Information Technology Reference
In-Depth Information
user in processing or understanding the content of a complex signal, such as an image.
Usually, an image consists of thousands of pixels. This information can hardly be
quantitatively analyzed by the user. In fact, some problems related to the subjective
factor or to the tiredness of the user arise, which may influence reproducibility. There-
fore, an automatic procedure for analyzing an image is necessary.
Although in some cases it might make sense to process a single image and to adjust
the parameters of the image processing algorithm to this single image manually,
mostly the automation of the image analysis makes only sense if the developed meth-
ods have to be applied to more than one single image. This is still an open problem in
image processing. The parameters involved in the selected processing method have to
be adjusted to the specific image. It is often hardly possible to select the parameters
for a class of images in such a way that the best result can be ensured for all images of
the class. Therefore, methods for parameter learning are required that can assist a
system developer in building a model [1] for the image processing task.
While the meta-learning task has been extensively studied for classifier selection, it
has not been studied so extensively for parameter learning. Soares et. al [2] studied
parameter selection for the identification of the kernel width of a support-vector ma-
chine, while Perner [3] studied parameter selection for image segmentation.
The meta-learning problem for parameter selection can be formalized as follows:
For a given signal that is characterized by specific signal properties A and domain
properties B find the parameters P of the processing algorithm that ensure the best
quality of the resulting output signal/information:
(1)
f
:
A
B
P
Meta-data for images may consist of image-related meta-data (gray-level statistics)
and non-image related meta-data (sensor, object data) [4]. In general, the processing
of meta-data from signals and images should not require too heavy processing and
should allow characterizing the properties of the signal that influence the signal proc-
essing algorithm.
The mapping function f can be realized by any classification algorithm, but the in-
cremental behavior of case-based reasoning (CBR) fits best to many data/signal proc-
essing problems, where the signal-class cannot be characterized ad-hoc, since the data
appear incrementally. The right similarity metric that allows mapping data to parame-
ter-groups and, as a consequence, allows obtaining good output results, should be
studied more extensively. Performance measures that allow to judge the achieved
output and to automatically criticize the systems performance are another important
problem [5].
Abstraction of cases to learn domain theory would allow better understanding the
behavior of many signal processing algorithms that cannot anymore be described by
means of the standard system theory [6].
The aim of our research is to develop methods that allow us to learn a model for
the desired task from cases without heavy human interaction (see Fig. 1). The specific
emphasis of this work is to develop a methodology for finding the right image de-
scription for the case that group similar images in terms of parameters within the
same group and map the case to the right parameters in question.
Search WWH ::




Custom Search