Due to the high robustness, Active Appearance Model algorithm (AAM) has been adopted to find face feature points. However, if the initial position of the AAM does not properly cover these feature points, AAM will converge to an illogical location. To resolve this problem, this paper develops salient region constraints on AAM initialization for facial feature extraction. In this approach, we first identify the mouth region and adopt a mouth-constrained mean model as the initialization of the AAM. By this approach, the initial position can be placed closer to the target. The experiments show that the method is effective.