In this section, we make on overview of existing works on gender
recognition from face images.
Early works on gender recognition from face images focused on
the case of frontal faces in a controlled laboratory environment. In
the beginning of the 90s, many authors tried neural networks to deal
with this problem. For example, Golomb et al. [6] trained a 2-layers
fully-connected neural network and achieved 91.90% accuracy on a
tiny test set of 90 images. The benchmark dataset of frontal faces in
a controlled environment is FERET [20]. With the emergence of SVM,
Moghaddam and Yang [18] used this classifier with an RBF kernel on
raw pixels and obtained 96.62% accuracy on FERET (though having
the same persons presented both in training and test sets). Rather
than using SVM, Baluja and Rowley [2] used AdaBoost on raw pixels
and obtained 96.40% on FERET without mixing people in training
and test sets. Li et al. [15] combined facial information with clothing
and hair components obtaining 95.10% accuracy on the FERET dataset.
Ullah et al. [29] used the Webers Local texture Descriptor to reach almost
perfect performance of 99.08% on FERET. This result suggests
that the FERET benchmark is saturated and not enough challenging
for modern methods.