Understanding human emotional states is indispensable for our daily interaction, and we can enjoy more natural and friendly human computer interaction (HCI) experience by fully utilizing human’s affective states. In the application of emotion recognition, multimodal information fusion is widely used to discover the relationships of multiple information sources and make joint use of a number of channels, such as speech, facial expression, gesture and physiological processes. This thesis proposes a new framework of emotion recognition using information fusion based on the estimation of information entropy. The novel techniques of information theoretic learning are applied to feature level fusion and score level fusion. The most critical issues for feature level fusion are feature transformation and dimensionality reduction. The existing methods depend on the second order statistics, which is only optimal for Gaussian-like distributions. By incorporating information theoretic tools, a new feature level fusion method based on kernel entropy component analysis is proposed. For score level fusion, most previous methods focus on predefined rule based approaches, which are usually heuristic. In this thesis, a connection between information fusion and maximum correntropy criterion is established for effective score level fusion. Feature level fusion and score level fusion methods are then combined to introduce a two-stage fusion platform. The proposed methods are applied to audiovisual emotion recognition, and their effectiveness is evaluated by experiments on two publicly available audiovisual emotion databases. The experimental results demonstrate that the proposed algorithms achieve improved performance in comparison with the existing methods. The work of this thesis offers a promising direction to design more advanced emotion recognition systems based on multimodal information fusion and has great significance to the development of intelligent human computer interaction systems.
Decision and feature level fusion of deep features extracted from public COVID-19 data-sets
