[Insight-users] Mutual Information

[Sarah Geneser]

I am working on an implimentation of Viola and Wells version of
registration by mutual information, and am having some difficulties.

I currently have an algorithm that follows the mathematics presented in
Viola and Wells 1995 paper _Alignment by Mazimization of Mutual
Information_.  However, when I supply my implimentation with the same test
and reference image and an initial transform that translates in the x and
y direction, I would expect the method to converge to the identity
transform, but this is not the case.

I imagine that this is a result of not knowing the correct parameters to
set for the method.  I currently have lambda set at 1.0, the 1D covariance
is 65536 and the 2D covariance matrix is diag(65536,65536).  My
understanding is that the covariance parameters must be large enough to
sufficiently blur the generated 1 and 2D histograms.  And lambda should be
small enough that it does not overstep maxima, but not so small that no
progress is made.  I feel that my current parameters are reasonable, but
they are not resulting in movement toward the known maxima.

Is there anything I am overlooking?

Any suggestions would be greatly
appreciated.

Thanks,

-Sarah Geneser