Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 220

System Theory

System Transfer Function H(x,y)

I(x,y)

O(x,y)

O(x,y)=I(x,y)*H(x,y)

Input Signal

Output Signal

Method

Meta Learning

Signal Description

System Behavior

Quality Measure

f : A ∪ B →

P

Fig. 1. Problem description in modeling

In section 2, we describe the image segmentation based on case-based reasoning.
In section 3, we review the theoretical and behavioral aspects of the Watershed Trans-
formation and how to use case-based reasoning to form the image segmentation
model. An overview about some of the test images and the corresponding best seg-
mentation parameters is given in section 4, where also the problems concerning the
evaluation of the results are briefly addressed. The derived image descriptions from
the theoretical and behavioral study in section 3 are given in section 5. A discussion
of the result is done in section 6. The process of generalization is described in section
7. Finally, we give conclusions in section 8.

2 CBR-Based Image Segmentation

2.1 Case-Based Meta-control of Image Segmentation

The segmentation problem can be seen as a classification problem for which we have
to learn the best classifier. Depending on the segmentation task, the output of the
classifier can be the labels for the image regions, the segmentation algorithm selected
as the most adequate, or the parameters for the selected segmentation algorithm. In
any case, the final result is a segmented image. The learning of the classifier should
be done on a sufficiently large test data set, which should represent the entire domain
well enough, in order to be able to build up a general model for the segmentation
problem. However, often it is not possible to obtain a sufficiently large data set and,
therefore, the segmentation model does not fit the entire data set and needs to be ad-
justed to process new data. We note that a general model does not guarantee the best
segmentation for each image; rather, it guarantees an average best fit over the entire
set of images.
Another aspect of the problem is related to the changes in image quality caused by
variations in environmental conditions, image devices, etc. Thus, the segmentation
performance needs to be adapted to changes in image quality. All this suggests the use
of case-based reasoning [7] as a basic methodology for image segmentation, since
CBR can be seen as a method for problem solving as well as a method to capture new