Research On Facial Feature Extraction

Human facial feature extration is the key of facial image analysis, which has been widely applied to face recognition, 3D facial reconstruction and face image compression. It's not easy to model the facial features, which have various apperances for different poses and facial expressions. Also it's very hard to locate the facial features, because the uneven lightings and occlusions in image make them disappear or deform. Like other facial image analysis techniques, facial features location needs to adapt to those various changes farthest. Although there are so many the further research on this problem, the previous techniques usually require even lightings, frontal and neutral-expression faces in image which prevent them from practical usage on speed and precision.In this dissertation, robust facial feature extraction is studied, which is detailed as follows.1) Firstly, Inverse Compositional Algorithm and its three important extensions are deduced. Secondly, the modeling and the fitting algorithm of active appearance model (AAM) are studied, and the disadvantages of the facial feature location based on AAM are point out.2) The Project-Out Inverse Compositional Algorithm based on AAM is an efficient mothod for facial features location. However, when the partial face is occluded in image, the accuracy and efficiency of algorithm become worse. An efficient Robust Normalization Inverse Compositional Algorithm with layered mask subdivision to eliminate disturbance is proposed. The algorithm not only keeps the superiority of the original Inverse Composition Algorithm but also enhances the ability of processing occlusion. It extends Project-Out Inverse Compositional Algorithm in two ways. (i) We present the Robust Normalization Inverse Compositional Algorithm which is applicable to mask technology. (ii) Layered mask subdivision is developed. The mask scale becomes gradually small to fix on the right place by estimating and blocking in the iterative process. 3) When head pose in image rotates in a large scale, the accuracy of facial features localization becomes worse for traditional 2D model AAM. In this thesis, 3D facial feature extraction algorithm based on 2D+3D Candide is presented, which can efficiently extract 3D information about the position of facial features and head pose. Experimental results show that the algorithm is faster and more robust. Innovations are as follows. (i) We unify AAM and Candide, thus 2D+3D Candide is established based on AAM. (ii) A constraint equation is established for keeping the deformation of 3D shape model consistent with 2D shape model. (iii) A novel algorithm is presented to optimize energy equation and constraint equation simultaneity.