Statistical Approach to Lip Segmentation Using Local MRF Model

Lip segmentation plays an important role in a lipreading system as it is the initial part of the whole processing, and its result can directly influence the performance of feature extraction and even classification and recognition. In this paper we present a lip contour segmentation based on local MRF models. For localization processing, an ellipse contour surrounding the lips is initialized with the minimum-bounding box. In our proposed method we specify the local region at every site of the ellipse contour with an appropriate radius. Within those local regions, MRF model performs lip segmentation robustly. Compared with traditional MRF model, a significant improvement of lip segmentation effectiveness is achieved for the local MRF model. The experimental results are very promising and have proved favorable performance during the segmentation process. Predictably this will enhance the subsequent feature extraction effectiveness and visual speech recognition rate considerably.