Real-Time Facial Feature Extraction by Cascaded Parameter Prediction and Image Optimization

We propose a new fast facial-feature extraction technique for embedded face-recognition applications. A deformable feature model is adopted, of which the parameters are optimized to match with an input face image in two steps. First, we use a cascade of parameter predictors to directly estimate the pose (translation, scale and rotation) parameters of the facial feature. Each predictor is trained using Support Vector Regression, giving more robustness than a linear approach as used by AAM. Second, we use the generic Simplex algorithm to refine the fitting results in a constrained parameter space, in which both the pose and the shape deformation parameters are optimized. Experiments show that both the convergence and the accuracy improve significantly (doubled convergence area compared with AAM). Moreover, the algorithm is computationally efficient.