A fisher discriminant framework based on Kernel Entropy Component Analysis for feature extraction and emotion recognition

This paper aims at providing a general method for feature extraction and recognition. The most essential issues for pattern recognition include extracting discriminant features and improving recognition accuracy. Kernel Entropy Component Analysis (KECA), as a new method for data transformation and dimensionality reduction, has attracted more attentions. However, as KECA only reveals structure relating to the Renyi entropy of the input space data set, it cannot extract effectively discriminant classification information for recognition. In this paper, we propose combining KECA and Fisher's linear discriminant analysis (LDA), utilizing descriptor of information entropy and scatter information of classes to improve recognition performance. The proposed method is applied to speech-based emotion recognition, and evaluated though experiments on RML audiovisual emotion databases. The results clear demonstrate the effectiveness of the proposed solution.