One of the most important problem that exists in pattern recognition
and classification is the development of a decision surface. Modern
algorithms rely on sophisticated statistical methodologies to perform
classification. The simplest approach involves finding the K-nearest
neighbor using the Euclidean distance measure. However, Euclidean
distance is not the best measure if feature vectors have correlation
with unequal variances. To counter this problem, data is transformed
to a feature space where the classes are better separated and the
features are independent. This makes the distance metric more
effective. The signal modeling techniques implemented in this work
help define the new feature space.
Linear Discriminant Analysis (LDA) and Principal Components Analysis
(PCA) are two common methods used for multi-group data classification.
These functions typically use a linear transformation which can either
be implemented in class-dependent or class-independent fashion. PCA
is a feature classification method in which data is statistically
decorrelated in order that distributions have equal variances. LDA is
also a transform-based method that attempts to minimize the ratio of
within-class scatter to the between-class scatter, thereby maximizing
class separability.
The work presented here compares the performance of these two
algorithms on an image processing problem involving classification of
forestry images. On a standard evaluation task, consisting of 478
training images and 159 test images, class-dependent LDA produced a
50.41% misclassification rate. Unfortunately, PCA produced an error
rate of 39.7% on the same task. For this image processing task, in
which features have unequal variances, LDA does not appear to be good
classification technique, and would be better applied as a
post processing step on the PCA features.