[Insight-users] "Look-up-table" suggestion

[Martin Kavec]

Thanks for reply, Shlomo.

I have implemented and already working SUSAN filter, which is a nonlinear 
noise reduction filter:

http://www.fmrib.ox.ac.uk/analysis/research/susan/


SUSAN nonlinear noise reduction reduces noise, ideally without blurring the 
underlying image. It is necessary to choose the spatial scale of the 
smoothing (just as in linear Gaussian filtering), and also an intensity 
threshold. The threshold is not, however, the step function but an 
exponential function of intensity difference between the working and center 
pixel in neighborhood. In the original, C implementation of the filter, the 
author uses a LUT for exp function, which speeds up the execution 
significantly.

As I mentioned, my ITK implementation filter is ready to use, but the use of 
std::exp function results about 5 times longer execution penalty. In 
comparison to the original my implementation is n-dimensional, and it gives 
great results on a time series of 3D volumes (such as fMRI). The problem is, 
however, the execution time.

Thanks for interest and ideas.

Martin 


On Thursday 29 June 2006 16:08, Shlomo Kashani wrote:
> Hi Martin,
> What type of filter are you working on? Maybe there is an alternative as in
> the case of a Gaussian derivative were instead of differentiating an image
> you differentiate the Gaussian kernel and convolve with the image. I am nor
> sure if your case can benefit but it would be helpful if you post the
> filtering operation you are trying to perform.
>
> Shlomo.