|
Items in cart [0] |
Questions? Call 888-624-8373 |
| Copyright © 2009. National Academy of Sciences. All rights reserved. Terms of Use and Privacy Statement |
Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter.
Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 493
Automatic Particle-Image Velocimetry
Utilizing Laser-Induced Fluorescent Particles
T. C. Fu, R. Bing and J. Katz
The Johns Hopkins University
Baltimore, USA
T. T. Huang
David Taylor Research Center
Bethesda, USA
Abstract
Microscopic, neutrally buoyant particles containing
fluorescing compound have been adapted as tracers for
velocity measurements of large scale turbulent flows. This
technique consists of illuminating a thin slice of the flow
field with a laser sheet, which is pulsed following a specific
illumination code. The multiple exposure image is recorded
on photographic film and later enhanced while being
digitized. Algorithms have been developed for analyzing
the resulting images. They rely on the illumination code and
the particle streak morphology in order to identify and
compute the tracer velocity vectors. A one-inch diameter jet
has been used as a flow field for preliminary tests.
Introduction
The present paper focuses on the development of a
quantitative flow visualization system which is particularly
suited for large scale towing tank experiments. Until recently,
flow visualization has been utilized for providing only
qualitative information, while quantitative data, namely
the velocity field, has been determined by single point
measurement techniques (hot wire anemometry, laser doppler
velocimetry, etc.). Due to their nature, as well as cost, these
techniques are limited to simultaneous sampling at a few
spatial locations.
Early experiments with quantitative flow visualization
have been performed by Kobayashi [1] and Marko & Rimai
[2]. They have all used long exposure photography to record
both the position and mean velocity of passive tracers within
a fluid. High speed photography has also been adopted to
record time series of tracer particle positions (Racca & Dewey
[3]). A further refinement has been to identify particles in
successive frames, and reconstruct the velocity with a time
base equal to the framing time (Racca & Dewey [3]). Another
approach has been to record multiple exposure images and
measure the displacements of small suspended particles to
obtain full-field velocity maps (Adrian [4]). Gharib &
Willert [5] have performed a similar analysis. However,
they have used a single extended exposure of particle streaks
with prescribed variations in the intensity of the light
emitted from each particle. The variation in intensity was
achieved by utilizing fluorescing particles and by varying
the wavelength of the illumination light. Finally, one
should mention the work of Khalighi [6] who utilized digital
image processing techniques to automatically analyze
particle streak images and produce full-field velocity maps.
The primary factors affecting the capability to adapt
particle displacement velocimetry to large scale complex
flows are the ability to handle large amounts of data, the
required processing time, variations in velocity scales in the
same flow field, the capability to record and identify fine
details in large scale images as well as the availability of
the appropriate particle tracers in large quantities. The
approach opted for utilizes double exposed images of laser
induced fluorescent particles which are analyzed
automatically by digital image processing. This method
enables one to resolve the entire velocity field
simultaneously, thus allowing the identification of large
scale flow structures as well as provide quantitative details
about the velocity, circulation, etc. This technique is
specifically suited for the study of large scale complex
turbulent flows. Similar to the techniques of Gharib &
Willert [5] and Khalighi [6], it consists of recording two
exposures on a single frame. This approach minimizes the
difficulty of identifying particle traces on successive frames,
and still allow for short sampling intervals. Unlike the
others, we have opted to identify the direction of flow by
keeping one of the exposures longer than the other. The
following paper focuses on this method.
An important assumption in all particle tracking techniques
is that the seed particles follow the flow without slipping
and do not alter the flow dynamics. This requirement
prescribes the size, concentration and specific gravity of
particles that can be used as tracers. Most of the past studies
(Adamcyzk and Rimai [7], Landeth, Adrian and Yao [8],
Khalighi [61), rely on the light reflected from the particles.
This limitation has prohibited the use of very small
particles, due to the low intensity of the reflected light. By
utilizing tracers containing fluorescing material (Gharib and
Willert [5]), the intensity of the emitted light is increased by
several orders of magnitude, so that even microscopic
particles can be used. The generation of particles with
imbedded fluorescing material has been one of the critical
problems, particularly when they are needed in large
quantities.
