Score Equalization in SVM Multimodal Fusion for Person Recognition

In the classificatory approach for multimodal score fusion, conventional techniques are commonly used for the normalization of features or scores. In this work, in addition to the conventional methods, two equalization methods, histogram equalization, which was recently introduced in multimodal systems, and Bi-Gaussian equalization, which is presented in this paper, are applied upon the scores provided by prosodic, speech spectrum and face experts in a SVM based multimodal person verification system. The equalization techniques have obtained the best results with both RBF and polynomial kernels; concretely, for the RBF kernel, Bi-Gaussian equalization outperforms in more than a 22.19 % the results obtained by Min-Max normalization, the most used normalization technique in SVM fusion systems. The prosodic and speech spectrum scores have been provided by speech experts using records of the Switchboard I database and the face scores have been obtained by a face recognition system upon XM2VTS database.