Master Thesis - Department of Computer Science

the response vectors on a validation set to compare with DP(x) (decision profile for a
test sample x). It can be stated that, these methods use prior knowledge about the
classifier’s behavior to improve combined performance. These methods will perform
well when classifier output preserves class specific contextual information. In case
if any classifier is highly sensitive to the features used and exhibits large intra-class
variance in the classifier output space, these methods do not perform well, and even
deteriorate beyond the performance of individual base classifiers.
In that case, the only way to overcome this problem is to conditionally improve
each row Di(x) (response vector of classifier Di) using response vectors of corre-
sponding classifier on a validation set as training data at fusion level and then use
class-conscious techniques to enhance the performance of classifier combination. Con-
ditional row-wise operation here implies the use of training data (generated on a vali-
dation set) as prior knowledge about classifier’s (Di) behavior to replace it’s response
vector Di(x) with ¯ Di(x) in DP(x) and thereby improving combined performance.
If any classifier’s output does not provide consistency in class-specific information,
training data does not help to improve the final decision. In such a case, it is wiser
to keep Di(x) unchanged for that classifier. This observation is the essence of our
approach to influence the final classification performance.
In our approach, we have accomplished the task of using training data (prior
knowledge about a classifier) by using the response vectors on a validation set as
training data at the fusion level for building a LDA or nonparametric LDA-based
eigenmodel. Now for a particular classifier, the ability of the eigenmodel in enhancing
class-separability in response vector eigenspace dictates the usefulness of training data
at fusion level. The improvement of class-separability by eigenmodel strengthens
the performance of it’s corresponding base classifier. This is evident from the fact
that LDA performs well when the training data uniformly samples the underlying
class distribution, or in other words training data represents class-specific information
[86]. Empirically, we can validate it by observing that the use of eigenmodel on
classifier’s output space provides better performance than the direct classifier output
on a validation set of response vectors.
To accomplish this task, we divide a database in four disjoint sets, namely train,
104