You are on page 1of 10

1098 IEEE Transactions on Dielectrics and Electrical Insulation Vol. 8 No.

6, Dece>mber 2001
Hydrophobicity Estimation of HV Polymeric
Insulating Materials
Development of a Digital Image Processing Method
M. Berg, R. Thottappillil and V. Scuka
Division for Electricity and Lightning Research
UppsaJa University, UppsaJa
Sweden
ABSTRACT
Image analysis of water drop patters on an inclined fat polymeric insulator surface has been
performed in order to fnd a simple mathematical function that indicates the level of hydropho
bicity of the insulator surface. A simple function, given the acronym average of normalized
entropies (AN E), seems to correlate well with hydrophobicity as defned by the Swedish Trans
mission Research Institute (SIRI) hydrophobicty classifcation. It is a composition of three
other functions, vz. the standard deviation, the Shannon inforation entropy and the 'frac
tion of small differences'. All these are in tur based on the histogram of horizontal nearest
neighbor pixel differences for a given digital greyscale image of a water drop patter. ANE
is fairly independent of illuminaton i ntensi ty (exposure) as well as total gain and offset in a
camera system (linear sensor). The experimental results also indicate that ANE i s fairly inde
pendent of limited changes in the surface inclination, although this needs further investiga
tion. Some of the various pitfalls associated with the photography of water drop patters and
processing of imdges are identifed, together with possible solutions for avoidi ng them.
1 INTRODUCTION
I
JIGH voltage (HV) insulators made of polymeric materials are be
lcorng popular due to their light weight, shatter-proof nature,
and hydrophobic surface. The hydrophobic surface gives them better
electrical flashover characteristics than hydrophilic porcelain and glass
in:mlators when being wet or polluted [1,2]. However, the hydropho
bicity of the polymeric insulators in service is degraded by many factors
such as pollution deposits, surface arcing, and aging. Therefore, peri
odic monitoring of the hydrophobicity of live HV insulators is helpful
for detecting signs of surface degradation [3]. The traditional labora
tory methods of measuring contact angles are not practical in the field,
because they require well-defined experimental conditions like fxed
illumination, optimal view of a single water drop, or small-fat, hori
zontal, surface samples. Furthermore, contact angles are hard to obtain
directly during sub-optimal photographic conditions like those prevail
ing outdoors. The SIRI method [47] offers a simple manual procedure
for obtaining a collective estimate of the hydrophobicity of an insulat
ing surface during field conditions. The drawback is that this method
is relying on human judgment, which can vary from person to per
son, and from time to time. Image analysis on the other hand, may
give objective and more precise estimates. In this paper, we present
some mathematical indicators of average hydrophobicity, based not on
a complicated physical model, but instead on the amollit of variaton
in a digital greyscale image of a waterdrop pattern on an insulating
surface. These hydrophobicity indicator function (HIF) may serve as
prototype tools for estimating hydrophobicity in the field as well as in
the laboratory.
2 IMAGE ANALYSIS AS A TOOL
FOR ESTIMATION OF
HYDROPHOBICITY
Digital photography can improve the quality of field inspection in
several aspects. The object of iterest can be photographed during op
eration from a safe distance. In the case of a power tnmsmission line,
a helicopter with a gyro-stabilized long-range camera could be used.
Another possible soluton is to have permanently installed digital cam
eras deliver images of the insulators to the control room of the utility.
Recorded digital images constitute objective measurements, and they
can be retrieved and analyzed in various ways after the inspection.
The hydrophobicity of an insulator surface can be estimated by nu
merically analyzing digital images of water drop patters. The scien
tifc discipline of image analysis has developed considerably in the last
twenty years, and a wide range of tools is available for the development
of a suitable algorithm. If successful, image analysis will lead to reli-
1070-9878/1/ $3.00 2001 IEEE
IEEE Transactions on Dielectrcs and Electrical Insulation
able and objective estimates of hydrophobicity with a higher degree of
precision and resolution than would be obtained by subjective manual
inspection.
However, it is not a trivial matter to determine the hydrophobic
ity of an insulator surface by analyzing one or a few images. What a
water drop willook like is highly dependent on the background, the
illumination and te direction of observation. The reason is that wa
ter is both transparent and reflecting. Furthermore, hydrophobicity is
normally defined in terms of contact angles, which are hard to measure
during field conditions and need to be determined for a large faction
of the drops on the surface. Therefore, it is better to look for an alter
native definition of hydrophobicity, based on those water drop features
that are visible in the recorded images. The STR] classification method,
described below, serves as a suitable starting point and reference.
3 THE STRI CLASSIFICATION
METHOD
Most accurate methods of measurement are based on stationary de
vices, which require that the insulators be brought to the laboratory.
Therefore, feld inspection of insulators is normally limited to a search
for visual signs of degradation, such as cracks, discoloration or layers
of polluting substances. If water is applied to the surface of a given
insulator, the hydrophobicity can be estimated also. The water drops
then act as probes, because their shapes depend on the local hydropho
bicity. This is the principle behind the STR] classification method [47],
specifcally designed for manual estimation of the hydrophobicity of an
electric insulator in the field. The method is based on two key features:
the receding contact angles of sessile droplets on the inclined surface;
and the area covered by a water film ('completely wetted area').
The STR! method gives a human-oriented defnition of hydropho
bicity that alleviates considerably the perception problem for this par
ticular property of an insulator surface. Still, the typical error is of the
order of one hydrophobicity classes (He).
Te STR! classification procedure is simple. Tap water is applied
to the surface of interest by the help of an ordinary spray bottle. The
tested area should be 50 to 100 cm2. The inspector carefully watches the
resulting water drop pattern from diferent angles, and then attributes
the surface to one out of seven He, ranging from He 1 (hydrophobic)
to He 7 (hydrophilic). The written criteria for the STR! classifcation
are given in Table 1. There is also an accompanying set of important
sample images that represent the different STR! He, but these are not
shown here.
The SIR] classification estimates the hydrophobicity of a certain area
of surface; i.e. it is a collective measure. Hydrophobic surfaces belong to
HC 1 to 3, whereas (partially) hydrophilic surfaces belong to He 5 to 7.
He 4 is an intermediate class, where wetted traces appear together with
discrete droplets. A newly manufactured silicone rubber insulator will
be of He 1, but once in outdoor service, the insulator gradually will lose
part of its hydrophobicity. The speed of degradation depends highly on
