Gender Classification by Information Fusion of Hair and Face

Various gender classification methods have been reported in the literature. These existing methods fall into two categories. The first kind of method is the appearance-based approach. Golomb et al. [1] used a two-layer neural network with 30 × 30 inputs and directly fed the scaled image pixels to the network without dimensionality reduction. Their database contains only 90 images with half male and half female facial images. Gutta et al. [2] used the mixture of experts combining the ensembles of radial basis functions (RBF) networks and a decision tree. Xu et al. [3] applied Adaboost to gender classification problem with the feature pools composed of a set of linear projections utilizing statistical moments up to second order. Wu et al. [4] also adopted Adaboost. Instead of using threshold weak classifiers, they used looking-up table weak classifiers, which are more general and better than simple threshold ones due to stronger ability to model complex distribution of training samples. Moghaddam and Yang [5] demonstrated that support vector machines (SVMs) work better than other classifiers such as ensemble of radial basis function (RBF) networks, classical RBF networks, Fisher linear discriminant, and nearest neighbor. In their experiments, the Gaussian kernel works better than linear and polynomial kernels. However, they did not discuss how to set the hyper-parameters for Gaussian kernel, which affect the classification performance. Kim et al. [6] applied Gaussian process technique to gender classification. The advantage of this approach is that it can automatically determine the hyper-parameters.Wu et al. [7] presented a statistical framework based on the 2.5D facial needle-maps which is a shape representation acquired from 2D intensity images using shape from shading (SFS). Saatci and Town [8] used an active appearance model (AAM) of faces to extract facial features and developed a cascaded structure of SVMs for gender classification. Lian and Lu applied min-max modular support vector machine to gender classification and developed a method for incorporating age information into task decomposition [9]. They also proposed a multi-resolution local binary pattern for dealing with multi-view gender classification probelms [10]. The second kind of approach is the geometrical feature based approach. The idea is to extract from faces geometric features such as distances, face width, and face length. Burton et al. [11] extracted point-to-point distances from 73 points on face images and used discriminant analysis as a classifier. Brunelli and Poggio [12] computed 16 geometric O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg