Utilizing Image Processing Techniques to Determine Alignment for Wind-On and Wind-

Off Images in PSP Experiments

By Catherine McCarthy

Supervised by Mike Bragg, Brian Woodard, Jeff Diebold

Aerospace Engineering, University of Illinois at Urbana-Champaign

Introduction:

One way that the pressure on a wing can be measured is through pressure taps. These are small

holes that are on the wing and measure the static pressure on the wing surface. Usually this static

pressure is then referenced to the static pressure in the freestream so the pressure coefficient can

be calculated. Another, more advanced, method in which the pressure on the wing can be

measured is through the use of pressure sensitive paint (PSP). This paint emits light at different

intensities based on the local pressure, and a continuous pressure distribution is obtained by

imagining the paint via an excitation light. Pressure sensitive paint requires a wind-off and wind-

on picture. Pervious experiments on swept wings at UIUC found that there was a significant

amount of noise on the tip of the wing model due to model deflection caused by aerodynamic

loads, resulting in misalignment between the wind-on and wind-off images. A solution to this is

to use image registration, which identifies physical markers on the surface of the model and

aligns the two images based on those marker points using computer software. Pressure taps on

the wing are commonly used as markers. Therefore, these pressure taps’ use are two-fold. They

allow collection of static pressure measurements, and for the pressure sensitive paint to be

aligned.

Studies on the effects on marker size and shape for large scale model deformation have been

completed by Wroblewski4. In her experiments, Wroblewski4 built a PSP Alignment Test

Apparatus which statically deformed a plate in different ways in order to represent the

deformation of a wing. In this apparatus, she utilized a plate with computer-generated markers to

align the wind-off and wind-on images. A Matlab code developed at UIUC would then be

utilized to align these images. This would be accomplished by finding the center of the markers

via a centroid method. There was a need for both alignment markers as well as markers that

would be used to check the accuracy of the alignment. These markers would be differentiated by

manually selecting the respective markers on the aligned image. One of the key conclusions of

this research was that more markers should be used for alignment, but at a certain limit no more

markers are necessary in order to obtain similar results. She also claimed that smaller, circular

markers improve accuracy. Finally, it was concluded that markers should be placed near the

edges of the surface to increase accuracy. In this experiment, the only two types of marker

patterns that were measured were large, low density vs. small, dense markers. More experiments

are required in order to isolate individual variables. By doing this, conclusions can be reached on

which specific variables most effect the accuracy in the alignment.

Bell and McLachlan1 compared different types of image registration techniques and tested the

limits of image registration for model motion that is larger than the models utilized in this

experiment. One of the measurement techniques outlined in this paper is the Delaunay

triangulation. In this method, “each image is divided into triangles whose vertices are at features

which can be precisely located in both images”1. They discuss how this Delaunay triangulation is

a more sophisticated method of transformation. The code utilized in the present work is

primarily utilizing a 2D projective transformation to align the two images. Different image

alignment techniques need to be tested to determine which of these methods is the most accurate.

A good starting point in exploring these different methods would be to compare the projective

transformation method with this outlined Delaunay method.

Finally, in Mantick, Quick and Quest2’s experiments, they utilized a shaking plate mechanism to

measure distortions of the plate and how that affected the accuracy of the measurements taken.

The use of a shaking mechanism is interesting, but they did not speak as much about isolating

different types of movement to see where the deformation reaches a significant level to where

the alignment is no longer accurate. In this report, linear rigid motion was analyzed in order to

define where the deformation becomes so extreme, that conventional alignment techniques are

no longer applicable.

Experimental Methods:

The first step of this setup was to create four different types of marker

patterns. The four types of plates utilized where a large, high density

plate, a small, high density plate, a large, low density plate and a large,

low density plates Examples of each pattern are shown in Figure 1.

