Unsupervised segmentation and classification of multichannel data

Contact person: Emanuele Salerno (ISTI-CNR)
Synopsis:

• The aim of this e-team is to study the problem of identifying and extracting the different components in a multichannel image without any other knowledge than the image itself. This is equivalent to \"segmenting\" the image (under a generalized notion of the term \"segmentation\") with no prior assumption on the emission spectra of the individual components. Different, generic, assumptions are to be made to overcome the underdetermination of this problem. In our case, a linear data model and a knowledge of some statistical properties of the component signals are common assumptions. Even in this restricted framework, each particular application presents additional difficulties. For example, the data can be nonstationary or particularly noisy, the mixing process can be linear but not instantaneous. All the partner groups have a long experience on this kind of problems.

Keywords:

• astrophysical images, Bayesian statistics, blind source separation, document analysis, MCMC, multichannel images, segmentation
  

Additonal info: E-Team's webpage