Experimental Procedures
During the past two years we have constructed a large scale
flow visualization facility with multiple light sheets in the
140 ft. towing basin at David Taylor Research Center (sketch
is provided in Figure 1). While developing the flow
visualization system in the towing tank, we have used water
jets as the flow field in order to generate data for the imaging
system. As a result the images provided in this paper have
been recorded in a steady flow water jet facility located in a
transparent test section. The jet was 1 inch in diameter. Thin
sections (approximately 1 mm) of the flow field were
493
OCR for page 494
illuminated by a pulsed 300 mW argon ion laser. The water
was seeded with microscopic (5-10 microns in diameter),
neutrally buoyant particles, containing imbedded fluorescing
dyes. These particles were invisible in most of the flow field,
but responded with intense spontaneous fluorescence within
the illuminated section. The production of these particles
will be discussed later. The temporal light modulation
followed a specific illumination code. Figure 2 shows the
pattern used in the present work. The signal consisted of a
long exposure (streak) followed by a shorter pulse (dot). The
magnitude of the velocity was determined from the distance
between the two traces of the same particle, while the flow
direction could be determined by comparing the lengths of the
two particles. As will be discussed later, the automatic
image analysis algorithm could match between the traces as
well as identify and remove streak patterns which did not
resemble the illumination code.
The images were recorded on 35 mm film. We have opted for
film since its resolution is much higher than that of video.
As an example, Figure 3 contains two digitized images, both
originated from the same negative. However, in the first one
the entire negative was translated to a single video frame
and then a portion of the frame was magnified. The second,
on the other hand, was magnified prior to translation to
video and as a result is clearer and sharper. A video frame
has a resolution in the order of 500 x 500 lines, which is less
than the resolution of 1 mm2 Of emulsion. Thus the
translation to video should be performed carefully to avoid
loss of details. Storing the original image in the form of a
film negative allows variation in magnification while
digitizing the image, and as a result enables us to control the
resolution. While analyzing the image, one can focus a video
camera on the negative and select the magnification
depending on the desired detail. This is especially
important when examining turbulent flows where wide ranges
of velocities are present. For example, Figure 4 shows a
typical image of the flow near the exit of a 1 inch diameter
nozzle. Portions of the flow are unresolvable at this
magnification. By focusing on a smaller portion of the
negative (Figure 5), the magnification and hence the
resolution are increased allowing for analysis of the
originally unresolvable traces.
Image Analysis and Processing
As noted before, the image analysis and processing
algorithms were developed with the specific objective of
examining large scale real turbulent flows. As a result, the
selected technique should be able to analyze images with a
large numbers of particles at a wide range of velocities. Thus,
it was necessary that the image resolution be such that both
very small (low velocity) traces as well as long (high
velocity) streaks could be clearly identified and handled
efficiently. It was also necessary for the algorithms to be as
simple as possible to maximize the speed of the analysis.
The images, recorded on film, were digitized by
illuminating the negative from behind and by focusing a RCA
video camera with a microscope objective zoom lens on a
section of the film. The camera was linked to a high
resolution video recorder as well as to an Imaging
Technologies Inc. Series 100 Image processor and frame
grabber which were installed in a Sun 4/260 workstation.
Each video frame was digitized to a 512x512 pixel, 8 bit
array. Each pixel was assigned an intensity value, ranging
from O to 255, corresponding to its relative brightness. The
digitized image was then enhanced by color filtering, a
smoothing convolution and contrast enhancement to reduce the
noise. The use of fluorescing particles has the added benefit
that the emitted light is of a wavelength which is higher
(in the yellow range) than the green light reflected from
bubbles and contaminants. This feature allows significant
enhancement of the input images by removing much of the
reflected laser light through color filtering.
The filtered image was then sharpened by convolving with
the following kernel:
-1
-1
-1
-1
12
-1
-1
-1
-1
Namely, each pixel value was multiplied by 12 and its eight
nearest neighboring pixels were multiplied by -1. Then, the
sum of these values was added to the original pixel intensity.