the local environment. When HC 4 or 5 is reached, this indicates that the
insulator is becoming hydrophilic, which in turn can be interpreted as a
waring sig: The insulator may have to be restored (cleaned) or even
replaced, especially i it is designed in such a way that the necessary
Vol. 8 No.6, December 2001 1099
Table 1. Written defnitions for the CK He (He 1 to 7). 7`~ receding
contact angle.
1L Description
1 Only discrete droplets are formed. 8r 80 for the
majority of droplets.
2 Only discrete droplets are formed. 8r [50, 80Jo for
the majority of droplets.
Only discrete droplets are formed. 8r E [20, .50] 0 for
the majority of droplets. Usually they are no longer
circular.
4 Both discrete droplets and wetted traces. 8r R 00 from
the water runnels are observed. Com
p
letely wetted areas
O
<2 cm2. Together they cover <90% of the tested area.
Some completely wetted areas >2 cm2, which cover <90%
of the tested area.
6 Wetted areas cover >90%. i.e. small UweUed areas (spots
or traces) are still observed.
Continuous film over the whole tested area.
margin to fashover to some extent depends on the insulator surface
maintaining its hydrophobicity.
4 EXPERIMENT
Image analysis needs digital images. For this reason, an experiment
was conducted, where water was sprayed onto an inclined silicone rub
ber plate and the water drop patterns subsequently photographed. I
order to simulate varying degrees of hydrophobicity the insulator plate
was blasted with fne olivine sand. Other methods such as immersion
in water, or irradiation by corona or plasma could in principle have
been used [8), but sand blasting was more convenient at the time of the
experiment. The plate was blasted during roughly 15 s it a pressure of
0.6 MPa and a distance of 1 U. After this, the plate was thoroughly
rinsed with de-ionized water in order to remove remaining sand parti
cles. The blasting transformed the originally hydrophobic surface (HC
1) into an effectively hydrophilic one (He 7). We believe that the surface
was made hydrophilic primarily through exposure of filler particles
and increased roughness (visible as specks with reduced reflectivity)
[9]. Microscopic sand particles also may have been embedded i the
surface. However, it was not our intention to study the effects of sand
blastng upon the insulator surface, but instead to produce data for im
age analysis.
After the sand blasting, the plate gradually recovered its hydropho
bicity. The initial rate of recovery was about one class in 4 to 5 days,
and then slowing down after HC 3. After about two months, the hy
drophobicity had returned to almost HC 2. This recovery rate is slower
than reported in [10], probably due to shorter blasting distance and a
different silicone rubber formulation. One explanation for the recov
ery of hydrophobicity is that low molecular weight (LMW) molecules
in the bulk diffuse to the surface [8). I our case, exposed hydrophilic
filler particles would be covered eventually by a thin layer of LMW, thus
leading to increased hydrophobicity. However, the hydrophobicity did
not recover completely, most likely due to remaining roughness.
After the hydrophobicity recovered and reached a new, lower He,
the surface was sprayed with distilled water and photographed 2Sx at
10 inclination and 25x at 35 inclnation from the horizontal. These
two inclination angles were chosen as representative extremes for the
1100 Berg et al.: Hydrophobicity of HV Polymeric Insulating Materials
upper surface of a shed on a vertical power line insulator. The hy
drophobicity classification was done according to the SIRI method, but
always at 35 inclination in order to maintain consistency.
The first pattern during a session was preceded by a prolonged
spraying during 30 s in order to apply water to the entire surface.
The spraying was done in two steps: First the previous drop pattern
was removed effectively by spraying the whole surface at close dis
tance, 50 Then a new drop pattern was applied by spraying at
a distance of 300 , until the first drop ran off the plate. H C 1 (a
virgin, hydrophobic surface) was treated differently because a running
drop tended to absorb most of its neighbors below. Two sets of drop
patterns were recorded for the sake of comparison, one set with run-off
and one without run-off.
5 EQUIPMENT
The setup consists of a video camera, a industrial type personal
computer equipped with frame grabbers, a focused spotlight, and a
tiltable platform, on top of which an insulator plate of interest is placed.
Te water drop patters are produced by an ordinary household spray
bottle containing distilled water. The spotlight and the camera were
placed close to each other, both pointing towards the insulator plate.
They were tilted by 25 from the horizontal in order to simulate obser
va hon by helicopter.
The insulator samples used in te experiment are thin circular plates
made of grey-colored Powersil311 (Wacker Chemie, Germany), a high
temperature vulcanizing silicone rubber (HTVSIR) based on polydi
methylsiloxane (PDMS). The filler content is 100 pph alumina trihydrate
(AT). The rubber plates have a diameter 220 and thickness 2 .
A insulator sample is placed on the platform, which can be tilted 0
to 45 from the horizontal. The relatively uniform illumination is pro
vhled by an ETC Source Four focused spotlight, equipped with a 575 W
tungsten halogen lamp. A Jai M10RSC video camera produces 50 full
greyscale images per second due to its non-standard design with two
amlog outputs. The camera sensor is of the charge coupled device
(CCD) type. The water drop patterns were photographed using a Rain
bow zoom video lens of focal length 8.5 to 51 |, giving an effective
horizontal feld of view of 150 at 1.1 m working distance. The
computer of the video system, Process Vision DVSP/l0, is specially
designed for remotely controlled video recording in industrial environ
ments. I is equipped with two lmagenation PX600 frame grabbers for
real-time digitization of the two analog video signals from the camera.
The 8 bit greyscale images are of the size 768 (W) x 576 (H) pixels. For
subsequent image analysis, the image fes are transferred to a Digital
AS400 workstation. The main software tool for the computations has so
far been Matlab v. 5 .3 and its image processing toolbox v. 2.2, together
providing a simple user-friendly environment for image analysiS.
6 RECORDING AND NUMERICAL
PRE-PROCESSING OF
DIGITAL IMAGES
The spotlight and the camera system were turned on half an hour
before each recording session in order to let them warm up. Mean
while, the relatve orientation of camera, platform and spotlight were
fe-checked and adjusted.
When the equipment had reached steady state, the camera was fo
cused and the dual outputs matched with respect to gain and offset by
using a specially designed computer interface, developed by the first
author. This interface is a display window (including a zoom window)
showing the recorded images in real time together with calculated val
ues of so-called 'target functions' for focus (the norm of the average lo
cal gradient of pixel values) and camera output channel mismatch (the
sum of the squared differences of frequencies i the two histograms of
pixels coming from different channels). For further details, see [ll]. All
other system parameters were fixed during the entire experiment: The
settings were such, that the greyscale distribution would peak roughly
in the middle of the allowed range from 0 ('black') to 255 ('white'),
