IMPAC
A Package of Programs for Segmentating Digital Images
Date of this documentation: 21-Dec-2001

Prof. Stanley L. Sclove, Ph.D.
Information & Decision Sciences Dept.
University of Illinois at Chicago

Description of IMPAC

IMPAC is part of the ISOPAC program suite, which also contains CLUSPAC for cluster analysis and IMPAC for segmentation of numerical images.

The input to programs in IMPAC is a two-way array of numbers

{x(i,j), i = 1,2,...,I, j = 1,2,...,J}, where x(i,j) may be a vector.

The algorithm used in IMPAC goes as follows. (More formal presentations are found in Sclove's ARO Technical Reports.) The elements of the segmentation model are the class-conditional time series or distributions, with their parameters; the labels; and the transition probabilities between the labels. Correspondingly, the algorithm alternates between estimation of the distributional parameters, estimation of the labels, and estimation of the transition probabilities. That is, given a tentative labeling, one can obtain tentative estimates of the parameters of the class-conditional distributions and of the transition probabilities. One then relabels the observations, using these updated parameter estimates. The relabeling moves from upper left to lower right and uses a one-step ahead Bayesian classification. This is a greedy algorithm which is suboptimal. The programs will be updated to use the Viterbi algorithm.

References

Sclove, Stanley L. (1984). Technical Report for ARO Conference on Design of Experiments and Applied Statistics, Las Cruces, NM, October, 1985.

Sclove, Stanley L. (1985). Technical Report for ARO Conference on Design of Experiments and Applied Statistics, Madison, WI, October, 1984.

latest update: 21-Dec-2001