Abstract:
|
Statistical inferences are basic to image understanding but they require probability models that capture real data and are tractable. A large number of models have been proposed over the years. There are low-level models that study the image pixels directly--for example, Markov random fields or the newly proposed Bessel K forms on the spectral components. And there are high-level models that capture object shapes--for example, Grenander's deformable template model. A general algorithm for image understanding should start at the low level, i.e. analyzing features of image pixels, and build towards higher level inferences. In this talk, we argue that a careful selection of scene representations and the corresponding probability models lend to such sequential inference procedures. That is, the problem of performing hypothesis selection at level n, conditioned on the given data and the hypothesis selected at level n-1, is well-posed for the proposed Bessel K forms. Some examples from infrared face recognition will be presented.
(This research is being done in collaboration with Prof. Ulf Grenander of Brown University and Prof. Xiuwen Liu of Florida State University.)
|