1. Results and Discussion
Identifying metal surface defects requires technologies that can
detect surface anomalies while maintaining the parts intact. We
present a comparative analysis of three vision systems to predict
defects on the surfaces of aluminum castings: a hyperspectral
imaging system, thermal imager, and digital color camera have been
used to inspect the surfaces. Hyperspectral imaging provides both
spectral and spatial information. Each material produces specific
spectral signatures which are also affected by surface texture. The
thermal imager detects infrared radiation whereby hotspots can be
investigated to identify possible trapped inclusions close to the
surface, or other superficial defects. Finally, digital color images show
apparent surface defects that can also be viewed with the naked eye.
The surface defect locations are analyzed and predicted using the
three systems, and verified by tensile testing. The final goal of the
project is to determine the most effective vision technology to
nondestructively detect defects.
Abstract
Introduction
Comparing Three Vision Systems for Metal Surface Defect Detection
Shawn Robinson*, Ruby Mehrubeoglu**, Petru-Aurelian Simionescu**
Hyperspectral Optical Property Instrumentation (HOPI) Laboratory
School of Engineering and Computing Sciences, Texas A&M University-Corpus Christi
*presenter, **mentor
Experimental Methods
Data Analysis Methods
Summary and Conclusion
Acknowledgement
References
Results and Discussion
Defect detection technologies have a wide range of variances
and an even bigger range of applications. From X-rays to ultra sound,
the most important goal of the technologies is to find the defects fast,
accurately, and as inexpensive as possible. The biggest market for
defect detection lies in the investigation of pipelines, aircraft wings,
support beams, and roadways. In this experiment we will consider
and compare 3 different vision technologies; Thermal, Hyperspectral
and Digital Imagers. Thermal systems are most commonly used as
night vision, or to identify moisture spots in housing structures,
because of their acute ability to detect infrared wavelengths (heat),
invisible to the human eye. This allows for the easy detection of hot
spots, which are usually problem areas. Hyperspectral imaging
systems are most commonly used in geographical scanning
equipment. They can take images over a wide range of acute
wavelengths and compile them together. They system computes a
relative brightness for each wavelength for each pixel of data. Form
this data you can deduce information like changes in elevation,
change in texture, and changes in material. High definition digital
Images allow for the complete capture of everything visible by the
naked eye, with little to no distortion. This allows you to analyze color,
shape, size, texture, and other obvious features accurately and make
reasonable predictions about them.
Photo of bars A3 and A4 that have been through tensile testing and
broken
Figure 2. Ground Truth
Digital Images -Lay each sample horizontally on the
countertop.
-Turn off overhead light and flash to reduce
bright spots.
-Keep the lens of the camera about a foot
away from the sample for each
measurement.
-Take pictures of each side of each sample
Thermal Images Put samples in an oven and heat them up to a
rage of 200-215 degrees F
-After the samples reached their new
equilibrium, take them out two at a time, quickly,
so that excess heat doesn't escape the oven.
-Put the hot samples on a sheet of black foam
board the reduce noise cause by the hot
samples exposure to external environment,
causing background temperature change.
-Take measurements of each side of each
sample keeping the lens about 3 feet away from
the samples.
Hyperspectral
Images
-Lay one sample at a time on the stage and
adjust till focused. Average width of each
sample is about 13mm
-Set system to reflection mode and set to
scan in 330 micron increments.
-Set scan length to 22mm and the
integration time is 50ms. Adjust stage 0.5cm
up when scanning sides.
Using the digital images we were able to visually identify
the most significant defects in each sample. When looking for
defects visually, you look for things like shrinkage, inclusions,
impurities, crack chips and fissures. These types of defects act as
stress risers, which means when a load is applied, they will affect
that area most and cause breakage.
On the thermal images the temperature rage was adjusted
to a higher and more specific range to reduce noise. Aluminum
has a high rate of thermal conductivity, which means it should be
very easy to distribute heat throughout the sample. This makes
identifying hot spot easier because they would become more
apparent and obvious to it surrounding. Specific points were
analyzed and picked as the hottest and most substantial hot
spots on the samples.
Digital Images For sample A3, the prediction was dead on. The
nick acted as a stress rise in the bar, kind like
perforated tear on a bag of chip. However the
prediction for A4 was incorrect, which leads the
assumption that the defect was not visible or
sub-surface.
Thermal Images Both of the thermal images were extremely
accurate in its ability to predict the break point. The
thermal images succeeded where the digital
images had failed.
Hyperspectral
images
These images were also to predict the break point
accurately, although not as specific as the thermal
images. Figure 5. a.(aeft) Dark area
background sample holder
with Spectral profile of the
region
Figure 5. b.(left) Pure
metal region with minimal
defect with spectral profile
Figure 5. c.(left) Suspect
defect area with spectral
profile of that region
The wavelength 880.223nm was selected cause it had a better
average resolution. Then profiles were made in areas were the
metal quality was exceptional, the background region of the
sample holder, and a region where there were suspected defects.
The profiles show a relative brightness to respective wavelength.
Each of theses profiles were taken of an average of a 9x9 pixel
area, to smoothen out the curve and reduce error while remaining
specific the chosen area.
Digital images are a very good source for defect detection,
but not good for prediction as to where it is going to break. They only
accurate o the surface and leaves out internal data. These types of
images are not very constant because they are highly subject to
human error, due to factors like poor eyesight, misrepresentation,
blurred or distorted images, and personal experience identifying
defects.
Thermal Images are very accurate when it comes to identify
defects and probable break points. It has the ability to class levels of
risk based upon the temperature range. It requires little image
processing, and does not depend on skill level or experience. The
affected area is highlighted with a different color to precisely show
the exact shape size and location of the defect/ break point. It is vey
easy to ignore unwanted information and background noise.
Hyperspectral images have the highest accuracy when it
comes to defect detection. They are not only able to show areas of
defect, but can show areas of high quality as well. It is easy to ignore
noise by changing the wavelength that is viewed. These images also
have the ability to adjust the level of accuracy, depending a number
of data collection procedures, as well as image processing. Although
it is quite time consuming to collet data from hyperspectral system,
there is a plethora of information you can deduce form it once
acquired. Acquiring data this way is very sensitive and not flexible for
various environments.
All three of these technologies are suitable for detecting
defects. However this experiment suggest that thermal imaging is the
best candidate out of the three.
Digital imaging is very good place to start identifying probable
defect areas, but is subject to human error. Even though defects may
be found there is no way to know which defect is most likely to cause
failure other than assumptions and speculation.
Hyperspectral imaging is very costly and time consuming.
Although you get more information with the possibility of higher
accuracy, the equipment is too sensitive and not flexible enough to be
deployed in different environment quickly. Also, image processing is
time consuming and subject to human error.
Even though thermal imaging is more expensive than digital
imaging, it is not nearly as costly as hyperspectral imaging. It can be
easily deployed in various environments, with little data acquisition
time and little image processing. It is exceptional in its ability to
identify defects to an exact area with shape. Future work includes
different image processing techniques, which may lead to better
results from digital and hyperspectral imaging, including automated
and manual processing, as well as different chart and graphing
techniques.
This material is based upon work supported by the National Science
Foundation und Grant No. 0960000
Figure 3.(above) On A3(left) a nick, and on A4(right) hairline
fractures have been identified as stress risers, and possible break points
Results and Discussion
Figure 1. a. Digital camera, b. Flir Thermal Camera, c. Headwall
Hyperspectral camera
Figure 4. a.(top) Thermal scan of the samples, b.(bottom)
Temperature range adjusted to show hottest hot spots with point
selected
M. Sharifzadeh, S. Alirezaee, R. Amirfattahi, Detection of steel
defect using the image processing algorithms, Edition of book,
Isfahan, Iran: Proceedings ofthe 12 th IEEE International Multitopic
Conference, 2008, p. 125-127.