Transformation from a given set of measurements to a new set of features
(transform-based features).

pixels as features -> very redundant information 
                         due to a large degree of correlations
                         in real world images

1) Fourier transform: 
most of the energy lies in the low-frequency components, due to 
the high-correlation between the pixels;
low-energy, high-frequency coefficients can be neglected 
with little loss of information

2)Kalhunen-Loeve transform
a tool to select m dominant features out of N measurement samples

But it does not necessarily lead to maximum class separability 
in the lower dimensional subspace.