Method Of Statistical Model-based Facial Feature Location

Facial feature localization is the key of realizing facial image analysis.It has been widely applied to face recognition, 3D facial reconstruction and face image compression. Active Shape Models(ASM) and Active Appearance Models(AAM) are effectively statistical tools for facial feature localization. Now they are the focus of image process and computer vision. The primary task of the thesis is to study the methods of facial feature localization based on ASM and AAM. The innovative achievements of the thesis include:1. An improvement ASM is proposed and this model is applied to facial feature localization. The basic principles and main characteristics of Principal Component Analysis(PCA) and Independent Component Analysis(ICA) are introduced in detail. In terms of the different optimization criterion in PCA and ICA, a new consecutive strategy of facial feature localization that combines PCA and ICA is presented. The PCA-based ASM is done for extracting facial features first and the global shape variabilities in the face images are searched. And then, the obtained results are refined using the ICA-based ASM, which show more local characters. Experimental results demonstrate that the improvement ASM not only can give an accurate localization of global face variations, but also can show localized face variations. Compared with the standard ASM, both the accuracy and the robustness are significantly improved. Meanwhile, the better face representation can be obtained under different perspective variations and facial expressions by this model.2. A new AAM—ICA-based AAM is studied and this model is applied to facial feature localization. The constructed process of the shape model, the texture model, and the appearance model is discussed. The parameters of texture model are analyzed in detail. Meanwhile, the principles and the realization of AAM search algorithm are concretely introduced .In the experiment, linear relationships of texture error, appearance parameter variations, model displacement , angle and scale are stated. Experimental results show that it is more accurate in localized shape localization and texture matching compared with the standard AAM. The ICA-based AAM is more...