[Insight-users] Re: Bug in bilateral filtering/ questions

[Luis Ibanez]

Hi Lisa,

Thanks for reporting this problem.

However,
That line was fixed on July 24 2006:
http://www.itk.org/cgi-bin/viewcvs.cgi/Code/BasicFilters/itkBilateralImageFilter.txx?root=Insight&r1=1.27&r2=1.28

Please update your version of ITK.


----

About your questions,


1) One potential mechanism could be to use a Mahalanobis distance that
    takes into account the Gaussian distributions of all components.
    You may want to look at:

         Insight/Code/BasicFilters/
           itkVectorConfidenceConnectedImageFilter.h

    and the
         Insight/Code/Common/
            itkMahalanobisDistanceThresholdImageFunction.h


2) The paper that you refer

    http://ieeexplore.ieee.org/iel5/10755/33900/01617216.pdf

    "Color Image Diffusion Using Adaptive Bilateral Filter"
     Jun Xie; P.A. Heng
     27th Annual International Conference of the
     Engineering in Medicine and Biology Society,
     IEEE-EMBS. 2005 Page(s):3433 - 3436


    is *not* published in an Open Access journal,
    therefore we have trouble getting access to it.

    Is there an equivalent publication in an Open Access venue ?

    For example, IEEE allows authors to post copies of the
    papers in their institutional web sites, as long as they
    made clear that IEEE owns the copyright of the paper.

    Curiously,
    if we go to that web site and download the paper
    we will be committing copyright infringement,
    http://www.copyright.gov/help/faq/faq-digital.html#p2p
    so even that is rather an entrapment option.  :-/


    "Anyone found to have infringed a copyrighted work
     may be liable for statutory damages up to $30,000
     for each work infringed and, if willful infringement
     is proven by the copyright owner, that amount may be
     increased up to $150,000 for each work infringed.
     In addition, an infringer of a work may also be liable
     for the attorney's fees incurred by the copyright owner
     to enforce his or her rights."



    Please let us know if you know of another publication
    E.g. Journals using the Creative Commons licenses are
    certainly better options.



    Regards,


       Luis


-----------------
Lisa Tang wrote:
> Hi all,
> 
> I noted that in itkBilateralImageFilter.txx, the following should be added
> after the reserve statement (line 246; in revision1.25) or a "subscript out
> of range" error would occur:
>   m_RangeGaussianTable.resize( m_NumberOfRangeGaussianSamples );
> 
> I also have some questions... 
> 
> 1) How might one extend the bilateral filter for nD/ vector-valued images?
> In the current implementation (for scalar), a table of Gaussians of range
> values are constructed; should such look-up table be generated for EACH
> component in the vector case or should there be one LUT? The major
> difference is the computation of the Gaussian product where the range
> difference is the component-wise vector difference so one could simply use
> the vectorneighborhood iterator to calculate the difference between 2
> neighboring vectors, correct?
> 
> 
> 
> 2) How might one implement the adaptive version of this filter, using the
> method proposed in the following paper?
> 
> http://ieeexplore.ieee.org/iel5/10755/33900/01617216.pdf
> 
> (The value of domainSigma changes according to eqn 9). 
> 
> Is it sufficient to recompute the Gaussian Image in each iterator's visit?
> i.e. 
>     b_iter = NeighborhoodIteratorType(m_GaussianKernel.GetRadius(),
> 				      this->GetInput(), *fit);
>  
> 2b) Could you also suggest a simpler formulation for the calculation of
> domainSigma that adapts to local image information?
> 
> 
> 
> 
> 3) The bilateral filter theoretically becomes a Gaussian smoothing filter if
> one were to replace...
> (a) gaussianProduct = (*k_it) * rangeGaussian;
> 
> with this...
> (b) gaussianProduct = (*k_it);
> 
> However, the images generated using (a) and (b) are quite different; could
> you suggest why that is so? 
> 
> 
> Many thanks in advance for answering these questions,
> Lisa
>