Neural-network-based gender classification using genetic search for eigen-feature selection

We consider the problem of gender classification from frontal facial images using feature selection and neural networks. We argue that feature selection is an important issue in gender classification and we demonstrate that by removing features that do not encode important gender information from the image representation of faces, the error rate can be reduced significantly. Automatic feature subset selection is used. First, principal component analysis (PCA) is used to represent each image as a feature vector (i.e., eigen-features) in a low-dimensional space, spanned by the eigenvectors of the covariance matrix of the training images (i.e., coefficients of the linear expansion). A genetic algorithm (GA) is then used to select a subset of features from the low-dimensional representation by removing certain eigenvectors. Finally, a neural network is trained to perform gender classification using the selected eigen-feature subset. Experimental results demonstrate a significant improvement in error rate reduction. Using a subset of eigen-features containing only 18% of the features in the complete set, the average NN classification error goes down to 11.3% from an average error rate of 17.7%