4. Proposed method
In this section, the Fukunaga Koontz transform (FKT) method is
first reviewed and then the proposed multi-class Fukunaga Koontz
discriminant analysis (FKDA) is described.
The traditional FKT is similar to PCA in terms of dimensionality
reduction but does a better job in discriminating two classes. The
FKT incorporates data from both positive and negative classes and
uses eigen decomposition on the joint correlation matrix in order
to find the optimal basis vectors that very well represent one class
while having the least representation power on the other class.
Let XARdm be the data set containing the images from Class 1,
with each column a vectorized image with dimension d. Let YA
Rdn be the data set containing all the images from Class 2. Both X