The discriminant analysis procedure breaks down when the within-class
scatter matrix 
becomes degenerate, which is our case due to a high
dimension of the input image and a much smaller number of training
samples. Weng [43] proposed DKL projection (short for
Discriminant Karhunen-Loeve projection). In the DKL projection, the
discriminant analysis is based on the space of Karhunen-Loeve projection,
where the degeneracy
typically does not occur.
For the details of Karhunen-Loeve projection of fovea vectors, the
reader is referred to [15].