[Insight-users] apparent subsampling artifacts in MultiResImageRegistration

[Severino Fernandez]

I tried to register two images (not from the medical domain, they are
satellite images) using the multiresolution approach. In fact I used
the MultiResImageRegistration1 example.
I used the Mattes MI because both images are of different sensors (a
radar and an optical one) and the 1+1 evolutionary optimizer because I
had already obtained acceptable results with this optimizer, that
seems to be robust to the "noisy" metric I obtain.
But as I was interesting in inspecting the metric surface, I generated
another example using the ExhaustiveOptimizer and producing images of
it.
My surprise was that the metric surface for the subsampled images
presented regularly distributed artifacts. These built a kind of grid
in the surface, much like the interpolation artifacs studied by Pluim.
In principle, it seems that the artifacts generate larger values (less
negative) for the Mattes metric. My problem in this case seems not to
be affected by these artifacts, as I search for a minimum (more
negative, higher absolute value) to find the optimmal displacement (I
just use the simplest transform, a global displacement).
My particular problem is that the minimum value of the Mattes metric
surface does not coincide with the real relative displacement between
both images.
Then I throught that I could have a problem due to different sizes of
the overlap between the images for each displacement for computing the
MI, so I decided to use as metric the NormalizedMutualInformation.
But in this case the situation was still worse, because this
implementation produces positive MI values and the artifacts
completely distort the results, because in this case the maximum of
the computed MI is always at an artifact.
After this long description my first question is if this behaviour is
normal, if anybody has noticed these artifacts in the subsampled
images. Or if I am overlooking anything.
My second question is related to the relevant problem that the optimal
of the surface does not correspond to the real relative displacement
between both images. May this be due to the different overlap size for
each displacemente? If this were so, which would be a way to proceed?
Using the normalized MI has the artifact problems I mentioned.
Regards