while at the same time avoiding excessive nonlinear clipping due to di
rect reflection in the water drops. The exposure time was deliberately
set to 0.1 ms in order to simulate recording of water drop patterns in
a 50 Hz electric field, where motion blur must be eliminated from the
images . As a consequence, the amplifcaton in the camera setup was
larger than would have been necessary during laboratory conditions
with stationary drops. The noise in the recorded images is still fairly
small, with a standard deviation of (2 0.2) grey levels (approximate
Gaussian distribution).
The water drop patterns were highly variable, although of the same
qualitative appearance for a given HC and surface inclination. In order
to avoid biased results, great care was taken to reduce correlation be
tween consecutive drop patterns. The insulator was sprayed intensively
at close distance before applying a new patter. Just after the spraying
stopped, the water drop pattern was allowed to settle for 10 s before it
was recorded as a short sequence of 24 images, out of which one image
was saved. Not less than 25 water drop patterns were recorded for each
combination of HC 1 to 6 and inclination angle (10 and 35).
After reCIdiOg session (one per HC), the images were transferred
to a workstation for subsequent analysis. The images were then pre
processed in order to attenuate the noise and the error due to the small
remaining output channel mismatch in the camera sy,tem. The pre
processing algorithms, developed by the first author, were designed to
affect the true image information as little as possible: The remaining
channel mismatch was reduced by applying a local cnvolutoH filter
that eliminates the difference in local average grey level between the
video channels. The noise was attenuated by applying a variable lo
cal weighted average filter (3x3 matrix), such that the local variance of
the grey level is reduced by the estimated noise variance. For further
details, see [11]. This filter can be regarded as a modification of a pix
elwise Wiener filter described in [12]. The reason for developing a new
filter was that the above Wiener filter, implemented in MATLAB v. 5.3,
was smoothing/blurring the drop details too much.
7 IMAGE ANALYSIS
More than 300 images were made during the cOUl'se of the exper
iment, corresponding to minimum 25 images per combination of in
clination angle and HC, except for He 5/35 that was missed due to
temporary equipment problems (while the hydrophobicity continued
to recover). As opposed to other HC, HC 1 images only include patterns
without bare patches caused by drops that have run off the insulator
plate (see Sction 8 below). He 7 was excluded because it represents
IEEE Transactions on Dielectrics and Electrical Insulation
the uninterestig case of a completely hydrophilic surface. Only the
central part of each original image was used in the image analysis, i.e. a
rectangular region of 568x330 pixels corresponding to a surface width
of around 110 mm. This was done in order to exclude the edge of the
insulator plate and areas slightly out of focus.
Figure 1. Water drop pattern on m insulator surface of ML 1 inclined
by 100.
Figure 2. Water drop pattern on an insulator surface of MI 1 inclined
by 35D
Figures 1 and 2 show two different drop patterns on a hydrophobic
surface of He 1 at the inclinations 10 and 350, respectively. In order to
show the details bttteI onl
y
the central part (270x202 pixels) of each
original image is included. (The contrast and brightness have been ad
justed for printing purposes). The drops have shapes that approximate
spherical caps, thereby acting as optical lenses that make the inner parts
of the drops appear brighter than the evenly grey insulator surface.
Note also, that the difference in inclination also corresponds to a differ
ence in perspective, since the camera was held at a fixed inclination of
250: More drops fit into the same view at 100 than at 35. Furthermore,
the drops reflect more light and the surface less light at 100 than at 35.
Figure 3. Water drop pattern on insulator surface of He 4 inclined
by 100
Figures 3 and 4 show drop patterns on a surface of He 4, repre
senting an intermediate state of hydrophobicity according to the STR!
classifcation. The drops are fewer, larger, flatter and more irregular in
shape, and a small fraction of the surface is covered by wetted traces.
For this He, refections inside the drops do not give a significant contri
bution to the variation in grey levels in the image. The drops are larger
and fatter at 10 than at 35, because they can grow larger before slid
ing off. Te difference in perspective leads to fewer direct reflection
Vol. 8 No. 6, December 2001 1101
Figure 4. Water drop pattern on an insulator surface of He 4 inclined
by 35.
spots at 10 than at 35. At He 7, the entire area of intereSt is covered
by a water fl. The sprayed surface appears as a more or tess uniform
patch of grey, without any significant features. The main effect of the
inclination is a moderate difference in average grey level due to the
angular dependence of the reflectivity of the insulator surface.
The purpose of the image analysis was to find a relatively simple
monotonic mathematical functon that would correlate well with the
SIRI classifcation, while being fai1ly robust with respect to variations
in the inclination angle of the insulator. This function, henceforth de
noted hydrophobicity indicator function (HIP), should be decreasing
with increasing He. This paper is the condensed result of many trials .
Only the best HIF found so far is presented here, together with its com
ponents. The paper is a continuation of [13], where the frst attempts at
finding good HIF are reported.
The overall theme for the search of candidate HIF was variation in
pixel values, since a hydrophobic surface forces the water drops to be
come rounded lenses, which refect and refract the incoming light. The
opposite is true for a hydrophilic surface: the pixel values vary rela
tively little as long as the insulator surface is evenly colored.
7.1 THE GREY LEVEL
DISTRIBUTION
The grey level distribution is perhaps the simplest way of character
izing a digital greyscale image. It is a function that maps each discrete
grey level E {O, 1, 2: . . . , 255} to the fraction of pixels in the image
having this same integer value. The grey level distribution is never
theless a valuable tool for comparing different water drop patterns, be
cause it is effectively sensitive to the hydrophobicity via the presence
or absence of optical effects produced by water drops. Furthermore,
it reduces the amount of calculations dramatically, since no extensive
neighborhood operations are needed.
Figure 5(a) shows a small portion of a water drop pattern on a flat
HTVSIR surface of He 1, inclined by 35. The water drops are isolated
and rounded in shape, thereby making them act as lenses. The corre
sponding grey level histogram is given in Figure 5(). The central peak
represents the evenly grey insulator surface, whereas the tails repre
sent the darker drop contours (no shadows) and the brighter interal
refections, respectively.
7.2 GREY LEVEL DIFFERENCES
However, it was found out that a histogram based on differences
of grey levels offers significantly increased sensitivity to the He. Fig
ure 6(a) shows a synthetic image, calculated from the original image
110'2 Berg et al.: Hydrophobicit of HV Polyeric Insulafing Materials
0.2
0.18
0.16
0.14
f
O.12
g
.. 0.1
"
.

