[Insight-users] RE: Multi-resolution + MI + Evolutionary Optimizer

[Erik Harg]

On Jul 13, 2005, at 4:04 PM, Li, George (NIH/NCI) wrote:

> Hi, itk users and developers:
>
> I have done some multi-resolution image registration using Mattes  
> matrix and
> 1+1Evolutionary optimizer. In the file That records all of the  
> intermediate
> metric values, it shows That the metric value consistently goes  
> down within
> each Pyramid level. However, it goes up between these levels,  
> Consistently.
> There seems something unexpected.
>
Hi George. I believe these are both valid and expected results. I was  
also
wondering about this behaviour in my own program a few months ago, but
after thinking it over, I found out that is actually quite logical,  
and due to the
nature of the Mutual Information-based metric.
That the Mutual Information value decreases suddenly when you go from  
one
pyramid level to the next is natural. The images contain more  
information
(have higher entropy) at the lower pyramid levels. This is likely to  
give a lower
MI value for the two images, as there are just more pixels, with more  
differences,
to be used in the calculation. This is, as far as I can see, just an  
indication that the
multi-resolution approach is working.

-- 
Erik Harg