Computer Aided Design software was used to create the patterns and they

were printed on sticker paper. The plate was placed in an apparatus on

one side of a table (Figure 2), and a camera was on the other side (Figure

3). The camera utilized was a USB2.0 CCD Mightex camera. A F11

setting with a shutter speed of 150 ms was utilized, and raw images were

taken. The plate was illuminated by a lamp, and was able to deform

rigidly when attached at the bottom. In addition, screws could be put into

the plate in the top two corners in order to create a radius of curvature

with the plate (Figure 4).

Fig 1: A sample of the

different plates utilized

in the experiment.

From left to right:

large, high density,

small, high density,

large low density, small,

low density

Fig 3: The camera setup

Camera

Plate in

Apparatus

Lamp

Fig 2: The plate setup on the

opposite side of the table

The plate was rigidly deflected in increments of 3 degrees, ranging from 3 to 30 degrees, with no

deformation representing the wind-off image. A series of 5 pictures at each deformation was

taken, and were averaged using ImageJ software. A background picture, or a blank photo, was

also taken to utilize as a base for the image so that no external portions of the image that weren’t

the plate would be included in the data collection. The averaged .tiff image would then be

processed in the Matlab alignment program previously mentioned. In this program, the center of

the individual markers is located using a centroid method.

First, the background, wind-off, and wind-on photo are uploaded. The background image is

subtracted from the wind-on and wind-off image, in order to only have the plate remaining in the

image. The program then computes a global threshold level for the plate to differentiate between

the markers and the background on the plate. It does this by utilizing Otsu’s method3. After this

is accomplished, the wind-on and wind-off images are changed to binary based on whether the

intensity is above or below this global threshold. This differentiated between the background and

markers on the plate. The center of mass, or centroid, of each of these markers is then found. The

x and y components are stored as an array, and reordered in order to align each dot in the wind-

off image to its corresponding dot in the wind-on image. Any duplicate markers are removed,

and then the user is prompted to select the measurement points, followed by the alignment

points. This is done on a screen that has the aligned plates represented, with moveable markers

assigned to each dot and its counterpart. These two different types of points cannot be identical,

as that compromises the alignment image. The program then determine how many pixels the

center points are away from each other pre and post alignment, and returns the root mean square

error in both the x and y directions. In addition to this, an alignment image is rendered. This

RMS error was then transferred to a table in excel for further analyzing.

Fig. 4: The apparatus used with

a plate attached

Results:

Figure 5 shows a contour plot of the local alignment error across the plate. If an image had

darker blue or darker red in it, then it was less aligned. Figure 6 shows the RMS error versus the

degrees of deflection.

In the majority of the data, there are some noticeable rises and falls that do not follow what was

predicted to take place. As the rigid deflection increased, it was believed that the RMS error

Fig 5: The contour plot on the left represents a plate that has only 3 degree deflection,

while the one on the right illustrates 30 degree deflection. The 30 degree deflection is a

much darker blue than the 3 degree deflection, showing that it is less aligned.

Fig 6: Degrees of Deflection vs. RMS Error after Alignment

0

1

2

3

4

5

6

7

0 5 10 15 20 25 30 35

RM

S Er

ror

Aft

er A

lign

men

t

Degrees of Deflection

Degree of Deflection vs. RMS Error After Alignment

Small, High Density Markers

Small, Low Density Markers

Large, High Density Markers

Large, Low Density Markers

would rise. However, we can see that in all the data, apart from the small, close together, there

are rises and falls with certain degrees of deflection. Even within the rises and falls there is no

clear pattern amongst all the data on when these will occur. However, despite these problems, the

error is lower for the small markers rather than large markers. Furthermore, the small, close

together markers are the only type where there are not any seemingly random rises or falls in the

data.

There are a few reasons that this inconsistent data could have arisen. A problem that arose from

the code was that it occasionally had a challenging time determining where the ‘beginning’ of

the plate was. Oftentimes the bottom row of the wind-off image would be aligned with the third

or fourth row of the wind-on image. The corresponding markers would be manually moved to

make sure that the two images were aligned properly. However, a problem with this method is it

leads to the possibility of human error, as the alignment points may no longer be located at the

exact center of the dot. This happened most often with large degrees of deflection, and may

account for the spike specifically in the large, far apart data at the 24 degree mark.