Performing this process on the entire image effectively
sharpens the edges of the traces (Figure 6). The image was
then equalized, namely the intensity values of the entire
image were normalized to range from 0 to 255, to improve
contrast.
The next step was to "threshold" the image. Pixel values
above a selected intensity level were set to 255 and values
below it were set to 0. The threshold level was determined
from an intensity histogram of the entire image. In an
optimal situation the intensity histogram would be bimodal,
with well separated peaks. That is, the particle traces
would be easily distinguishable from the background and
their edges would be distinct and clear. In practice this was
not usually the case. In fact, it was not uncommon that the
brightest background pixel would be brighter than the
faintest pixel of a particle. This phenomenon occurred when
the background illumination was not uniform. Therefore
construction of an accurate binary image using threshold
analysis required the use of local threshold intensity levels.
Additional techniques could be utilized, if needed, to further
aid in distinguishing particles from the background. These
techniques include the use of gradient and Laplacian
operators, to provide edge enhancement (Rosenfeld and Kak
[9]) and examination of the slope of the thresholded average
intensity vs. threshold curve (Prasad & Sreenivasan [10]).
The thresholded image was then reduced to a binary bit
map of "0's" and "l's". The "0's" represented the background
(pixels of value = 0) and the "l's" (pixels of value = 255) were
parts of particle traces (Figure 7). This step reduced the
needed computer memory space and processing time in that
the image was now represented by an array of 1-bit integers.
The bit map was then searched pixel by pixel, row by row
until a pixel representing part of a particle trace was found
(a "1"). Then the total size of the trace, as well as its length
and orientation, were determined by examining connected
pixels. The centroid of the trace was also found at this time.
The program assumed that the trace found was the longer
streak. The length and orientation of this trace, coupled
with the illumination code, was then used to determine the
probable location of the matching trace. In the photographs
presented in Figures 3 6 the illumination code was such that
the exposure time for the streak and the delay between
exposures were 4 and 5 times the exposure time of the dot,
respectively (Figure 2). Once the search distance and
orientation of the streak were calculated, the matching dot
was searched for only in a limited space. It was necessary
however, to examine the image on both sides of the slot, since
the flow direction was not known a priori. If a second trace
was identified within the prescribed area, the ratio of the
slot length to separation distance was then compared to the
respective time ratios in order to insure that they were traces
of the same particle. The last position of the center of the
particle within the streak was then estimated by
determining the width of the trace at three positions along
the trace's major axis (Figure 8). The variation in their value
494
OCR for page 495
had to remain within a specified range for them to be
accepted as the actual width of the trace. The position of the
center of mass at the end of the streak was then determined to
be at a half width from the edge of the trace and centered
along its major axis (Figure 8). The same process was done for
the shorter trace. The velocity was calculated from the
estimated separation distance, center of mass to center of mass
(Figure 8). This sequence was repeated until the entire image
had been analyzed and a map of the entire velocity field
(Figure 9) was produced.
Since both sides of the traces were searched there was a
small possibility that dots which fulfill the above
mentioned criteria would be found on both sides of the streak.
If this situation occurred the velocities would be compared to
neighboring values to determine the correct direction. If the
correct direction and magnitude could not be inferred, this
data point was not used for the final velocity map. The
computer program also contains additional procedures to
handle unmatched traces, variations in slopes, flow near the
core of a vortex, etc.
The digitization, enhancement and processing phases were
done interactively which allowed for operator control of the
thresholding and scale of digitization. The analysis of the
bit maps was done automatically with its output being the
velocity vector for each particle trace found. A 512x512 pixel,
8 bit image was usually completely analyzed in
approximately 2 minutes of CPU time. The more "trouble
spots" (unmatched traces, multiple dots for one slot, noise,
etc.) there were, the longer it took for the computer to
complete the analysis. The entire process consisting of:
selection of an image, digitization, enhancement,
thresholding, and image analysis took approximately 10
minutes per image. More sophisticated automatic edge
detection techniques as well as automatic thresholding are
being implemented at the present time.
