Face gender classification: A statistical study when neutral and distorted faces are combined for training and testing purposes

•Study of gender recognition from neutral, expressive and occluded faces•Comparison of global/local approaches, grey level/PCA/LBP features and three classifiers•Three statistical tests over two performance measures are employed to support the conclusions.•Local models surpass global ones with different types of training and test faces.•Global and local models perform equally with the same type of training and test faces.

This paper presents a thorough study of gender classification methodologies performing on neutral, expressive and partially occluded faces, when they are used in all possible arrangements of training and testing roles. A comprehensive comparison of two representation approaches (global and local), three types of features (grey levels, PCA and LBP), three classifiers (1-NN, PCA + LDA and SVM) and two performance measures (CCR and d′) is provided over single- and cross-database experiments. Experiments revealed some interesting findings, which were supported by three non-parametric statistical tests: when training and test sets contain different types of faces, local models using the 1-NN rule outperform global approaches, even those using SVM classifiers; however, with the same type of faces, even if the acquisition conditions are diverse, the statistical tests could not reject the null hypothesis of equal performance of global SVMs and local 1-NNs.