Beruflich Dokumente
Kultur Dokumente
Image analysis
1400 200
800
100
600
400
50
200
0 0
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200
4. Ratio of handsheet ppm to target ppm vs target ppm. Note that 5. Wire side ppm minus felt side ppm vs target ppm. Data randomly
the scatter of data decreases with increasing ppm (whereas the above and below the y-axis demonstrate that no significant differ-
absolute scatter of actual ppm increases with increasing ppm, Fig. ence in ppm existed in this study between the two sides of the
2). No significant difference can be observed in the wire side and felt handsheet.
side data.
6 500
WIRESIDE ppm FELT SIDE ppm
HANDSHEET ppm / TARGET ppm
Wire side
5
Felt side
3 0
0 500
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
just because an automatic and seem- Statistics of a where Ai is the area of a single spot.
ingly more precise method is used. random distribution Since the average area of a spot, AM,
Moreover, theoretical considerations is calculated by:
in this work will allow an estimate of
of impurities in pulp
AM = (1/N) x Ai (2)
the variability of the evaluated The total area covered by impuri-
samples beforehand, when appropri- ties, AT, can be calculated by: with N being the number of spots,
ate estimations of the pulp quality the expected value of the total area
AT = Ai (1) covered by impurities, E(AT), can be
are available.
expressed by:
170 February 1995 Tappi Journal
6. Variance of the number of spots vs number of spots. The data are 7. Predicted error vs actual error. The predicted error was calcu-
scattered about the 45 line, as expected for a random distribution lated using Eq. 8. The actual error is the standard deviation of the
of impurities in pulp. For random distributions, 80% and 50% of the experimentally detected ppm of handsheet sides of a target ppm
data are expected to fall within the respective limits plotted on the batch. Results for the wire side and felt side are similar. The scatter
graph. of data about the 45 line provides validation of Eq. 8.
400 200
VARIANCE OF NUMBER OF SPOTS
200 100
150
100 50
50
0 0
0 50 100 150 200 250 0 50 100 150 200
E(AT) = E(AM) x E(N) (3) because n = m = N, the number of 7 or 8. For this, the number of spots,
The variance (equal to the square of spots on a sample. Using the vari- the average spot size and the vari-
the standard deviation) of AT can ance of the area covered by single ance of the spot size have to be de-
then be written as: spots, Var(Ai) = (Ai AM)2/(N 1) termined.
obtainable from an automated dirt This error can then be compared
Var(AT) = Var(N x AM) (4) count, Eq. 6 yields: to the standard deviation of AT ob-
It can be shown that the variance of tained from the measurement of a
Var(AT) = [N2/(N + 1)][AM2 + Var(Ai)] (7)
a function h(X,Y) of two indepen- series of handsheets made from the
dent stochastically distributed vari- Equivalently: same batch. By repeatedly scanning
ables, X (estimated from n mea- SDEV(AT) = {(N2/(N + 1)) the same handsheet, one can also
surements) and Y (estimated from [ AM2 + Var(AI)]}0.5 (8) gain an estimation of the error in the
m measurements), can be written as measurement of AT introduced by
where SDEV is the standard devia- the imaging device. The three er-
(12): tion, and SDEV(AT) = [VAR(AT)]0.5. rors can then be contrasted to each
Var(h(X,Y)) = [1/(nm 1)] x [(n 1) x m x By factoring out an AMN from the other. If Eq. 3 and Eq. 9 are com-
(h/X)2 x Var(X) + (m 1) x n x ( h/Y)2 right hand side of Eq. 8 and rear- bined, one can gain an equation that
x Var(Y)] (5) ranging, the relative error of the to- estimates the relative error in the
Since our system from which we tal area covered by impurities, measurement of AT as a function of
sample contains mostly water, less SDEV%(AT), can then be determined the level of impurities in ppm (parts
than 1% pulp and an even smaller as: per million):
amount of impurities, the assump- SDEV%(AT) = [SDEV(AT)/AT] x 100% =
SDEV%(AT) = 100% x {[1 +Var(Ai)/AM2]/
tion of independence is justified as 100% x [(1 + Var(AI)/AM2)/(N + 1)]0.5 (9)
[(ppm x AS/(106 x AM)) + 1]}0.5 (10)
long as vigorous agitation is assured.
