[Insight-developers] Classifier woes

[John Melonakos]

Jim,

I believe that the Bayesian Plug-In Classifier does what you are suggesting
here, basically using the MaxRatioDecisionRule to apply Bayes' Rule.

This will not work for me in my application.  If I recall correctly, the
Bayesian Plug-In Classifier jumps from the prior/membership probabilities to
the resulting labelmap without explicitly producing a vector image of the
posterior probabilities.  In my application, I desire to apply a smoothing
filter to those posterior probabilities prior to applying the decision rule.

Therefore, we are in the process of creating the
itkBayesianClassifierImageFilter which will provide the user with the option
to smooth the posteriors prior to executing the decision rule.  The general
flowchart that we (Luis, Karthik, and myself) have designed for this project
can be found in the SandBox as flowchart.png.  Note that when the filter is
complete, the user will not be constrained to using uniform priors or
Gaussian distributions.  Any arbitrary vector image of prior probabilities
and vector image of membership probabilities will work as inputs to the
itkBayesianClassifierImageFilter.  The code in the SandBox is about halfway
complete, but we are hoping to have the full flowchart implemented in the
next few months.

Hope this helps.  Let me know if you have any suggestions.

Thanks,

John

----
John Melonakos, ECE PhD Student
Laboratory for Computational Computer Vision
Advisor Dr. Allen Tannenbaum
Georgia Institute of Technology, Atlanta, GA
<lab> 404.385.5062
<cell> 801.885.3024
<email> jmelonak at ece.gatech.edu





-----Original Message-----
From: Miller, James V (Research) [mailto:millerjv at crd.ge.com] 
Sent: Friday, October 21, 2005 1:03 PM
To: Karthik Krishnan
Cc: Insight-developers (E-mail); John Melonakos
Subject: RE: [Insight-developers] Classifier woes

It seems like ImageClassifierBase with a MaxRatioDecisionRule
would be more efficient than the Bayesian filter in the sandbox.

If I understand the code in the sandbox (I just looked briefly at it),
it generates a posteriorPixel image which holds for posterior probabilities
of each pixel for every class.  Then the decision rule is applied 
to this vector image.

ImageClassifierBase runs through each pixel, calculates a vector of 
probabilities for the pixel, and applies the decision rule, then moves
onto the next pixel.  So no vector image of posterior probabilities is 
needed.

The handling of the priors is also not quite what I want.  The
MaxRatioDecisionRule allows the for a single prior per class
to be specified.

Jim