Automatic Segmentation of Lip Images Based on Markov Random Field

This paper addresses the problem of lip segmentation in color space that is a crucial issue to a successful lip-reading system. We present a new segmentation approach to lip contour extraction by taking account of the maximum a posterior Markov random field (MAPMRF) framework. We first examine various color models and select a simple color transform derived from LUX and 1976 CIELAB color space as an effective descriptor to characterize the lip region by its discriminative properties. Thus, the initial label set with respect to lip and skin region is available. Based upon the identified lip area, we further refine the lip region using both color and label information, as those two are combined within a Markov random field (MRF) framework. Finally, we extract the lip contour via convex hull algorithm with the prior knowledge of the mouth shape. Experiments show the efficacy of the proposed approach in comparison with the existing lip segmentation methods.