Biomedical Engineering Reference
In-Depth Information
Figure 9.
Filter pipeline for the confidence rate approach of the STOP and GO active
models.
4.3. Confidence Rate Approach
The confidence rate approach is based on the output of a classifier. Several
classification techniques are able to provide the confidence of the output label.
The confidence rate is a measure of the certainty that the algorithm has in its
label choice. According to this confidence rate we can build a
confidence map
in the same way we created the likelihood map. Given a classification technique
h
(
x
), we consider that the output of this function is
y
=
{
C,
c
r
}
, where
C
is the
label assigned to each feature vector
x
and
c
r
is a vector of the confidence rates
associated with each of the class labels. Of course, the label selected is the label
associated with the maximum value of confidence rate. The confidence map is
the result of classifying each feature vector associated with each pixel
I
(
x, y
) of
a given image
I
,
C
(
x, y
)=
h
(
x
(
I
(
x, y
)))
Many classification schemes are able to produce these confidence rates —-
e.g., adaptive boosting (adaboost), support vector machines, etc. In this section we
introduce the one we have selected to demonstrate the performance of the STOP
and GO models with confidence rate spaces.
We will focus the following paragraphs on the adaboost process. The adaboost
algorithm is an ensemble method for supervised classification. The basic idea of
the method is to combine a set of
weak
classifiers until some desired low training
error has been achieved. The combination of the
weaks
is done using a weight
independent of error associated with that weak.
Adaboost defines a systematic way to create these
weaks
: each feature point
in the training set has a weight associated with it. This weight depends on how
accurate the data point is being classified up to that point by the combination of
weaks — the combination of weaks will be called
strong
from now on. If a data
point is accurately classified, then its probability of being used by subsequent
