[Insight-users] User inputs Mean and Variance for K-means
Karthik Krishnan
Karthik.Krishnan at kitware.com
Tue Jun 13 11:34:07 EDT 2006
Hi Eric,
A variant of K-means that allows you to specify a mean and variance
could be accomplished by fitting a Gaussian Mixture Model to your data.
GMM's are a soft / probablistic alternative to k-means. As the name
implies it fits a set of gaussians to your data. As with k-means, you
would start with k initial means for the gaussians and k initial
variances (actually covariances for ND data, to be general). So you can
see, it lets you plug a known mean and variance. Unlike k-means clusters
can overlap (you can discretize them later by choosing a threshold).
Also, unlike k-means its less affected by noise or outliers in your
data, since samples far away from the mean are weighted by the gaussian...
In ITK, there are a couple of ways you can go about fitting a GMM to
your data :
1. See Examples/Statistics/ExpectationMaximizationMixtureModelEstimator.cxx
This uses the EM method to estimate GMM's that fit your data. You
can specify an initial set of gaussian distributions. Its an iterative
process. In the E-step you would find the weights from each gaussian for
a MeasurementVector. And in the M-step you would use those weights to
update the gaussian. So the gaussian would shrink/grow/shift to fit your
data.
2. If you already have an initial classification, you could use the
classes Code/Algorithms/ImageKmeansModelEstimator or
Examples/Statistics/BayesianClassifer.cxx
HTH
-karthik
Eric John wrote:
> Hello all,
>
> Is anyone familiar with a Kmeans or similiar method for classifying in
> ITK that allows the user specify a known mean and variance for each
> label in an image? Or how to modify an Itk example to do so. Thanks!
>
> Eric
>
>------------------------------------------------------------------------
>
>_______________________________________________
>Insight-users mailing list
>Insight-users at itk.org
>http://www.itk.org/mailman/listinfo/insight-users
>
>
More information about the Insight-users
mailing list