Newsgroups: sci.image.processing
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!europa.eng.gtefsd.com!howland.reston.ans.net!wupost!waikato!canterbury.ac.nz!otago.ac.nz!bifrost.otago.ac.nz!nmein
From: nmein@bifrost.otago.ac.nz (Nick Mein)
Subject: Re: A question on Region Merging
Message-ID: <CyKFvs.3xr@news.otago.ac.nz>
Sender: usenet@news.otago.ac.nz (News stuff)
Nntp-Posting-Host: bifrost.otago.ac.nz
Organization: University of Otago
X-Newsreader: TIN [version 1.1 PL8]
References: <3931ik$k4s@bingnet1.cc.binghamton.edu>
Date: Tue, 1 Nov 1994 01:52:39 GMT
Lines: 72

Te-Shen (Dickson) Liang (liang@mipl.ece.arizona.edu) wrote:

>       [...] I would like to determine if these two regions should be 
>       merged into a bigger region. [...]  

ali@bingsuns.cc.binghamton.edu wrote:

> First of all merging two regions based on having low contrast is not a 
> very good idea. What happens when a region and a background have low
> contrast?? You would merge them together which is not correct. If your
> application is in object recognition, you should start that process now.
> and merge regions with each other if the new region makes sence and 
> has a model in the data base. I beleive we should do as little image
> processing as possible on an image, since by that we lose more and
> more information at each stage, and go to the matching part and match
> regions or the combination of several of them and merge regions 
> according to the result of the match.
> ali.

I have a problem similar to Te-Shen. I am working on a project to
classify (human) red blood cells by surface shape, from scanning
electron microscope images. Shape classes are:

	- "Normal" biconcave disks.
	- "Flat" cells.
	- Cells with irregular margins.
	- Cells with surface changes such as peaks and ridges.
	- "Cup shaped" cells.

Using the watershed algorithm (Vincent & Soille, IEEE Trans PAMI 13(6)
583-598) gives a good initial segmentation, but like Te-Shen I must now
merge some regions to get complete cells. The types of regions that I
have are:

   ******         ******       ******            ***       *** **
  *      *       *   *  *     *      * **         * *     *   *  *
 *   A    *     * B * C  *   *   D    *  *         *F*   *   G  *
 *        *     *   *    *   *        *   *        * *   *       *
  *      *       *   *  *     *      * E  *       * *     *     *
   ******         ******       ******    *       ***       *****
                                  *     *
                                   *****

	A is a perfectly segmented cell.
	B & C should be merged (B is a ridge, for example).
	E is occluded by D, and should be discarded.
	F is only part of a cell, and should be discarded.
	G is a perfectly segmented cell with an irregular boundary.

As Ali points out, merging regions based on simple gray-level statistics
is not appropriate. My current thinking is that I should try to classify
the initial regions using measures such as boundary shape, concavity,
.. ? and then try to reconstruct & classify complete cells.

I would be very grateful for:

	- Comments / pointers to literature on "higher level" region
          merging.
        - Pointers to previous work on segmentation of images of (blood)
          cells.
	- General comments.

Cheers,

Nick Mein
Dept of Computer Science
University of Otago
Dunedin
New Zealand.



