Clusters of Features Using Complementary Information Applied to Gender Classification From Face Images

Face recognition performance by computers has been shown to be more accurate than that of humans. However, a bias with soft-biometrics features has been detected. This bias reduces recognition performance when gender is used. Feature selection for gender classification from face images is a difficult problem since faces contain symmetrical and redundant features. We argue that traditional methods, based on mutual information using pairs of features to estimate the relevance and redundancy among features, fail to select the right set of features in cases where there are strong spatial correlations among features, which is the case with facial images. In this paper, a new method is proposed fusing a filter and a wrapper to measure the relationships among image features, and to select feature clusters based on mutual information for gender classification. We applied this method on nine face datasets using an SVM classifier. We were able to achieve 98.2% correct gender classification in the testing partition using the UND, 95.56% with the Morph II, 98.33% on the LFW, and 98.66% on celebA databases. We validated the results using a cross-test with three different datasets: COFW, Adience, and Image of Groups, that were not used to define the parameters of our method. Additionally, the method was tested with a Random Forest. All the results achieved are better than those previously published on the same databases, and with a significantly smaller number of total features.