Particle Production
For the technique to be practical, particularly for large
scale towing tank flows, an efficient method of manufacturing
microscopic fluorescent particles was needed. A substantial
effort has been invested in developing a reliable and
controllable manufacturing process. The particles were
composed of a specific mixture of acrylics and several
fluorescing dyes. The mixture was adjusted to produce a
neutrally buoyant substance. They were manufactured by
dissolving the acrylics and then mixing the solution with the
dyes. The mixture was atomized and the resulting "dust" (5-
10 microns in diameter) was then collected and used as
velocity tracers.
Error Estimation
An important aspect of the analysis is to estimate the error
of the measurement. Geometric distortion due to the lenses
can be corrected for by using the techniques described by Green
[11]. Other errors are predominantly due to the digitization
which sets the accuracy of each measurement to + 0.5 pixels.
The velocity is equal to the separation distance divided by
the time lag between exposures. The separation distance is
found by calculating the edge to edge distance between traces
plus half the thickness of each trace. The error in edge to
edge distance is 1 pixel and the error in estimating the center
of the particle is 0.5 pixels, so the total error is 1.5 pixels.
This is a rough analysis since filtering and enhancement may
also introduce an error. As a result, future calibrations will be
utilized for determining the error more accurately. In the
future, the images will be digitized to provide generally a
separation distance of 15 pixels, resulting in an error of
495
separation distance of 15 pixels, resulting in an error of
approximately 10%. The error can be further reduced by
reducing the digitization scale. For example, if the image is
digitized such that the separation distance is increased to 60
pixels the error decreases to 2.5 %. However, the processing
time increases accordingly. Thus, a judgement of what is the
best digitization scale must be made. By recording the
original image onto film, different digitization scales can be
used for different portions of the image. This method
provides the capability to optimize between the processing
speed and the error. This feature is especially important if
gradients of the velocity are desired, i.e. during vorticity
analysis. Hesselink [12] estimated the maximum acceptable
relative error to be less than 0.5% to insure an accurate
vorticity determination. This error level is quite impractical
for the a technique presented in this paper. However, errors
on the order of 1-10% can be achieved, depending on the
length of the analysis. It should be noted here that to
achieve an error of loo the traces of a single particle should
occupy the entire 512x512 frame.
Summary and Future Work
A particle displacement velocimetry technique utilizing
digital image processing has been developed for examining
large scale complex turbulent flows. The technique consists of
illuminating a section of the flow field with a sheet of Argon
ion laser while seeding the water with microscopic
fluorescing neutrally buoyant particles. These tracers are
invisible in most of the flow field, but respond with intense
fluorescence within the illuminated section. By pulsing the
laser twice while recording a single frame each particle
leaves two traces on the same frame. The velocity is
determined from the distance between the traces, and the
direction of the flow is identified by keeping one of the pulses
longer than the other. Each recorded film negative is
analyzed by digitizing a portion of the frame, to a 512x512
pixel array which is then enhance, thresholded and then
translated to a bit map. The analysis of the bit map consists
of searching the array for traces. Once a trace is found, its
position, orientation, width and length are determined. From
this information, as well as the illumination code a search
distance and direction of the matching second trace are
calculated. If the second trace is found within the
predetermined space the velocity is then determined from
the separation distance. This procedure is repeated until the
entire image is analyzed.
At present, the system is installed in a 140 foot towing tank
at the David Taylor Research Center and is being utilized in
the study of the three dimensional separated flows. Further
refinement of the image processing and analysis procedures
are also currently underway. In the future the image
processing and analysis will be fully automated and expert
systems will be utilized to determine proper levels of
enhancement, thresholding and accuracy of analysis.
Acknowledgement
This work was sponsored in part by the Office of Naval
Research Applied Hydrodynamic Research Program and in
part by DARPA's Submarine Technology Program. Their
support is gratefully acknowledged.
