Feature selection for facial expression recognition based on mutual information

This study compares three different methods of feature selection for facial expression recognition from still images. The log-Gabor filter features were reduced to smaller sub-sets of representative features using three alternative approaches: selection of 4 optimal filters (log-Gabor/4-OF), principal component analysis (log-Gabor/PCA) and optimal feature selection based on the mutual information criterion (log-Gabor/MIFS). Six different facial expressions were considered. The selected features were classified using the K-NN classifier. The training and testing was performed using images from the Cohn-Kanade database. The percentage of correct classifications varied across different expressions from 23.1 % to 66.7 % for the log-Gabor filter/PCA approach, from 41.3 % to 96 % for the log-Gabor filter/4-OF, approach, and from 61.5 % to 96.1 % for the log-Gabor filter/MIFS method. The average correct classification rate was increased from 52.5 % for the PCA-based selection, to 68.9 % for the log-Gabor/4-OF, and to 75.5 % for the log-Gabor/MIFS method. An overall improvement in the discrimination between different facial expressions was observed when using the optimised (log-Gabor filter /MIFS) sub-set of features.