Improving the Recognition Performance of Moment Features by Selection - Feature Selection for Data and Pattern Recognition - page 311

The aforementioned drawback of theGA-based selectionmakes the filter selection
methods [ 24 ] an attractive alternative approach. These methods do not use the mining
model (they are independent of the classification model), instead the internal data
properties/characteristics (dependency, correlation etc.) are taken into consideration.
For the sake of the experiments presented in the next section, the GA-based
(wrapper) and the Relief [ 15 ] algorithm (filter) selection methods will be applied
for the selection of the proper moment subsets that better discriminate the patterns
of some benchmark pattern recognition datasets. These two algorithms are briefly
discussed in the next sections.

13.4.1 GA-Based Selection

The main operational element of a Genetic Algorithm is the chromosome. The chro-
mosome corresponds to a candidate solution to the problem at hand, consisting of the
set of variables appropriately coded. For the case of the GA-based moment selection
method, the chromosome consists of the indices (Fig. 13.8 ) of a predefined number of
moments. The indices correspond to the moment id belonging to the moment feature
vector, which is constructed by arranging the computing moments according to the
zigzag scanning operation [ 11 ].
Initially, a pool of 100 moment features is constructed by computing all the
moments up to a specific order. Considering that a number of n moments are required
to be selected, the k th chromosome structure of the GA is depicted in Fig. 13.8 .
Furthermore, the objective function being minimized by the GA is equal to
the recognition error (Wrong Recognized Patterns/Total Patterns) derived when the
selected moment sets are fed to the classifier model.

13.4.2 Relief Algorithm

Relief algorithm [ 15 ] is a popular feature selection method due to its simplicity. It
is based on the computation of the relevance between pairs of feature vectors. The
relevance is measured by applying the L-dimensional Euclidean distance, where L
is the dimension of the feature vectors being compared. This algorithm selects those
features which are relevant subject to a defined threshold in linear time.

Fig. 13.8 Chromosome structure