European Journal of Scientific Research - EuroJournals

Feature Selection Based on Statistical Analysis 429
KG-rule retains only those PCs whose variances, i.e. eigenvalues that are ≥ 1. Nevertheless, for
large variable spaces p, the KG-rule usually retains too many PCs [4], [8][10].
c) Cumulative Variance utters the criterion for choosing m is to select a cumulative variance
threshold, t where t is at certain percentage of the total variance that the first m PCs should account
for. The required number of PCs is then the smallest value of m for which the chosen percentage is
exceeded [7][10]. From PCA theory, the variance of the i-th PC (eigenvector) is equal to its
corresponding eigenvalues λi [6].
Three different methods have been elucidated on how to retain the optimal number of PCs. In
this study, we investigated the required number of PCs to be retained using these three methods.
Statistical Analysis
ANOVA is a standard technique for measuring the statistical significance of a set of independent
variables. It takes a single feature and the class associated with the data samples and measures the
significance of the class variables in predicting the means of the feature. The measure that ANOVA
produces is the p-value for the feature set. On the other hand, Multiple Comparison Procedure (MCP)
test is one that can be used to determine which means amongst a set of means differ from the rest. The
ANOVA leads to a conclusion that there is evidence that the group means differ but the main goal is to
determine which of the means are different. Therefore, the MCP is applied.
The MCP test compares the difference between pair of means with appropriate adjustment for
the multiple testing. The results of MCP present the p-value or confidence interval for each pair. In this
study, after performing ANOVA on the eigenpostures, MCP is applied to assess which ‘eigenpostures’
means are significantly different. Further, multiple range test is realized for testing homogeneous
subsets of groups based on their group means. In doing so, the groups that differ significantly are
revealed. Finally, we will determine the optimized number of eigenpostures that will act as inputs to
the ANN for classification of the four main postures.
Artificial Neural Network (ANN) as Classifier
Artificial Neural Networks (ANN) are known for their ability to express highly nonlinear decision and
makes them appropriate for recognition of complex pattern and the ability to maintain accuracy even
when some input data are inapt. In this study, the multilayer perceptron is chosen for recognition
purpose. A multilayer perceptron is a feed forward network structure in which neurons are connected
only between two succeeding layers [9][12]. A feed forward neural net that performs the recognition
part consist of one input layer with one neuron per feature, two hidden layers with four and three nodes
in each layer and a single output.
Experiments and Results
A collection of 400 images of various human postures constitutes the database to generate the
eigenpostures for this study. The various postures include all four main postures namely standing,
sitting, bending and lying position for both gender with the subjects are either facing front or facing
either side and no restriction impose on the type of clothing being worn. Initially, each image has m x n
pixels, but eventually reshaped to a column vector of x mn. Then, the eigenvectors and eigenvalues are
computed according to [3][11]. Implementing the three rules mentioned previously, we select the most
suitable eigenpostures required as inputs to the classification system. In other words, we select the
most relevant eigenvalues or PCs to be retained in this study.
Figure 3 illustrates the results of the Scree test. The decrease in magnitude for successive
eigenvalues implies that the first few principal components can approximate a large part of the original
data’s variance. In this case, decision to retain the first thirty-five PCs is appropriate and they
reasonably represent good approximation of the original data set. Next, from the PCA results, applying