[Insight-developers] Classifier woes

[Karthik Krishnan]

Miller, James V (Research) wrote:

> I was just experimenting with some pixel classifiers in ITK.
>
> I had a program that I had written a few years ago that built up a 
> decision rule and set it on an ImageClassifierBase and passed that 
> ImageClassifierBase to the MRF filter.  This works fine.
>
> Now I just wanted to run a pixel classifier without running the MRF 
> filter.  So I was looking for a filter that would just run through all
>
> the pixels in an image and apply a decision rule to each pixel.
>
> The first filter I found was BayesianClassifierImageFilter.  This 
> sounded perfect.  However, this filter assumes the membership 
> functions are Gaussian and has no means to set a prior.  The filter 
> takes membership functions as opposed to a decision rule.  Most 
> troubling, however, is that this filter doesn't actuall do anything.  
> The GenerateData() method simply iterates over the input pixels.  It 
> doesn't evaluate anything or set any output pixels.
>
The BayesianClassifierImageFilter was a premature addition to the 
toolkit... The actual class itself is in the NAMIC sandbox. and does 
pretty much what you want.

http://www.na-mic.org:8000/websvn/filedetails.php?repname=NAMICSandBox&path=%2FBayesianSegmentationModule%2FitkBayesianClassifierImageFilter.txx&rev=0&sc=0
http://www.na-mic.org:8000/websvn/filedetails.php?repname=NAMICSandBox&path=%2FBayesianSegmentationModule%2FitkBayesianClassifierImageFilter.h&rev=0&sc=0

It allows you to set your own membership functions (they are not 
restricted to be Gaussian).
The class assumes a prior image with a uniform distribution, but again 
this is not a requirement. I think John's intention is to allow you to 
specify a prior image, with the default being a uniform one.
The membership functions using information from the prior generate the 
posteriors.

The decision rule is hardcoded to be a maximum decision rule, but it is 
not a requirement (although it generally makes sense, given that you 
want a MAP decision rule).  It could probably be set by the user, 
defaulting to MAP.

John's put in a test for it and the filter seems to classify just fine.

The goal might be to have a very generic BayesianClassifierFilter and 
subclass it for specific needs... ??

Regards
Karthik

> So I looked at ImageClassifierBase.  I was already giving one of these 
> to the MRF filter.  ImageClassifierBase has a GenerateData() method.  
> ImageClassifierBase, however, does not have the standard 
> SetInput()/GetOutput() methods.  Instead, it uses 
> SetInputImage()/GetClassifiedImage().  So it unclear to the user that 
> they could use ImageClassifierBase as an ImageToImageFilter.  Along 
> those line, ImageClassifierBase is not named like an 
> ImageToImageFilter (in particular, it does not end in ImageFilter).
>
> Do we have any ImageToImageFilters that you can give a decision rule 
> to and it will classify each pixel?
>
> *Jim Miller*
> */_____________________________________/*
> /Visualization & Computer Vision
> GE Research
> Bldg. KW, Room C223
> 1 Research Circle, Schenectady NY 12309-1027
>
> _millerjv at research.ge.com <mailto:millerjv at research.ge.com>_/
> /(518) 387-4005, Dial Comm: 8*833-4005/
> /Cell: (518) 505-7065, Fax: (518) 387-6981/
>
>
>------------------------------------------------------------------------
>
>_______________________________________________
>Insight-developers mailing list
>Insight-developers at itk.org
>http://www.itk.org/mailman/listinfo/insight-developers
>  
>