[Insight-users] Performance of PatchBasedDenoisingImageFilter

[Kristen Zygmunt]

Hi Ben,
   Did you try it first with the default settings?  What size is your image, how many components does it have and what are you setting patch radius to?  A patch radius of 3 or 4 might be a good place to start.  Also, I'd recommend starting with 1 iteration. Since kernel bandwidth estimation defaults to off,  explicitly turning it off won't make a difference.

The other main strategy is to play with the sampling strategy.  The default sampling choice is set to look at all patches within a radius of 25 from the current patch.  You can change this by getting the sampler from the PatchBasedDenoisingImageFilter, and then setting its radius to something smaller.  Depending on your image characteristics, as you lower the number of patches to sample, you may want to consider changing the sampling strategy.  Other sampling choices include sampling near the current patch using a gaussian distribution (it's configurable what "near the patch" means, or sampling across the entire image with a uniform distribution.  You can see the sub samplers defined in Modules/Numerics/Statistics for more details.

Hope this helps,
Kris


> 
> Hi,
> 
> I am trying to use the itk::PatchBasedDenoisingImageFilter and am experiencing poor performance. I am currently only setting SetPatchRadius() and SetNumberOfIterations(). I also tried setting KernelBandwidthEstimationOff() which did not appear to help. Are there any recommended parameter settings for improved performance of this filter?
> 
> Thanks,
> Ben Alpert
> 
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL: <http://www.itk.org/pipermail/insight-users/attachments/20131002/6a3e4211/attachment-0001.htm>
>