An Automatic Approach to Facial Feature Extraction for 3-D Face Modeling

Creating a friendly human interface for visual communication has become a popular and significant topic. One key issue is to construct a three-dimensional (3-D) face model as a visual representation. A 3-D face model can present the desired 3-D facial images related to a specified person by using the realistic 3-D structure and texture description. In order to automatically model the human face, we propose a coarse-to-fine method to extract the facial features from a 2-D head-and-shoulder color image and to map these facial features onto a 3-D template model. A skin-color-based scheme is first adopted to extract salient regions for selecting the face region, where the salient region is invariant to different situations such as scaling, rotation, and skewing. To more precisely extract the facial features, face geometry and proportions obtained by training database are used for reference. After locating desired regions such as eyes, lip, nose, and eyebrows, the corresponding facial features in the located regions are extracted based on a feature detection method, where a morphology-based scheme is used to enhance the corner features. Finally the complete facial features are mapped onto the adopted shape model and each facial feature is adjusted automatically to the proper position according to the training data. Our experiments demonstrate the feasibility of the proposed approach.