Digital Signal Processing Reference
In-Depth Information
2.1
Multilevel Bayesian Decision Fusion
Looking back to Equation (5), once a threshold is set in the likelihood
ratio test, one can claim that if the log likelihood ratio is much
larger or much smaller than the confidence of the decision is stronger.
Hence the absolute difference between the likelihood ratio
and the
threshold
can be used as a measure of confidence
In the multimodal scenario, the confidence measure can be used
beneficially in the decision fusion if we have enough a priori information on
the different modality streams. Let us define a multimodal scenario with
three different modalities and feature vectors and There are also
three streams of log likelihood ratios and
correspondingly. If we have a priori information such that the reliability of
the modalities are in an order, such that the first modality is the most
reliable and the last modality is the least reliable source under some
controlled conditions (such as low acoustic noise, frontal face stream, etc.),
then the confidence of the decision that is coming from first modality as
defined in (8) would be beneficial for the decision fusion. Keeping this fact
in mind a Bayesian decision system can be built. In this system a decision
tree is utilized as:
a)
A decision (accept or reject) is taken according to the modality if
the confidence measure that is coming from the most reliable
modality is high enough (i.e. if
Otherwise a decision is taken according to the modality with the
highest confidence among and if
Otherwise a decision is taken according to the modality with the
highest confidence among all three modalities.
Note that the decision scheme uses three confidence thresholds
and
b)
c)
d)
that have to be determined experimentally.
