How is EM used in the sense of data mining on images? I understand how EM is used in the sense of estimating the Gaussian model that underlies a set of data, but its unclear how this is applicable. 
I am trying to understand how EM might be used to perform data mining on any sort of computer vision/image processing task.  This is the domain I am most familiar with, so it is helpful when learning about new algorithms to frame them in a familiar framework.
So one again, how and where is expectation maximization used in data mining images?  Thank you
 A: EM is a very general method to estimate the parameters of probabilistic models with latent variables. In fact, it is so general that it can come along your computer vision way in so many different ways.
E.g., as Anony-Mousse said, EM is essentially what is behind K-Means which can be used to find good representations of patches of pictures.
Also, EM can be used to estimate the parameters of... 


*

*a Kalman Filter, which you might want to use for tracking of objects in a video,

*factor analysis or probabilistic PCA, which are ways of reducing the dimensionality of data (e.g. for face recognition),

*a student-T distribution, which is better at representing the statistics of natural images than other classic distributions.

A: Well, in order to mine images, you need to transform them somehow into a more manageable format. For example into a histogram or a bag of words representation. k-means is a popular choice for the latter. An this is where these classic algorithms come in, despite being based on assumptions that generally will not hold for image data: a preprocessing step. Pixel errors can somewhat be assumed to be Gaussian distributed, so can sift features. So, using EM is not that stupid to simplify features, in particular when it improves performance for the actual task.
