Happiness Analysis with Fisher Information of Dirichlet-Multinomial Mixture Model.

Emotion recognition requires robust feature representation and discriminative classification models. In this paper, we consider Fisher vectors for feature representation and Fisher scoring algorithm for learning the proposed model. We first propose a new Fisher scoring algorithm using an exact Fisher information matrix for the Dirichlet-multinomial (DM) mixture model. Subsequently, we present an exact derivation of the Fisher vectors for images representation and we analyze the intensity of happiness from EMOTIC database by applying the proposed framework. The obtained results prove the effectiveness and the robustness using Fisher vectors for emotion recognition.