Biomedical Engineering Reference
In-Depth Information
the estimated error for the whole
strong
classifier is lower than a given error rate,
or if we achieve the desired number of
weaks
. The final decision is the result of
the weighted classification results of the
weaks
.
The most commonly used weak learner is the
decision stumps
. This simple
learner looks for the most discriminant feature of the training set and classifies
using a threshold. Formally, the weak classifier,
h
j
(
x
), consists of a feature
f
j
,
a threshold
θ
j
, and a parity value
p
j
. The classifier is trained using ROC curve
evaluation. Note that, although the threshold separates the two classes, it is not
enough to identify which class is on either side of the threshold. Therefore, a
parameter
p
j
(parity) is needed to indicate the direction of the inequality sign
when classifying:
1
if
p
j
f
j
(
x
)
<p
j
θ
j
,
h
j
(
x
)=
0
otherwise
.
Observe that the label resulting from the evaluation of the strong classifier is based
on the comparison of the weighted combination of weak results with a threshold.
If we omit this last comparison, we can consider that the number resulting from
this process is related to the confidence rate of that output. In fact, the greater the
number of weak classifiers that agree, the more extreme the value is. The output
of this process can be used to feed a STOP and GO active model.
4.4. Geometry-BasedGeneral Enhancement ofContinuous Potentials
Up to this point we have seen several scenarios with different spaces in which
STOP and GO active models can be used. However, several of them do not provide
maps accurate enough for feeding the STOP and GOmodels. In this sense, general
tools for further enhancing the results are useful. In this subsection we provide
an independent method for image and local contour enhancement based on the
geometry. The method is based on a new representation of an image proposed in
the work of Salembier and Garrido [30].
In their approach, Salembier and Garrido represent an image as a tree (
Max-
Tree
) composed by flat regions and linking information among regions. Each flat
region is a node
C
h
in the tree. The process for creating the tree is divided into
two steps:
Binarization step:
For each temporary node
TC
h
, the set of pixel belong-
ing to the local background is defined and assigned to the max-tree node
C
h
.
Connected components definition step:
The set of pixels belonging to the
complement of the local background (
TC
h
\
C
h
, where
is the set differ-
ence defined on connected components) are analyzed and its connected
components create the temporary child nodes
TC
h
+1
\
.
