A Discriminant Correntropy Analysis For Multi-Feature Fusion

In this work, a discriminant correntropy analysis (DCA) method is proposed with application to multi-feature fusion. Benefiting from the joint strength of discriminant power and correntropy descriptor, not only is the discriminant representation explored but also the localized similarity is utilized to measure the structural relation between the given multiple features, generating a new multi-feature representation with high quality. Different from the most existing multi-feature fusion techniques, such as canonical correlation analysis (CCA) and kernel CCA (KCCA), the correntropy is used to reveal the intrinsic relation of input data sources instead of correlation. Moreover, unlike the traditional entropy-based algorithm (e.g., kernel entropy component analysis (KECA) method), DCA is able to be applied to multiple variables instead of a single data source only, enabling a more powerful tool for multi-feature fusion. The performance of the proposed DCA method is verified through experiments on audio emotion recognition and face recognition tasks. The results demonstrate DCA outperforms other deep neural network (DNN) and statistics machine learning (SML) based methods.