Another problem that could lead to this datum is the method in which the code is finding the

center of the markers. The Matlab code is initially creating a gray threshold, then converting the

picture to binary in order to locate the center of mass for each dot. However, depending on what

that grayscale is, the program may not be able to have a perfect circle with the binary image,

leading to an incorrect center point. This would also lead to misalignment, and could account for

some of the spikes that are seen in the data.

Finally, another potential reason for the poor data could be due to the cropping technique utilized

with the close-together plates. The program ran too slow, and was causing problems to occur

with the image processing. To resolve this, the image was cropped down length-wise, creating a

long, skinny plate equivalent. However, while there were plenty of markers required in order to

get an accurate alignment4, the shape of the image could have led to some inaccurate data, since

the final plate that was run through the program was only 2 or 3 rows of markers wide.

Conclusions:

It is unlikely that we are able to draw any definite conclusions from the current data. There are

too many factors that might lead to inaccurate data that need to be addressed in further

experiments.

First, a new method of finding the center of the markers needs to be explored. The grey threshold

algorithm along with the binary images will be analyzed to determine if the circle is uniform

enough in order to create an accurate representation of the center point. One way in which this

can be accomplished is by creating an artificial plate, which would be simulated on a computer,

where the center point of the markers is known. The ‘plate’ could then be run through the

program, and the alignment could be compared to what is known. This way, the amount of

accuracy that is prevalent utilizing the centroid method could be measured.

Another problem with the method utilized is the fact that the current amount of manual input

increases the potential for inaccurate data. To resolve this, a pattern recognition method may

need to be input into the program. With this, the user could identify one point, such as the bottom

corner dot, and the computer will have these two aligned points and be able to align the

remaining points based on how far apart they are from each other. This could also help with the

amount of time that each trial consumes, leading to the ability to create and analyze more data to

come up with better conclusions.

If it is determined that the centroid method is an accurate way of finding the center of the points,

then a better method of cropping the photos needs to be explored. At this point, the close-

together plates were cropped so that they were only about 2 or 3 rows wide. This may not be

wide enough in order to get an accurate alignment. The different cropping methods will be

explored, and the point where the width is not a factor in the alignment method will be identified.

The biggest factor in that method is to be sure to include the top of the plate, where the deflection

is the most apparent.

It could also be determined if there is a large difference in the data when a different contrast

color is utilized. Wroblewski4 utilized black markers on white background. In PSP experiments,

the paint is not white. It is pink, and a black and white camera is used in experiments. Mantick,

Quix, and Quest2 state that there is a noticeable difference in accuracy if there is a high contrast

between the markers and the background. Utilizing different colors of markers, or having a

gradually changing difference between background and marker color could allow us to see what

the best experiment setup is to get the most accurate results.

Overall, there is still a lot of work to be done with this experiment. The top priority at this point

would be to improve the program in order to increase its accuracy without the need for human

input. This would hopefully eliminate some of the initial errors. Beyond that, the method of

which the pictures are run through the code may need to be reevaluated. This would be the best

way to begin the elimination of the current errors accrued by utilizing the Matlab code.

References [1] Bell, J. H., & McLachlan, B. G. (1996). Image registration for pressure-sensitive paint

applications. Experiments in Fluids, (22), 78-86. [2] Mantik, J., Quix, H., & Quest, J. (2013). Enhancement of the stereo pattern tracking

technique for model deformation assessment at ETW. European Transonic Windtunnel

GmbH. [3] Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE

Transactions on Systems, Man, and Cybernetics, 9(1), 62-66. [4] Wroblewski, G. (n.d.). Custom pressure sensitive paint image registration software testing

error.