O
.
0
8
0
0.06
0.04
0.02
"
L--
50
-1 00150
=
2 OO
-
2

50
(b)
f0yl0V6| [0 -0|30K,255 -W0I
Figure 5. (a) Detail of a water drop pattern image (He 1, inclination
35). () Grey level histogram corresponding to the water drop patter
in (a).
in Figure 5(a) as the difference between each pixel and its right-hand
neighbor. I order to make a displayable image, the interval of differ
ences [-255, 255J was mapped linearly onto the interval [0,255]. The
difference image is mostly mid-grey with little contrast, because most
differences are close to zero. This is also reflected by the sharply peaked
histogram in Figure 6(b). Significant deviations from zero are caused
mainly by edges associated with drop contours and direct reflection
spots.
There are also other good reasons for using horizontal nearest
neighbor pixel differences:
1. The unwanted sensitivity to the inclination of the insulator surface to
wards the camera is reduced.
2. Horizontal nearest-neighbor differences are among the simplest edge
detectors, and quick to calculate [14].
3. The influence of sligtly non-uniform illumination is reduced, because
nearest neighbor differences are not sensitive to slow spatial variations
in the background grey level.
In principle, it is also possible to normalize the grey level distribu
tion so that it is mapped onto the fl range of [0,255]. However, the
test results were not satisfactory, so this idea was abandoned at an early
stage. Instead, we sought other ways to normalize the output without
ad verse side effects.
Next, we proceed by presenting a few candidate hydrophobicity
indicator function (HlF), which are calculated from the histogram of
0.18
0.16
0.14
f
O.12
C
g 0.1

