[Insight-developers] Normalized Gradient Fields Image To Image Metric

[Gert Wollny]

Hello all, 

like I pointed out before, I'm currently implementing the Normalized
Gradient Fields Image To Image Metric in ITK [1]. Of course this will
become an IJ submission with the code and necessary tests included. 

I have now a working implementation, but before I want to submit it, I'd
like to clean up the code and make it run faster. 

Now I have two questions:

1. For a 3D data set the program to register rigidly is awfully slow
(compared to an house internal implementation that doesn't rely on ITK).
With valgrind I measured that it spends 28% of the time in
VectorNeighborhoodOperatorImageFilter (with the DerivativeOperator) and
17% in NeighborhoodOperatorImageFilter called from GradientImageFilter
during a simple rigid registration (measured with valgrind). So that
would be 55% only to evaluate gradients.
Is there a better way to evaluate the gradients of an image?

2. I need to evaluate the noise of an image and currently I'm using a
function that takes an image as input and returns a value (Based on
ImageNoiseFilter). Where do I put a function like this, in an extra
file?

Many thanks, 

Gert 


[1] E. Haber and J. Modersitzki, "Beyond Mutual Information: A Simple
and Robust  * Alternative", in Bildverarbeitung für die Medizin 2005,
eds. Meinzer, Handels, Horsch and Tolxdorff, 2005, pp. 350-354