Equation 9 is useful because it al- where AS is the total scanned area.
It can be shown that the variance
lows an immediate measure of the The only difference between Eq. 9
of the count of randomly distributed
dispersion of the measurement, and Eq. 10 is that N in Eq. 9 is
features is equal to the number of
SDEV(AT) relative to the measure- substituted by ppm x AS/106 x AM in
the features (13). Using this and Eq.
ment value AT. For each handsheet, Eq. 10. To use Eq. 10, the values of
5, Eq. 4 can be rewritten to yield:
one can calculate an estimation of Var(Ai) and AM must be known or
Var(AT) = [1/(N + 1)][N2 x AM2 + N3 x the error in AT due to the random estimated. This equation can be di-
Var(AM)] (6) distribution of impurities using Eq. rectly used to estimate the error of
200 200
RELATIVE ERROR, %
Actual error wire Eq. 10: Var(Ai)/AM2 = 0.0
Actual error felt Eq. 10: Var(Ai)/AM2 = 1.5
ERROR, ppm
80 80
40 40
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000 1200
an analysis given the population pa- the resulting ten handsheets were size of the spots, and a histogram of
rameters of the spots, the level of each approximately 1.5 g. The batch spot size versus frequency. The
impurities and the total scanned was kept well mixed during sam- smallest size spot that could be dis-
area. pling. The handsheets were stored tinguished under the conditions was
at ambient conditions before image 0.007 mm2 (one square pixel, with
Experimental analysis. edge 85 microns).
Each batch of 10 handsheets was
Image analysis of handsheets analyzed using the following proce-
Handsheet preparation dure. Random numbers between 1
Images of the resulting handsheets
Six batches of handsheets were made and 40 were generated; two of these
were acquired using a flatbed scan-
with target impurities of 10, 20, 100, random numbers were assigned to
ner (HP Scan Jet Plus, Model HP
200, 600, and 1000 ppm. The paper each side of each handsheet. The
9195A/AB) at 300 dots per inch
used was a typical office xerographic handsheet sides were then scanned
(about 85 micron dot width) optical
paper, Matrix Xerographic D. P. Hi- in ascending order of their assigned
resolution. The scanner was run
Speed, 75 g/m 2, distributed by number, 1 through 40. This proce-
through a warmup procedure before
Zellerbach, A Mead Co. The paper dure was performed to remove any
use as suggested in our previous pa-
was printed with a standard image auto correlation between successive
per (5). The scanner produced a gray
using Xerox 5100 copier with Xerox scans on a single handsheet. It al-
scale image data file of the handsheet
Dry Ink 5100 toner. lowed for a randomized estimate of
in which each image element, a pixel,
Each batch of handsheets was pre- the error of the scanning.
could have a gray scale value (GSV)
pared as follows. A total of 60 oven- The scanned area was a 3 by 3-in.
between 0 = dark and 255 = white.
dry grams of paper, with a ratio of square approximately centered in
The image analysis was performed
blank sheets to printed sheets that the circular handsheet. To ensure a
using a DOS-based, file-oriented im-
would approximately provide the tar- repeated scan of the same area on
age analysis program, Imagine, de-
get ppm, was mixed with tap water the handsheet, the handsheets were
veloped at the Dept. of Wood and
to achieve a 1.5% consistency. No cut slightly larger than the scanned
Paper Science, North Carolina State
deinking chemicals were used. The area and placed on the same posi-
University. A GSV threshold of 155
mixture was disintegrated using a tion of the flatbed scanner. The cut-
was utilized to distinguish the con-
TAPPI Disintegrator. From this ting of the samples also made the
taminants (GSV < 155) from the
batch, ten 100-mL specimens were samples lie flat on the scanner, a
background (GSV 155). The image
removed and made into handsheets procedure which decreases the mea-
analysis software provided the num-
using standard TAPPI procedures; surement error.