O.OB
u
0.06
0.04
0.02
0
(b)
-200 -100 0 100
,f0yl0V0|R0t0nO0
200
Figure 6. (a) Horizontal difference image of the water drop patter
in Figure 5(a). The contrast has been exaggerated for the sake of print
ing quait. (). Grey level histogram corresponding to the difference
image in (a).
horizontal nearest-neighbor differences. All HIF are presented as aver
ages over 25 different images for each He, because the variability of the
drop patterns tured out to be both significant as well as hard to control
with the simple means used in accordance with the SIRI classification
method.
7.3 STANDARD DEVIATION OF
GREY LEVEL DIFFERENCES
When the hydrophobicity increases, the spread of grey level dif
ferences increases as a result of the water drops acting as lenses and
partially reflecting mirrors. A simple measure of spread is the standard
deviation O of the horizontal grey level differences. In Figure 7, the
unbiased estimate of O is plotted as a function of He for the inclinations
10 and 35, respectively. Each point in the plot is the average of over
25 images of different drop patterns at the same He. The error bars
correspond to one standard deviation. As a HIP, O is not very good.
I is strictly decreasing with increasing He, but it depends too much on
the inclination angle. Both curves in the Figure are nonlinear, which
is to be expected, since the SIRl classification by its very definition is
nonlinear (see Section 3 above). Note also, that the spread of de
creases with increasing He: At He 3 to 7, there are no longer any strong
lenses among the drops, which have partially turned into water film. A
hydrophilic surface covered by a water film will thus not produce any
large deviations from the grey levels of the background. Furthermore,
IEEE Transactions on Dielectrics and Electrical Insulation
it is evident that noise in the images must be reduced if HC >3 is to be
detected reliably. (The standard deviation of differences of grey levels
would for this experiment be ' 3 without noise reduction).
Figure 7. Standard deviation of horizontal grey lcvd differences v
,according to SIRI. Surface inclinations: 10 (solid line), 35 (dash
dotted line).
7.4 ENTROPY OF GREY LEVEL
DIFFERENCES
The more variation (apart from noise) there is in an image, the more
information the image contains, at least in a mathematical sense . Im
ages of sprayed water drop patters will contain more information if
the insulator surface is hydrophobic, since rounded drops (given the
geometrical configuration of the setup) will act as lenses that produce
intensity deviations from the evenly colored background. A rough es
timate of the amount of information in an image can be obtained by ap
plying information theory [15, 16]. For the sake of simplicity, the image
is regarded as a collection of symbols transmitted through a hypotheti
cal one-dimensional communication channel. Each symbol corresponds
to a possible pixel value, here assumed to be a horizontal grey level
difference. The distribution of pixel values is taken as an estimate of
the probability distribution for the available set of 511 symbols . The
estimated average amount of information per symbol is then given as
the Shannon information entropy E
2OO
E = - ' f(d) log2 f(d) (1)
d=-
2
55,!"O
where f(d) are the relative frequencies of the grey level differences d
in the image. Note that only non-zero frequenCies are included, due to
the presence of a logarithm.
The word 'entropy' is borrowed from statistical physiCS, mainly be
cause Equation (1) has the same form as certain formulations of entropy
but the entity as such is meant to be a measure of information [15]. The
Shannon entropy is now a well-established tool within image analysis.
If only one grey level difference is represented in the image, then E=O
bit/pixel. The maximum value of E, " 9 bit/pixel, is obtained when
the difference values are evenly distributed. The above formula is only
a frst order (over)estimate of the ideal amount of information in a dig
ital image. More advanced versions of image entropy exist, but they
Vol. 8 No.6, December 2001 1103
require more computation. The tested alteratves to E did not offer
any signifcant advantages! so they were abandoned.
D
.

...

.
Hydrophobicity C|33
Figure 8. Entropy of horizontal grey level differences S. L. Surface
inclinations: 10 (solid line), 35 (dash-dotted line).
Figure 8 shows how the entropy E varies with the HC for the surface
inclinations 100 and 35, respectively. The two curves are fairly close
to each other, which means that E is less sensitive to the inclination.
Furthermore, the spread of E for a given HC is smaller than for above,
and the sensitivity (slope) for higher He is somewhat better too.
7.5 INFLUENCE OF GAIN AND
OFFSET
The HIF ideally should be independent of system gain and offset,
because these are controllable but not easily measurable parameters
that depend not only on the electronic circuits in the camera system,
but also on the intensity of the illumination, the exposure time and the
lens aperture. Changing the offset (brightness) corresponds to adding
a constant to all grey levels. The simple solution is to base the HIF
either on grey level differences or on teir relative frequencies in te
histogram. To increase the gain (contrast) means that all grey levels
are multiplied by the same positve factor> 1. This leads to an increase
in both the spread and the average of the grey level distribution. The
problem is that an increased system gain could be interpreted falsely
by an HIF as increased hydrophobicity. Therefore the HIF should be
designed in such a way that the gain dependence is effectively canceled.
When digitization effects (noise or saturation) are small enough!
both and E are approximately independent of offset, because O is
based on differences of grey levels, whereas the relative frequencies
in the formula for E are independent of the addition of d arbitrary
constant to all grey levels. The standard deviation is proportional to
the system gain (amplifer gainx exposure); A multiplication of all
grey level differences by a factor G leads to a new standard deviation
0* :: GO. I the case of the entropy we instead get an additive loga
rithmic dependence,
E* :: E + log
2
(G)
(2)
The proof is based on treating E as an approximate integral with re
spect to continuous grey levels.
1104 Berg et al.: Hydrophobicit of HV Polymeric Insulating Materals
7.6 SCALED ENTROPY
The gain dependence can i principle be approximately canceled by
subtracting log2 (C) from E. Then a gain C would lead to a new value
E* log
2 (*) E + log2 (C) log2 (CC)
= E log2(C)
(3)
However, this idea did not give good test results, so it was abandoned
in favor of a slightly different function with better properties, given the
name of 'scaled entropy' Es
7OO
Es = L af(d) log2[af(d)]
(
4
)
d=-
2
55J"O
Note that the function can be rewritten as Es=a(E log2(a)). The
scaled entropy Es combines three simultaneous and interconnected
trends: The variation of grey level differences, the presence of local gra
dients and the information content in the image. All these increase with
hydrophobicity. The scaled entropy Es is plotted o. He i Figure 9.
The surface inclination dependence is markedly smaller for He 2 to 6
when compared with the entropy E. The price is that the sensitivity
is rather low for these He. At He 1 the curves deviate from each other,
because the internal reflections from the water drops are much stronger
at 10 than at 35 due to different relative viewing angles. Furthermore,
E _ is still proportional to the gain, since there is an extra factor a in
the formula.
20
!18
"

pA
"
14

j 12

:10
0
>
c

"
&
r 4
0
2
0
0
I
Z 4
|C|0D0D|C|\yC'866
Figure 9. Scaled entropy of horizontal grey level differences VS. He.
Surface incliations: 10 (solid line), 35 (dash-dotted line).
7.7 FRACTION OF SMALL
DIFFERENCES
In order to resolve the gain dependency problem, it is necessary
to proceed further. Both the standard deviation O and the entropy E
increase with hydrophobicity. By instead focusing on the absence of
variation in the difference image, another type of HIF was found, the
fracton of small differences 4, defined as the sum of relatve frequen
cies of the smallest grey level differences in the drop pattern image
n
4 = L f(d) (
5
)
d=-1
The threshold n is a small positive integer E [0, 255]. In practice, n=l
gave the best results, probably because this threshold roughly coincides
with the remaining noise level. The fraction of small differences is
plotted VS. HC in Figure 10. It is a function that increases monotonically
with He, since the more hydrophilic (less hydrophobic) a surface is, the
larger the fraction 4 becomes. The curves are qualitatively similar to
although upside down. At HC 1, there still is a significant difference in
the response to different inclinations. One reason for this is a perspec
tive effect: More of the lens-like drops fit into the camera view at 10.
4 is a very suitable indicator in that it is simple, and by its definition
resticted to a normalized range of [0, 1].
O.9
08

0.7
t
. 0.

0.5
w
0.4
<
0.3
.
0.2
0.1

0 o b
HydtOhObI0|lyC|a33
Figure 10. Fraction of small horizontal differences VS. He. Surface
inclinations: 10 (solid line), 35 (dash-dotted line).
7.8 NORMALIZED ENTROPY
The fraction of small differences now can be used to normalize te
entropy E in order to cancel the gain dependence. This new, composite
HIF is given the name of normalized entropy En
E
n = - f f(d) log2
(I(
d
) )
(6)
U=- 255,f"O
1
where the fraction of small differences 4 is given by Eq uation (5). Note
that the normalization of the entropy is accomplished by dividing the
relative frequencies f (d) by 4, but only inside the baBe 2 logarithms.
I Figure 11, the normalized entropy En is plotted '. the He at the
surface inclinations 10 and 35, respectively. The nonlinear response
remains as expected. The two curves now coincide well for He 1 and 2,
at the expense of some difference for higher He. The invisible change is
that the normalized entropy is approximately independent of offset and
gain, in contrast to the previously presented HIF. Without round-off
errors from the digitization, the independence of gain end offset would
be exact.
7.9 AVERAGE OF NORMALIZED
ENTROPIES
With three basic HIF (a, E, 4) it is natural to investigate various
gain-independent combinations. The previously mentioned idea that
E can be normalized by subtracting log_(C) from E turned out to
IEEE Transactions on Dielectrics and Electrical Insulation
3.
_<.O
g
g Z

'5

U.
----3--4----
Hydrophobicity 6|999
Figure 11. Normalized entropy of horizontal differences >. !L. Sur
face inclinations: 10 (solid line), 35 (dash-dotted line).
L---
2
-3-
4
--
5
--
6
-
Hydrophobicity 0|aS9
Figure 12. Complementary normalized entropy of horizontal differ
ences :. !L. Surface inclinations: 10 (solid line), 35 (dash-dotted
line).
improve considerably if the expression also was multiplied by q1. This
new function is given the name 'complementary normalized entropy'
2O
Een = -0 L [o}(d)] log2 [I! (d)]
(7)
d=-255,!#O
Een also can be interpreted as the product of the fraction of small dif
ferences and the scaled entpy Es. The function is plotted in Fig
ure 12, where the curves for the inclinations 10 and 35 coincide well
at He 1 to 2, but deviate at He 3 to 5. In the latter interval, the differ
ence between the inclinations is roughly opposite to the corresponding
difference for the normalized entropy En. Therefore it was natural to
try the average of En and Een. This is the last and the best HlF found
so far, giVen the name 'aerage of normalized entropies' A
A = (E" + Ec) (8)
where En and Een are given by Equations (6) and (7), respectively.
A is plotted b. HC in Figure 13. The function is strictly decreasing
Vol. 8 No. 6, December 2001 1105
with respect to He, with only a small sensitivity to the inclination of
4.5
u
2
:.3.5
0
.
i
w

