A comparison of techniques for robust gender recognition

Automatic gender classification of face images is an area of growing interest with multiple applications. Appropriate classifiers should be robust against variations such as illumination, scale and orientation that occur in real world applications. This can be achieved by normalizing the images in order to reduce those variations (alignment, re-scaling, histogram-equalization, etc.), or by extracting features from the original images which are invariant respect to those variations. In this work we perform a robust comparison of eight different classifiers across 100 random partitions of a set of frontal face images. Four of them are state-of-the-art methods in automatic gender classification that use image normalization (SVMs, Neural Networks, ADABOOST and PCA+LDA). The other four strategies use invariant features extracted by SIFT (BOW, Evidence Random Trees, NBNN and Voted Nearest-Neighbor). The best strategies are SVM using normalized images and NBNN, the latter having the advantage that no strong image pre-processing is needed.