ber of spots detected, the average
172 February 1995 Tappi Journal
Results and the felt side for a handsheet error. (c) The error that is predicted
(Table I). This demonstrates that by Eq. 8 will be referred to as the
Image analysis parameters even inside the handsheet mold, ran- predicted error. To calculate this
dom mixing was observed in this error, the number of spots, the aver-
Handsheets from the 100 and 1000 study. age spot size and the variance of the
ppm target are plotted in Fig. 1. Note spot size determined for each
that the detected area covered by handsheet was determined and sub-
Random distribution of spots
spots is plotted logarithmically and stituted into Eq. 8.
changes over six order of magnitude As mentioned above, it can be shown
that the variance of counted features The predicted error versus the
by changing the gray scale value actual error is plotted in Fig. 7. Data
threshold. This again demonstrates is equal to the number of counted
features (13). For our theoretical for both the felt and wire sides are
how important it is to select a mean- distinguished in the plot; however,
ingful threshold. Since there is no consideration, we applied this law.
It should therefore be possible to the results of the two sides are simi-
standardization available, we de- lar. The predicted error is shown to
cided to use a constant threshold of demonstrate this behavior using the
experimental results. If one plots the be equal to the actual error, i.e., the
155 (indicated by arrows in Fig. 1) data are scattered about a 45 line
which is approximately 75% of the variance of the number of spots per
handsheet for each batch versus the with an intercept of zero. This ob-
average gray scale value of the batch servation lends evidence to the
targeted with 1000 ppm. average number of spots per
handsheet, one should essentially theory used to predict the actual er-
find a straight line at a 45 angle. ror.
Sample concentrations The actual error and the scan-
Figure 6 shows that the observation
of impurities ning error are plotted versus the ac-
of random distribution statistics
The actual ppm concentrations of holds true for our samples. In the tual ppm values in Fig. 8. The results
impurities and the number of spots plot, also included are the confidence demonstrate that the scanning er-
versus target ppm concentrations for intervals of 50% and 80% (two-sided ror is much smaller than the actual
sides of individual handsheets are intervals). It is expected that 50% error at high ppm values. It is im-
plotted in Figs. 2 and 3, respectively. and 80% of the data will lie within portant to note that the scanning
Data acquired on the felt and wire the 50% and 80% confidence inter- error is small and reproducible us-
sides of the handsheets are both plot- vals, respectively, if random distri- ing a single image analysis tech-
ted and may be distinguished by dif- bution is true. nique; however, as demonstrated in
ferent symbols. Figure 2 demon- Fig. 1, if different image analysis
strates that the absolute scatter of techniques are used the differences
Comparison of the predicted
detected ppm and number of spots in results can be significant. Equa-
and actual error in ppm
increases with increasing ppm. How- tion 10 allows prediction of the rela-
ever, the scatter relative to the value Three errors in the area covered by tive error in ppm given the average
of ppm actually decreases with in- contaminants that will be discussed particle size and standard deviation
creasing ppm, as shown by plotting in this section are briefly reviewed of particle size. The prediction of the
actual ppm divided by target ppm in the following. All of these errors relative error versus ppm is plotted
versus target ppm, Fig. 4. are standard deviations and have in Fig. 9. The value of Var(Ai)/AM2
It was of interest to determine if units of ppm. (a) The error that origi- was approximated as 1.5, as found in
significant differences existed in par- nates from the limitations in the re- our data. To demonstrate the sensi-
ticle population between the felt and producibility of the image analysis tivity of the prediction with respect
the wire side. No significant differ- system is referred to as the scan- to this ratio, the curves generated
ence in ppm and number of spots ning error. This error was calculated using ratio values of 0 (i.e.,
detected between the felt and wire by performing image analysis twice Var(Ai) = 0) and 5 are also plotted.