2
,5

c Z
g
1
w
A

0.
d J 4
Hydrophobicity class
7
Figure 13. Average of normalized entropies 05. He. Surface inclina
tions: 10 (solid line), 35 (dash-doted line).
the surface within the range of [10, 35]. At HC 3 there is a knee,
separating two almost linear sections. This is not surprising, because
the STRI classification is a composite one, based on the receding contact
angles for HC 1 to 3, and the relative amount of wetted patches for He
4 to 7. Finally, A is approxiately independent of the gain (including
exposure) and offset in the photographic setup.
8 RESULTS AND DISCUSSION
The experiment was designed to simulate certain aspects of a field
inspection of power line insulators by helicopter, i.e. short exposure
time in order to avoid motion blur, the limited contrast, and the typical
relative angles between the camera, the insulator and the source of illu
mination. In other words, the intention was to fnd out if simple image
analysis could be used under less than optimal conditions.
The pre5ented hydrophobicity indicator function (HlF) are mono
tonic functions of the He (according to the STR! method). The sensitivity
is largest for He 1 to 3, which is to be expected, since these H!F require
variations in the images, variations that are produced by lens-like drops
rather than by wetted patches, given the geometrical Cnfguration of
the setup during the experiment.
A better sensitivity for HC 4 to 7 would require another camera view,
where specular refections in wetted patches can be detected. However,
it is not easy to maintain such a geometrical relation between the cam
era, the illumination and the insulator surface in the feld. Instead, the
orientation angles were chosen to simulate a hypothetical inspection of
a powerline by helicopter, where the source of illumination is on board
the helicopter and the incoming light rays are nearly anti-parallel to the
refected/recorded light rays. As a consequence, the sensitiVity of the
presented HIF to flat wetted patches on partly hydrophilic surfaces is
lower.
Any measure of hydrophobicity can be used as calibration reference
for A, as long as corresponding images of water drop patters are avail
able. Here, the STRI classification method was employed as a referLnLe,
because it was simple to use. It would have been much more cumber
some to measure contact angles at many points on the surface. The
1106 Berg et al.: Hydrophobicity of HV Polymeric Insulating Materials
recorded image sets are not ideal representatives of the HC, but they do
correspond to a strict decrease of hydrophobicity.
-
The input to the HIF is a single greyscale image. However, it was
quickly discovered that averaging over several images is needed in or
der to get reasonably precise HIF values due to the variation between
drop patterns for a given He. The problem is that it is impossible to
repeat the positions and sizes of the droplets from one pattern to an
other when spraying water on an insulator. Also, the hydrophobicity
can vary over the surface, thereby increasing the patter variation.
A particular problem for hydrophobic surfaces is run-off: when the
spray continues to add water to the surface, drops already on the sur
face grow. When a drop is large enough, it runs off the insulator plate
and effectively absorbs most of the drops below it, thereby creating an
uncovered patch. Without water on this patch, there are no 'probes'
left, and the local absence of variation in the image shifts the HIF value
in the wrong direction. I practice this was mostly a problem with He
I, for which it was decided to test two alteratives; either to ignore
rurl-off, or to stop spraying just before run-off occurred. With run-off,
the variance of the HlP values increased, but their averages at the incli
nations 10 and 35 were closer to each other. Run-off was allowed as
part of the spraying process for all other HC during the experiment.
The STRI classification method is qualitative and intended for a hu
man observer, who can view the drops in three dimensions from differ
ent directions and at close distances. The inclination of the surface is
only mentioned as a factor that infuences the receding contact angles
of water drops, but is not accounted for as such [4]. The experiment, on
the other hand, showed that the inclination must be taken into account
in order to get useful results from image analysis. The shapes and the
maximum volumes of the drops, and also the number of drops that fit
into the image (with a fixed camera) all depend on the inclination of the
surface. Hence, also the image contents are highly dependent on the
surface inclination, which is why water drop patterns were recorded
at two different surface inclinations for comparison, and why consid
erable effort was put into finding a HIP that is reasonably robust with
respect to inclination.
However, larger inclinations than 35 will also need to be investi
ga ted, as outdoor insulators have variable geometries and orientations.
Drops hanging from the undersides of weather sheds probably have to
be treated separately, because their shapes are in general different from
those of sessile drops. On the other hand, ANE may work rather well
also for surface inclinations : 90, since the variations in drop size and
drop shape are relatively limited already at 35. Furthermore, the in
clinations of the camera and the illumination lamp can be adjusted to
maintain a relative view close to the ones used in this paper.
The brightness and contrast of the recorded images are controlled
by the operator, who can adjust freely the intensity of the illumination,
the lens aperture, the exposure time, and also the electronic gain and
ofset in the camera system. All these parameters in principle can be
measured, but this is not a practical approach. Instead, it is better to
look for HlF that are approximately independent of system gain and
offset, so that the spread of indicator values is reduced. The functons
En, Een, and A meet these requirements, as long as digitization effects
(noise, saturation) can be neglected.
Probably the best approach to develop image analysis algorithms for
hydrophobicity classifcation is to look for purely geometrical features
of water drops while trying to compensate for, or adapt to, variable ex
perimental conditions. Two attempts in this direction have been made
by Tokoro et al., who calculated the size (area) distributions and aver
age circular form factors Fe of sprayed water drops on s ilicone rubber
sheets
Fe =
41a
(9)
L
