Personal tools
 

A Meta-Learning Framework for Improving Segmentation Evaluation

— filed under:

Jason Fritts, SLU

What
  • Computer Science Seminar
When Mon, Oct 17, 2005
from 03:10 PM to 04:00 PM
Where RH 202
Add event to calendar vCal
iCal

Abstract: Object segmentation is a fundamental early step in many target recognition systems. A good segmentation of the object(s) from the optical image or sensor data greatly increases recognition accuracy. Many image segmentation methods have been used, but there is still no satisfactory means for measuring the effectiveness of these various methods. The lack of a good evaluation measure makes it hard to compare different segmentation methods, or even different parameterizations of a single method, thus hindering further improvements in the accuracy of recognition systems. It is well known that evaluating the effectiveness of object segmentation is a hard problem. Oftentimes, segmentation effectiveness is judged solely on the recognition rate of the imaging system, which is inaccurate because the performance of the segmentation itself is not necessarily directly correlated with overall system performance. A few standalone evaluation methods have been proposed, but these techniques examine different fundamental criteria, or examine the same criteria in varying ways. Consequently, these methods work well in some cases, but poorly in the others. Using machine learning methods, we can determine the circumstances under which the various metrics perform well versus when they perform poorly, and thereby leverage the appropriate evaluators’ quality measures to achieve reliable evaluation results. Consequently, we propose a meta-learning framework for segmentation evaluation, in which different effectiveness measures judge the segmentation performance according to their various criteria, and then a meta-learner coalesces the results to generate an overall effectiveness score based on the image characteristics. This co-evaluation framework is being targeted for dual-use, so segmentation evaluation results will be presented across a variety of single- and multi-spectral images in both the military and commercial domains.

« December 2017 »
December
SuMoTuWeThFrSa
12
3456789
10111213141516
17181920212223
24252627282930
31