sides is distinguishable in the data on the exact same sample, i.e., two Also plotted in Fig. 9 is the actual
of Figs. 24. Figure 5 is a plot of the scans of each side of each handsheet. relative error of the data. The actual
difference in ppm of the wire minus With these results one can calculate relative error was determined for
the felt side for 10 handsheets of the standard deviation of the mea- each side of the handsheets at each
each batch. If an actual separation sured ppm for a single side of a ppm level by dividing the
was taking place inside the handsheet. (b) The error that was STDEV(AT) by the measured value
handsheet mold, one would expect calculated by simply determining the of AT and multiplying by 100%. The
that the data would cluster either standard deviation of the experimen- predicted relative error adequately
above or below the y = 0 line. More- tally detected ppm of wire and felt fits the experimental results and fur-
over, no correlation was found be- sides of handsheets of a target ppm ther supports the theory presented
tween ppm detected on the wire side batch, thereafter called the actual in this work. Note that both the ac-
Vol. 78, No. 2 Tappi Journal 173
Image analysis
II. Image analysis results on 25% post consumer recycled copy paper
Sheet Side
1 1 10 1 0 0 9.78E-03 6.63E-03 11 18.5
2 1 8 1 0 0 1.19E-02 7.17E-03 9 18.4
3 1 9 0 0 0 7.96E-03 2.39E-03 9 12.3
4 1 11 1 0 0 1.08E-02 8.37E-03 12 22.3
5 1 11 0 0 0 1.04E-02 4.93E-03 11 19.7
6 1 10 2 0 0 1.37E-02 8.35E-03 12 28.3
7 1 8 0 0 0 8.96E-03 5.07E-03 8 12.3
8 1 11 0 0 0 7.82E-03 2.16E-03 11 14.8
9 1 7 0 0 0 8.19E-03 2.71E-03 7 9.9
10 1 13 0 0 0 9.93E-03 4.66E-03 13 22.2
1 2 11 0 0 0 9.78E-03 4.83E-03 11 18.5
2 2 6 1 0 0 1.33e-02 1.05E-02 7 16.0
3 2 8 1 0 0 1.35E-02 9.10E-03 9 20.9
4 2 4 0 0 0 8.96E-03 3.58E-03 4 6.2
5 2 7 0 0 0 9.22E-03 3.50E-03 7 11.1
6 2 12 0 0 0 7.77E-03 2.07E-03 12 16.1
7 2 7 1 0 0 1.08E-02 7.66E-03 8 14.9
8 2 12 0 0 0 8.96E-03 3.24E-03 12 18.5
9 2 19 1 0 0 9.32E-03 5.74E-03 20 32.1
10 2 9 0 0 0 8.76E-03 3.16E-03 9 13.6
35.25 <=SDEV%
(actual)
33.96 <=SDEV%
Predicted Eq. 10
Var( Ai)/AM2
= 0.28
Upper bound: upper border of spot class (in mm2)
Lower bound: lower border of spot class (in mm 2)
No spots were detected with areas greater than 0.05 mm 2.
tual and predicted relative errors suitable for high speed xerographic reported plot size for TAPPI Method
increase dramatically as the ppm equipment and laser printers. The T 437 om-90, Dirt in Paper and Pa-
level decreases toward zero. In con- paper is specified as containing 25% perboard. For this reason the ppm
trast to the relative errors, the abso- post-consumer fiber. Ten sheets of level reported here is expected to be
lute error, SDEV(AT), both predicted paper were analyzed. The same im- higher than that reported using the
and actual, increase with increasing age analysis procedures as described TAPPI method.
ppm, see Eq. 8 and Fig. 8, respec- earlier were followed, including scan- The value of SDEV% of a sheet
tively. ning 3 x 3 in. areas of the sheets. was found to be 35.3%. The value of
The results of the image analysis SDEV% predicted by Eq. 10 with
Analysis of a commercial are shown in Table II. Var(Ai)/AM2 = 0.28, from data in
recycled copy paper The average ppm of the area Table II, was 34.0% (Table II). The
scanned was found to be 17.3. The excellent agreement between the
To further explore the utility of the
ppm was calculated using detected actual and predicted values gives fur-
statistical tools as described above,
spots with sizes between 0.0007 and ther validation of the derivation of
image analysis was performed on a
0.05 mm2; many detected spots were Eq. 10.
commercial recycled content paper
less than 0.04 mm2, the minimum