with U the area and I the circumference. The contact areas of the drops
tend to increase when the hydrophobicity of the rubber and/or the ef
fective surface tension of the drops decreases [17]. Water drops on sili
cone sheets aged by immersion in distilled water also were more irreg
ular (smaller Fe) when compared to the drops on virgir, sheets. How
ever, no information is given on how the edges of the water drops were
detected, a difficult image analysis problem that can be tackled in many
ways. I [18], the dependence of water drop shapes on the inclination of
the insulator surface was largely avoided by only studying very small
droplets, which are not much influenced by gravity. The drawback of
this approach is that the working distance of a microscope often is too
short for field use, unless the powerline is temporarily switched off.
9 CONCLUSI ONS
N
UMERICAL image analysis of applied water drop patterns can be
used to assess the hydrophobic state of a polymeric insulator sur
face. Because this is a new research topic, we decided to start from the
beginning by adopting a simple approach, i.e. to fd a simple real
valued function of a digital image of a drop pattern on an inclined flat
surface, such that it is strictly increasing with average hydrophobicity.
The best result so far is a HIF given the name of average of normalzed
entropies (ANE), (see Equations (6) to (8)), which, according to experi
ment, correlates well with the DK hydrophobicity classifcation.
ANE is designed to be approximately independent of gain and offset
in the photographic setup. It is also robust with respect: to small varia
tions in the inclination of the surface, at least for the confguration used
in the experiment (camera and spotlight tilt 25, surface tilt 10 or 35).
ANE and the other presented HIF are based on qualitative physical
reasoning about water drop shapes and their optical effects. The ad
vantage of ANE over the STRI method is that it provid,es a continuous
numerical estimate of the hydrophobicity instead of discrete class.
Furthermore, ANE is potentially suitable for remote monitoring of out
door insulators.
Future image analysis algorithms for field estimation of hydropho
bicity are not likely to be based on detailed physical models, which
require precise experimental conditions and an enormous amount of
computational effort. Instead, the development should be focused on
general image analysis methods while at the same timE considering the
special properties of drop pattern images.
ACKNOWLEDGMENT
'The work has been financed by Elforsk, within the frame of project
no. 3134 i the Elektra program. The authors thank SIRI (Ludvika,
Sweden) for the use of certain equipment, and in particular Anders
Eklund for his advice on the STRI hydrophobicity classification method.
IEEE Transactions on Dielectrics and Electrcal Insulation
Finally, the authors thank Dan Windmar and Henrik Hillborg at ABB
Corporate Research for providing insulator samples.
REFERENCES
[1] R. Hackam, "Outdoor H Composite Polymeric Insulators", IEEE Transactions on
Dielectrics and Electrical Insulation, Vol . 6, No. 5, pp. 557-585, 1999.
[2J R. S. Gorur, "Status Assessment of Composite Insulators for Outdoor H applica
tions", Pruc. of the 5th Interational Conference on Properties and Applications of
Dielectric Materials, Vol. 1, pp. 35-38, 1997.
[3J D. Birtwhistle, P Blackmore, A. Krivda, G. Cash and G. George, "Monitoring the
Condition of Insulator Shed Materials in Overhead Distribution Networks", IEEE
Transactions on Dielectrics and Electrical Insulation, Vol. 6, No. 5, pp. 612-619, 1999.
[4] STRl Guide 92/1, Hydrophobicity Classication Guide, 1992.
[5J A. Erikssun and D. Wikstrom, "Influence uf the Hydrophobicity of Outdoor Insu
lators on the Flashover Voltage", Proc. of Nord-IS (Nordic Insulation Symposium)
1994, pp. 39-51, NTH, Trondheim, Norway, 1994.
[6J R. Hartings, "Hydrophobicity of composite insuators: Measurement and inJ1uence
on flashover performance", Stockholm Power Tech Interational Symposium on
Electric Power Engineering, IEEE, New York, Vol. 1, pp. 246-51, 1995.
[7J S. Gubanski and R. Hartings, "Swedish Research on the Application of Composite
Insulators in Outdoor Insulation", IEEE Insulation Magazine, Vol. 11, No. 5, pp.
24-30, 1995.
[8J H. Hillborg and U. W. Gedde, "Hydrophobicity Changes in Silicone Rubbers", IEEE
Trans. on Dielectrics and Electrical Insulation, Vol. 6, No. 5, pp. 703- 717, 1999.
[9J X. Wang, S. Kumagai and N. Yoshimura, "Fractal Analysis onthe Recovery of Con
taminant Propeities of Silicone Rubber Insulator Agaist Rain", Proceedings of 1998
International Symposium on Electrical Insulating Materials, pp. 61 9-622, 1998.
Vol. 8 No. 6, December 2001 1107
[10J A. Eklund, I. Gutman and R. Hartings, "Conditiouing of Silicone Rubber Insula
tors: Loss and Recovery of Hydrophobicity", Nit Interational Symposium on
H Engineering (ISH), Ins!. H Eng., Graz, Austria, Vol. 3, pp. 3236/1-4, 1995.
[l1J M. Berg, Estimation ofHydrophobicity ofInsulating Suraces by Studying Sessile Water
drops, Acta Universitatis Upsaliensis, Vppsala Dissertations from the faculty of Sci
ence and Techology 33, Uppsala, Sweden, 2001 ISBN 91-554-5060-1.
[12J J. S. Lim, Two-Dimensional Signal and Image Processing, Englewood Cliffs, NI: Prentice
Hall, pp. 536-540, 1990.
.
[13J M. Ber
g
, R. Thottappillil and V Scuka, "A Digital Image Processing Method for Es
timating the Level of Hydrophobicity of H Polymeric Insulating Materials", 1999
Conference OU Electrical Insulation and Dielectric Phenomena (CIDp), Vol. 2, pp.
756-762, 1999.
[14J J. c. Russ, The Image Processing Handbook, 2nd ed., CRC Press, pp. 233-249, 1995.
[15J C. E. Shannon, "A Mathematical Theory of Communication", Bell System Technical
Joural, Vol. 27, pp. 379-423 and 623-656, 1948.
[16J R. C. Gonzalez and R. E. Woods, Digital Image Processing, Addison-Wesley Ch. 6, pp.
324-343, 1992.
-
[17] T. Tokora, M. Nagao and M. Kosaki, "Image Analysis of Hydrophobicity of Silicone
Rubber Insulator", 1999 Annual Report Conference on Electrical Insulation and Di
electric Phenomena (CEIDP), IEEE, Piscataway, N, USA, Vol. 2, pp. 763-6, 1999.
[18J T Tokom, M. Nagao and M. Kosaki, "Image Analysis of Hydrophobicity of Polymer
Insulators by Observing at Microscopic Area", 2000 Annual Report Conference on
Electrical Insulation and Dielectric Phenomena (CEIOP), IEEE, Piscataway, Nj, USA,
Vol. 1, pp. 226-9, 2000.
Manuscri
p
t was received on 15 December 2000, i fal for 27 September 2001.

You might also like