A Bayesian Model for Extracting Facial Features

A Bayesian Model (BM) is proposed in this paper for extracting facial features. In the BM, first the prior distribution of object shapes, which reflects the global shape variations of the object contour, is estimated from the sample data. This distribution is then utilized to constrain and dynamically adjust the prototype contour in the matching procedure, in this way large or global shape deformations due to the variations of samples can be tolerated. Moreover, a transformational invariant internal energy term is introduced to describe mainly the local shape deformations between the prototype contour in the shape domain and the deformable contour in the image domain, so that the proposed BM can match the objects undergoing not only global but also local variations. Experiment results based on real facial feature extraction demonstrate that the BM is more robust and insensitive to the positions, viewpoints, and deformations of object shapes, as compared to the Active Shape Model (ASM) algorithm.