Cutset Sampling and Processing

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

PROJECT ABSTRACT Cutset Sampling and Processing D.L. Neuhoff (P.I.) This research project is focused on the development of efficient new methods for sampling and processing multidimensional data such as images, video, or spatially distributed sensor data, especially the first and last. Traditional methods of digital image processing rely on the taking and processing of regularly spaced samples (pixels), typically at the sites of a lattice, for example a square grid of points. In contrast, this project is investigating digital image processing based on samples taken in a cutset, the simplest example of which is a square grid of lines that spans the image. More generally a cutset is defined in terms of a set of neighborhood relations among a discrete set of potential sample sites. Cutset sampling is expected to be especially effective when the capturing of image edge information is important. The principle motivations for cutset sampling derive from the facts that two-dimensional data can often be well modeled by Markov random fields and that unsampled pixels of Markov random fields can be efficiently estimated with belief propagation algorithms. This research project is investigating different forms of cutset sampling, and it is developing reconstruction algorithms tailored to cutset sampling, image compression methods based on cutsets, distributed sensor network algorithms for scenarios in which sensors are deployed on cutsets, and analysis techniques to determine the strengths and weakness of cutset sampling, and to enable comparisons with conventional lattice-based sampling.