|
The proposed approach consists of two steps.
First, facial images are represented in a low dimensional space, spanned by the eigenvectors of the covariance matrix of the data, computed using PCA. Then, for each of the classifiers (Bayes classifier, Neural Network, Linear Discriminant Analysis and Support Vector Machines) , a GA is used select gender-related features automatically to reduce error rate.
It has been found in several studies that different eigenvectors encode
different kind of information.
For example, the first few eigenvectors seem to encode lighting
while other eigenvectors seem to encode features such as glasses
or moustaches . For example, Fig.1 shows some
of the eigenvectors computed from our training data. Obviously,
eigenvectors 1-4 encode light variations while eigenvectors 10 and
20 encode information about glasses.
Fig.1. Eigenvectors (from left to right and from top to bottom):
No. 1-6, 8, 10, 12, 14, 19, 20, 150, 200 and 250.
|
Genetic Feature Selection
In our encoding scheme, the
chromosome is a bit string whose length is determined by the
number of eigenvectors. Each eigenvector is associated with one
bit in the string. If the ith bit is 1, then the ith
eigenvector is selected, otherwise, that component is ignored.
Each chromosome thus represents a different eigen-feature subset.
The goal of feature subset selection
is to use fewer features to achieve the same or better
performance. Therefore, the fitness evaluation contains two terms:
(i) accuracy from the validation data and (ii) number of features
used. Combining these two terms, the
fitness function is given as:
fitness=10^4*Accuracy +0.4*Zeros
where Accuracy is the accuracy rate that an individual achieves,
and Zeros is the number of zeros in the chromosome.
Dataset
The dataset used contained 400 frontal images
from 400 distinct people, representing different races, with
different facial expressions, and under different lighting
conditions. The 400 images were equally divided between males and
females.
|