References
1 Kobayashi, T. 1983 Proc. 3rd Int. Symp. Flow Visualization
Sept 6-9, Ann Arbor, Michigan, p. 261
2. Marko, K. A. and Rimai, L. 1985 Appl. Opt. 24, 3666 3672
3. Racca, R.G. and Dewey, J.M., 1988, Experiments in Fluids,
Vol. 6, pp. 25-32.
OCR for page 496
4. Adrian, R. J. 1984 Appl Opt. 23, pp. 169~1691
5. Gharib, M. and Willert, C. ,1988 AIAA paper 88-3776-CP,
pp. 193~1943
6 Khalighi, B. 1989 Experiments in Fluids, Vol. 7, pp. 142-144
7. Adamcyk, A. A. and Rimai, L. 1988 Experiments in Fluids,
Vol. 6, pp. 373-380
8. Landreth, C. C., Adrian, R. J. and Yao, C. S. 1988
Experiments in Fluids Vol. 6, pp. 119-128
CL — CYLINDRICAL LENS
F — FILTER
— MIRROR
L — LENS
iSTRUT SUPPORT
BACK ILLUMINATION PROBE
(NEW)
J ~ AL S/E R b R E E T
lo,
CL ~ c
LASER SHEET ~ /
L
/ ~=
/ / - HYDROFOIL
. ('
9. Rosenfeld, A. and Kak, A. C. 1982 Digital Image
Processing, 2nd edition New York: Academic Press.
10. Prasad, R R. and Sreenivasan, K.R., 1989 Experiments in
Fluids Vol. 7, pp. 259 -264.
11. Green, W. B., 1983 Digital Image Processing A Systems
Approach, New York: Van Nostrand Reinhold.
12. Hesselink, L., 1988 Ann. Rev. Fluid Mech. Vol. 20, pp.
421~85
ACOUSTO—OPTICAL SWITCH (NEW)
AR—ION LASER
i , V //
~ OPTICAL FIBER (NEW) a
'_ ~ , _ — BEAR SPLIT
_—TRAILING CAMERA / (NEW)
f 1 ~ ~ (E XISTING) /
TRAILING PERISCOPE /
~ (EXISTING) /
'`L /
, ~ ~~
^~ ~
_—SIOE CAMERA (NEW)
TRAILING PERISCOPE
(EXISTING) ~
SIDE ILLU~INAT10N PROBE
(EXISTING)
I SIOE PERISCOPE (NEW)
_ ~
Figure 1: Sketch of the large scale flow visualization facility in the
140 ft. towing tank at the David Taylor Research Center.
~ ~ ~ Laser beam modulation
,
,
,
496
Pa rti cl e t race Figure 2: Incident illumination
modulation code with a typical
pair of traces of the same particle.
OCR for page 497
Figure 3: (a) Image of particle traces enlarged after digitization.
(b) Same image as (a) but magnified prior to digitization.
Figure 4: Typical double exposed image of the flow
near a 1 inch diameter nozzle.
Figure 5: Enlarged section of Figure 4.
Figure 6: Figure 5 sharpened through convolution.
497
OCR for page 498
Figure 7: Binary bit map of Figure 6.
1) Measure trace length.
it- A
2) Determine trace width and center of particle.
~3
w/2 ~ ~ 1~ ~ ~ W/2
Wit W2 W3
3) Search for the partner trace.
4) Measure parti cl e di spl acement.
I..
o
Figure 8: General sequence of image analysis steps.
(1) Determine the length and orientation of the trace.
(2) Determine the width and the position of the
center of the particle at the end of the trace.
(3) Search a small area at the calculated search distance.
(4) Measure the particle displacement.
~~ ~~ ~ ~~ ~~ ~: at. ~~ .. ~~ ~~ ~ ~ ~ ~~ .~ ~~ ~.~.~.~.~. ~~ i: i- i: ~ ~.~ ~~: ~~ ~ .~ : ?~ ~ ~~ ~~.~ ~~ I.... ~~ ~
Figure 9: Map of the velocity field of the flow near
a 1 inch nozzle, determined through analysis
of Figure 7, by digital image processing.
498
Representative terms from entire chapter:
image processing
