[Insight-users] Problems with bilateral filtering

[Stuart Golodetz]

Stuart Golodetz wrote:
> Hi,
>
> I think I may be having some trouble using BilateralImageFilter 
> appropriately and just thought I should check. Essentially, I'm trying 
> it out as a replacement to multiple iterations of anisotropic 
> diffusion filtering for edge-preserving smoothing of my input image, 
> using the following code:
>
> ###
> ...
> typedef itk::BilateralImageFilter<RealImage,RealImage> BilateralFilter;
> BilateralFilter::Pointer bilateralFilter = BilateralFilter::New();
> bilateralFilter->SetInput(realImage);
> double domainSigmas[] = {6.0,6.0,6.0};   // test values (not connected 
> to the GUI yet)
> double rangeSigma = 5.0;   // ditto
> bilateralFilter->SetDomainSigma(domainSigmas);
> bilateralFilter->SetRangeSigma(rangeSigma);
> bilateralFilter->Update();
> realImage = bilateralFilter->GetOutput();
>
> if(is_aborted()) return;
> increment_progress();
> ...
> ###
>
> I was hoping it was going to be faster than the anisotropic filtering 
> I'm currently doing, as that takes about a second per iteration per 
> 512x512 slice (i.e. 30 seconds per slice for 30 iterations, which is 
> not ideal):
>
> ###
> ...
> // Smooth this real image using anisotropic diffusion filtering.
> typedef 
> itk::GradientAnisotropicDiffusionImageFilter<RealImage,RealImage> 
> ADFilter;
> for(int i=0; i<m_segmentationOptions.adfIterations; ++i)
> {
>    ADFilter::Pointer adFilter = ADFilter::New();
>    adFilter->SetInput(realImage);
>    adFilter->SetConductanceParameter(1.0);   // test value (not 
> connected to the GUI yet)
>    adFilter->SetNumberOfIterations(1);   // done this way so that I 
> can update the progress after each iteration
>    adFilter->SetTimeStep(0.0625);
>    adFilter->Update();
>    realImage = adFilter->GetOutput();
>
>    if(is_aborted()) return;
>    increment_progress();
> }
> ...
> ###
>
> Unfortunately, it takes 3 minutes for a 512x512 slice, whilst 
> producing "worse" results (in my context) than the anisotropic 
> diffusion filtering I was doing before. So I have three questions:
>
> 1) Would you expect bilateral image filtering to take this long? If 
> not, do you have any idea what I might be doing wrong please? (If not, 
> is there anything I can investigate to try and track down the problem 
> myself?)
> 2) Are the domain and range parameters having any effect on this, and 
> how should I go about setting them appropriately please?
> 3) Am I barking up the wrong tree by trying to use bilateral image 
> filtering as a drop-in alternative to anisotropic diffusion filtering 
> like this?
>
> Cheers,
> Stu
A separate (but related) question I have is: why would 3D anisotropic 
diffusion filtering on a 512x512x1 volume be substantially slower than 
2D anisotropic diffusion filtering on a 512x512 slice? I know the time 
step has to be smaller for the 3D version, but I recall it still being 
substantially slower than before. However, 3D anisotropic diffusion 
filtering on a larger volume e.g. 512x512x20 wasn't all that slow 
considering what it was doing. I was wondering whether boundary checking 
might be slowing it down? Or is there actually some other reason I 
haven't thought of?

Cheers,
Stu