Research On Facial Feature Points Locating Technology Using Active Appearance Model

In 1998, F.T. Cootes et al proposed the AAM (Active Appearance Model) and used it to locate human facial feature points. In 2001, S. Baker et al of CMU improved original fitting algorithm of AAM. In the improved algorithm, the Hessian and grads can be caculated in advance and the interative computational complexity reduces. AAM has been regarded as a kind of effective human facial feature points locating methods because it has good 3D expansibility, good locating effect and fast speed which reaches 230 frames per second. Our research follows the work of F.T. Cootes and S. Baker et al, and has analysed the problems in detail which will be met in application, then proposes the improved method.We have improved the technology of facial feature points location using AAM in three aspects: 1) We have proposed the multi-scale mask method for the problem that locating effect becomes worse when some other objects occlude any parts of human face. In the method, the different areas are obtained by calculating the difference between the facial appearance of the input image and the average facial appearance of AAM model, and the results decide the scale of the original mask. Then in the iterative process, the mask scale gradually becomes smaller to fix on the right place by estimating and blocking, and the disturbance is to be eliminated. 2) For the problem of illegal transformation, we use shape limitation to deal with it. In the method, we project the sample shape of face in the training set to the shape vecotor of AAM model, then get a shape coefficient area. We set shape limitation with the shape coefficient area and use it to amend the change of shape coefficient during facial feature points location, then eliminate the illegal transformation. 3) Adopt the method based on statistics mask of the human facial complexion to solve the sensitive problem of the background. The results show that the improved algorithm can be used to locate facial feature points in different background.Finaly, we have developed the system of locating facial feature points using the improved method which we proposed in this paper. Through designing each function module rationally and adopting Matlab tool to carry on system development, the system developed in this paper not only is suitable for being as the experiment platform of follow-up study, but also has settled good foundation for developing the application software in future.