Dismiss
InnovationQ will be updated on Sunday, Jan. 21, from 9am - 11am ET. You may experience brief service interruptions during that time.
Browse Prior Art Database

# Unsupervised Segmentation of Multiband Images Using Information Theoretic Criteria

IP.com Disclosure Number: IPCOM000122628D
Original Publication Date: 1991-Dec-01
Included in the Prior Art Database: 2005-Apr-04
Document File: 2 page(s) / 59K

IBM

## Related People

Dom, BE: AUTHOR [+3]

## Abstract

Described is a method for image segmentation. The method is unsupervised (i.e., it does not require "training data"), and it works on multiband images. It is based on the minimum description length (MDL) principle [1] from information and coding theory and attempts to find the segmentation that yields, using a particular encoding scheme, the shortest encoding of the image data.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 61% of the total text.

Unsupervised Segmentation of Multiband Images Using Information Theoretic
Criteria

Described is a method for image segmentation.  The method
is unsupervised (i.e., it does not require "training data"), and it
works on multiband images.  It is based on the minimum description
length (MDL) principle [1] from information and coding theory and
attempts to find the segmentation that yields, using a particular
encoding scheme, the shortest encoding of the image data.

The method consists of the following steps:
1.   Fix an initial large set of small regions completely
covering the image.  For each region, compute (or assume) a
covariance matrix.  For example, the initial regions may be the
individual pixels, and the initial covariance matrices may be a
common diagonal matrix with diagonal elements equal to the global
variance in each band.
2.   Begin merging regions.  In each step, the two regions are
merged that give the greatest codelength decrease.  The codelength
decrease for any two regions t and v is computed as:
where Ri is the sample covariance matrix of region i, Rtv is the
sample covariance matrix of the combined region, i is the boundary
length between the two regions, and ni is the number of pixels in
region i.  K is equal to 1/2b(b+3), where b is the number of bands in
the image and is derived using Rissanen's "universal prior" [2].
(This codelength decrease term is based on assuming all regions have
a multivariate...