A method of visual speech feature area localization

In audio-visual bimodal man-machine interaction, extracting Region Of Interest (ROI) that carries visual speech features is a very crucial step. In this paper, our work about ROI localization is described in detail. First, we propose a simplified human skin color model to segment input images and estimate the location of human face. When we locate ROI in the human face area, the traditional linear methods' performance cannot satisfy system's need, especially for unseen subjects. Then we propose a new localization method that is a combination of Support Vector Machine (SVM) and Distance of Likelihood in Feature Space (DLFS) derived from Kernel Principal Component Analysis (KPCA). Results show that the new method outperformed traditional linear ones. All experiments are based on Chinese Audio-Visual Speech Database2 (CAVSD).