Gabor Feature Selection and Improved Radial Basis Function Networks for Facial Expression Recognition

This paper presents an improved radial basis function neural network with effective Gabor features for recognizing the seven basic facial expressions (anger, disgust, fear, happiness, sadness, surprise and neutral) from static images. The proposed improved RBF networks adopt a sigmoid function as their kernel due to its flexible decision boundary over the conventional Gaussian kernel. This study uses an M-estimator instead of the least-mean square criterion in the network updating procedure to enhance the network robustness. A growing and pruning algorithm adjusts the network size dynamically according to the neuron significance. Additionally, entropy criterion selects informative and non-redundant Gabor features. This feature selection reduces the feature dimension without losing much information and also decreases computation and storage requirements. The proposed improved RBF networks have demonstrated superior performance compared to conventional RBF networks. Experiment results show that our approach can accurately and robustly recognize facial expressions.