Various Minerals And Image
Artifacts
Testing the Imaging Of
Different Minerals And The Appearance Of Apparent Fossils
BACK
When images are made,
people often assume that the picture is completely accurate and that we can
implicitly trust what we see in the final image. This is incorrect.
Rarely does an image convey absolutely believable information. We also
impress a certain bias on what we see based on expectation and experience,
rather than an impartial judgment of the image contents.
This is a
simple investigation into the appearance of artifacts that might be mistaken for
traces of biological activity. By testing different minerals using the
recreation of the microscopic imager's field and resolution, we can ascertain
whether some materials might appear to harbor structures that, when digitally
imaged, appear to be fossils or traces of fossils.
We have
demonstrated that small features can be compromised by both camera resolution
and also by image compression methods (such as JPEG algorithms). Can this
effect influence larger structures as well? To answer that question, I
selected minerals of differing types. Some are igneous and have no trace
of fossils. Others are man-made (crushed concrete) and one is limestone
gravel, which is rich in small fossils and seashells.
| This is the raw
image (scaled at 30%) of a group of magnetite samples added to the
simulated Mars soil.
I mixed them, shook the
container to allow them to rise to the surface, and blew low velocity air
over the setting to distribute the sand.
This will be converted to
monochrome and then enhanced to check for artifacts. Any artifacts
will then be compared to the original samples in an effort to find a
correlation. |
|
 |
| This is a set of
crushed concrete made with small stone chips. Some are probably
marble or limestone. I also noticed that there are machining marks
on some of these rocks that were made by the crushing machinery.
This is originally from some
concrete structure that was demolished and then converted to fill gravel
for parking lots. |
|
 |
| After monochrome
conversion and sharpness enhancement, this is the resulting image of the
magnetite samples.
Looking at this cropped
sample image, we can see that some circular or ring features did indeed
emerge. Examination of the original sample showed that these were
fractures or surface geometry that coincided with reflective or rougher
surface areas to create the illusion of features that were not actually
present.
When a single-viewpoint,
monochrome image is observed, we cannot distinguish between areas that are
lighter or darker due to actual material properties, illumination, or
geometry.
In some samples, this can
easily create illusory features. |
|
 |
| Similar results
are achieved with the concrete and other mineral types. Here is a
close up of the concrete chunks and some apparent features.
One result of this test is
that the size of the false feature is directly related to two factors- the
area in pixels of the feature, and the probability that random values can
produce a coherent image.
The implication is that the
larger a feature is, the less likely that a specific image will appear
there. Furthermore, the chance of a duplication of that image is
less likely as the square of the number of such occurrences.
In other words, 2 are only a
fourth as likely as one, 3 are only 1/9 as likely, and so on. |
|
 |
In other words, the
likelihood of an artifact is controlled by only a few factors. First is
the size or area of the artifact. This is not just the features of the
artifact, but the background that it is recognized against as well. In
this area is a certain number of pixels, each of which might have any arbitrary
value. If we created an identical area and filled it with random pixels,
what would the chance emergence of the artifact be?
This is a tough
question that depends on the recognition ability of the human visual
system. But simple math shows that for a given area 100 x 100 pixels,
using a gray scale of 256 values, there are 256 ^ 10,000 possible images, and of
that huge number that as many as 5% might be seen as containing recognizable
features, such as lines, curves, polygons, or other complex shapes.
Our visual system is
optimized to see faces, circles, lines, and other basic forms that appear
commonly in our day-to-day lives. Because of this fact, we tend to
"put together" many such images where there are in fact none.
The simple act of looking at clouds is the best example- we see shapes and
objects that are truly not present, but that we are biased to recognize.
This effect greatly influences what we see in random images, as well as real
world ones.
However, as the
field size increases, the chance of a larger coherent image emerging actually decreases,
because it takes more "co-opted" pixels to create the image, and the
chances decrease that this will be so- it takes more random hits to make a
larger coherent image, but probability is against this.
The second factor is
duplication- a single, unique artifact is one story, but two, identical
artifacts are quite unlikely, unless they are small in total number of
pixels. Why? Because the smaller number of pixels constrains the
possible combinations that they can have, greatly reducing the number of random
tries needed to create a given artifact. So large, repeated features are
far more likely to be the genuine article, while small, repeated features are
the converse- less likely to be genuine. Add in the effects of image
compression algorithms, like JPEG, and it becomes very likely that some small
features will emerge that have no bearing on the actual image content.
As it turns out, a
complex, symmetrical artifact is very unlikely, because it requires duplication
of features internally. We can treat such an image in the following
manner- suppose that an image feature has 3 or 4 identical parts- perhaps they
are rotated, as in the case of a cross section of an orange. Each section
of that image is identical to the others, and while a small segment of the image
is not too unlikely, having 6 or 10 or 20 of them is geometrically more
unlikely. A 20 segment rotationally symmetrical image is at least
400 times less likely than a single segment one, and in most cases actually many
millions or billions of times less likely. If the chances of a small
feature appearing at random are, say, a million to one, and then 10 such items
appear in a small patch of the image, the odds explode to a million to the tenth
power (or a one followed by 60 zeroes) against such an occurrence at random.
So, the best defense
is to have many images, larger images, and many pixels. It also helps if
features are symmetrical because this squares the odds against random factors
creating such an image. These items help
to trim away any random or extraneous factors that might contribute to false
recognition of features.
References:
Image Processing and Pattern
Recognition in Soil Structure
Daisheng Luo
Department of Electronics and Electrical Engineering University of Glasgow
http://www.mech.gla.ac.uk/Research/Control/seminars/1mar95.html
Size and shape
recognition using measurement statistics and random 3D reference structures.
Arnab Sinha and David
J. Brady, Duke University
http://www.opticsexpress.org/abstract.cfm?URI=OPEX-11-20-2606
Image recognition in single-scale and multi-scale decoders
Dirck Schilling, Pamela Cosman, and Charles Berry
Department of Electrical and Computer Engineering, University
of California at San Diego
http://code.ucsd.edu/~pcosman/conf-paper-25.pdf
Image generation and shape recognition toolkit
Mike Schauf and Selim Aksoy
http://www.ee.washington.edu/research/isl/IAPR/ICPR00/shape_recognition/shape_recognition.html
END