[Insight-users] (no subject)

[Sayan Pathak]

Hi Jim,
Since there are different flavors of MRF filters in the literature, at the
time of the implementation of this filter we tried make the interface as
extensible as possible. At that time we chose an implementation based on the
Besag et al paper to be a representative candidate.  According to that paper,
a classifer is used to calculate the probability of a given pixel to belong
to a particular class. The label image is an initialization to the MRF. The
label image is used to determine the neighborhood influence on a pixel
location.

If you are looking into some other implementations, you may consider deriving
your class from this filter and then override the following virtual
functions.

virtual void GenerateData();

virtual void ApplyMRFImageFilter(); 

virtual void MinimizeFunctional();

The MRF filter provides a basic interface to define the model you may want to
use. This is done via the virtual void SetMRFNeighborhoodWeight(
std::vector<double> BetaMatrix ); 

By modifying the weights you can change your model. 

-Sayan

 

 

Date: Thu, 02 Dec 2004 13:37:18 +0800
From: "Jimmy Wong" <good_piggy at msn.com>
Subject: [Insight-users] About the MRFimagefilter
To: insight-users at itk.org
Message-ID: <BAY5-F1474A8B5D56A4DEF6052DAF9B00 at phx.gbl>
Content-Type: text/plain; format=flowed

Dear All,

For this MRFimagefilter, what kind of Ising model (or other models) is it
using? Why we need a traning image/label image?

Is the traning image the initialization for the MRF filter? or do we have to
obtain it by clustering algorithm?

Thanks.

Zhimin