Automated Initialization and Automated Design of Border Detection Criteria in Edge-Based Image Segmentation
[Karl Note: The concept of an MRI looking for the EDGE of something is becoming well recognized in the "science of MRI." When that "edge" is the wall of a cell, the MRI has a very long way to go -- to detect the "edge" which is 100 times smaller than a human hair, and too small to be seen in any microscope. the article on this page acknowledges the need, within MRI examination, for improvement in "edge detection."]
Involved researchers: Milan Sonka (my advisor), Marek Brejl (myself)
Project Description -
Introduction
Recent applications of image segmentation and
image understanding techniques require
increased robustness, better reliability and
high automation of t he algorithms. While
traditional methods of region based (region
growing, texture based segmentation) and edge
based (snakes, dynamic programming) ima ge
segmentation algorithms continue to be
explored with attempts to improve their
performance by adding expert knowledge to
their detection criteria, new model driven
techniques have been developed (active shape
models, active appearance models).
Performance of traditional approaches to
edge-based i mage segmentation is being
improved by utilizing advanced segmentation
criteria that reflect higher level of
knowledge about the segmented object. In most
applications, however, this expert knowledge
is not derived automatically, but introduced
by an observer and then translated to
mathematical form . This approach requires
expert knowledge of the given problem and is
usually time consuming when new application
have to be designed. Use of informat ion
about object shape is usually very limited or
is not considered at all. Active Contour and
Active Appearance Models are based on
statistical model of object shape and border
appearance derived from examples and this
information is used to improve image
segmentation. Unfortunately, both tradition
al and new approaches require manual
initialization - approximate location of the
object of interest - in order to succeed.
This work is attempting to solve both above-mentioned limitations of today's edge-based segmentation algorithms. Our approach is based on creating statistical model that contains information about object shape and its variance and information about border appearance. The model is constructed from training images with available segmentation. The information is then used to:
-
Find approximate position of the object of interest in previously unseen im ages, thus providing automatic initialization for any edge-based image segmentation method
-
Compute cost function in previously unseen images that utilizes information from the border appearance model (This cost function can be used in segmentation criterion of any edge-based segmentation method, substituting insufficient purely gradient information usually used in those criteria.)
Project Report (August 1999) [ pdf 300kB, 7 pages]
RESULTS - comprehensive verification of the algorithms
Feasibility study (1997)
We have performed a feasibility study to
prove that it is possible to learn
automatically image segmentation criteria
from segmentation examples.
We have developped a
radial-bases-neural-network based system for
automated design of a cost function used by
Dynamic Programming. The approach was based
on training a neural network to typical
border features (as for example gradient
profile of the border) and calculating the
cost function according to difference between
new and trained profiles.
The method was tested and proved to be
applicable. Border detection errors achieved
while using our system with automated design
of the cost function were comparable to
results achieved while using previously
developped systems with manualy designed cost
function.
Details of this feasibility study can be found in our paper.
PUBLICATIONS
Journal papers
-
M. Brejl, M. Sonka, "Medical Image Segmentation, Automated Design of Border Detection Criteria From Examples", special issue of Journal of Electronic Imaging, vol. 8(1), pp. 54-64, January 1999 [pdf 1MB]
Conference papers
-
M. Brejl, M. Sonka, "Edge-based Image Segmentation: Machine Learning from Examples ", in Proceedings - IEEE World Congress on Computational Intelligence, Anchorage, Alaska, pp. 814-819, May 1998 [pdf 270kB]
-
M. Brejl, M. Sonka, "Automated Design of Optimal Border Detection Criteria: Learning from Image Segmentation Examples", in Proceedings - 19th international conference - IEEE/EMBS, pp. 542-5, 1997 [pdf 260kB]
Poster presentations
-
M. Brejl, M. Sonka, "Automated Design of Border Detection Criteria from Example Contours", 1998 Symposium on Cardiovascular Imaging, Iowa City, Iowa, 1998
CEIG Talks
-
April 1999 - Presentation and software demonstration [PowerPoint presentation 800kB]
|
I promise to answer your message -- click here to send me a personal message
|
SUBSCRIBE: The Wednesday Letter is a free electronic monthly newsletter written and published by Karl Loren. You can view more than 50 back issues of this publication by clicking here. The Wednesday Letter subscription list is maintained on a secure server, no name is ever given or sold to anyone, and it is never used except for this Newsletter. It is automatically published on the Tuesday night just before the first Wednesday of every month. You can subscribe to this free monthly electronic letter by entering your eMail address and name below. You will then automatically receive a request for confirmation, sent to whatever address you have entered. If you do NOT receive this confirmation request, then you will not be subscribed. There may have been an error with your address and you should resubmit. The letter is never sent twice to the same address -- so you do not have to worry about a duplicate subscription. When you receive this confirmation request you must reply to it, or your subscription will not become active. No one can subscribe your name, and address, without you being notified, and if you get an unwanted notice of subscription you only need to DO NOTHING and the subscription will NOT be active.
REMOVAL: You can remove yourself from the subscription list in several different ways. Click here to read about this entire newsletter system. Every edition of The Wednesday Letter is delivered to your address with YOUR name and address in view on the letter, with a link that allows you to remove THAT name from the subscription list. If you try to send this removal message from an address different from the one you used to send in your original confirmation, then you will get a warning notice first, sent to the subscription address, asking you to confirm that you want to be removed from the list -- by replying to THAT request for confirmation, you will then be automatically removed. Thus, no one else can unsubscribe you, from some other computer, without your knowledge. But, if you send in the unsubscribe notice from the same machine used to receive the Letter, then the removal from the subscription list is automatic.
Personal Message: When you send a personal message to Karl Loren, you will receive a personal reply as per his instructions. Karl pledges that every personal message will get a personal answer. When you provide your mail address, we will send you free information including our free catalog and a cassette tape lecture by Karl Loren about heart disease, no charge, by mail, even if outside the US. You can select particular information you would like to receive, along with the free cassette tape and catalog.
You can reach Vibrant Life in many ways, including by mail to Vibrant Life, 2808 N. Naomi St., Burbank, CA 91504. Within the US and Canada, use the toll free number: (800) 523-4521, the local number: (818) 558-1799, the FAX: (818) 558-7299, eMail to kimberly@oralchelation.com or any one of the hundreds of message forms throughout the 50 web sites. Vibrant Life normally ships the same day we get an order. There are message forms on each of the 100,000+ pages on this and other sites where you can communicate with Vibrant Life. Check out our companion site, at: http://www.oralchelation.net where Karl's 2000 page book is published. Karl Loren is the author and webmaster for this BOOK, as well as for another web site about ORAL CHELATION. His personal philosophical articles are at PHILOSOPHY.
Copyright © May 20, 2008 6:24 AM by Karl Loren on behalf of Vibrant Life, ALL RIGHTS RESERVED. Permission is granted for non-commercial downloading, copying, distribution or redistribution on two conditions: One, that some form of copyright notice is included in every copy distributed or copied, showing the copyright belonging to Vibrant Life, Burbank, CA, at www.oralchelation.com . The second condition is that the material is not to be used for any purpose contrary to the purposes and objectives of this site. This permission does not extend to materials on this site which are copyrighted by others.