
does not work correctly. The quality
parameters must be redefined based on
larger scale tests. Histogram and linear
averages based matching algorithm needs
to be improved to gain higher efficiency
and processing speed.
8. REFERENCES
1. Põlder, A., Juurma, M., Tamre, M.
Wood Products Automatic
Identification Method Based On
Fingerprint Method. In Mechatronics
systems and Materials: Abstracts
(Skiedraute, I. Basukutiene, J.,
Dragašius, E., eds.). Kaunas University
of Technology, 2011, 55.
2. Põlder, A., Abiline, I., Tamre, M.,
Automatic Visual Code Quality
Evaluation For Wood Industry. In 7
th
International Conference Of DAAAM
Baltic Industrial Engineering:
Proceedings (Küttner, R., eds.). Tallinn
University of Technology, 2010, 554.
3. Põlder, A., Modified fingerprint
method for traceability enhancement of
marked products. In Topical Problems
in the Field of Electrical and Power
Engineering * Doctoral School of
Energy and Geotechnology II:
Proceedings (Lahtmets, R., eds.).
Tallinn University of Technology,
2010, 121
7. DATA ABOUT THE AUTHOR
Põlder Ahti, doctoral student,
Department of Mechatronics, Ehitajate Tee
5, Tallinn 19086, Estonia,
ahti.polder@ttu.ee
+372 620 3207