An initial position correction and model instance selection method for AAM based face alignment

Active appearance models (AAM) is very useful for extracting attention objects from objects, e.g. faces from images. Traditional improved methods of AAM based face alignment always concentrate on fitting efficiency without any concrete analysis of characteristics of the initial position and model instance, thus the accuracy and speed are both not ideal when the face has a certain degree of deflection. An initial position correction and model instance selection method based on facial features detection and simple 3D pose estimation is proposed in this paper. Adaboost algorithm was applied to pre-detection of facial features in the images firstly, then features of images that could not be detected or had been incompletely detected were extracted by facial skin properties in YCbCr color space. Finally, we calculated the coordinate of the nose tip and deflecting angle of the face according to feature region, next properly adjusted the AAM fitting center position and model instance and introduced linear algebra software ATLAS into fitting process for matrixes optimization. Simulation experiments on IMM face database show that our method increased the fitting accuracy rate by about 43% and the time consumption was decreased by about 76% comparing with standard AAM algorithm.