A Multilevel-Multiresolution Method for Image Processing. A Bayesian Framework for Reconstructing and Representing Shapes

Abstract : During the period of the grant, 1/15/90 - 1/14/93, we have developed: (1) a coherent multiresolution framework for image analysis tasks, in particular, for estimating 3-D shapes from a single video or SAR image; the algorithm has been applied to constructing topographic maps of Venus' terrain, and to segmentation/classification of textures, (2) efficient procedures for estimating the parameters of Markov Random Fields (MRF's) from noisy and degraded data, (3) a fixed-length noiseless source coding for MRF's using large deviations, and (4) a multi-grid type algorithm for maximum-likelihood estimation in tomography. In addition, we have begun a new non-parametric approach to speech recognition.