Integration of Discriminative Information from Expressive and Neutral Face Image for Effective Modelling of Facial Expression Classification Problem

The contribution of appearance feature variations between expressive and neutral face image towards facial expression classification problem has received limited attention in the literature. The prime focus of the proposed work is to investigate the abstract, robust and discriminative feature space to effectively model expression classification problem. The significant contributions of the work are, hybrid feature space developed by the integration of discriminative features derived from histogram of oriented gradients (HOG) and local gabor binary pattern histogram sequence (LGBPHS) feature descriptor; computing the shape and texture feature variations between expressive and neutral face image (feature difference); novel feature space developed by combining the hybrid feature space acquired from the expressive face image with the feature difference; stacked deep convolutional autoencoder is employed as an efficient feature selection algorithm for dimensionality reduction and multiclass support vector machine (MSVM) for classification. The combination of HOG and LGBPHS has improved the recognition accuracy, robustness, and generalization capability of the model. The work is carried out on three benchmark datasets (CK+, KDEF, and JAFFE) and the model has shown remarkable recognition rates on all the three benchmark datasets (96.43% on CK+ , 96.03% on KDEF, and 88.53% on JAFFE).