Research And Application Of Facial Feature Points Detection

Active Appearance Model is a Parametric Statistical Model proposed by Cootes et al.It is an important algorithm in the field of computer vision,which has been widely used in modeling and analysis of two dimensional non rigid object,such as feature detection,recognition and posture correct,etc.Active Appearance Model uses the principal component analysis to build the shape model and the texture model of non-rigid object,and uses the objectives function optimization to match the new objects.The thesis makes a comprehensive and profound study of the Active Appearance Model in the face image along with the Local characteristic operator and Cascade Regression.The face image understanding is a technology that uses the computer technology,automation technology for digital face image to extract the high-level semantic information.Because the face image information is informative,non intrusive,simple acquisition equipment,etc,it has been widely used in face recognition,video surveillance,national defense security and many other fields,and becomes a hot issue in present research.A complete face image understanding process has five main components,face detection,face feature point localization,face alignment,face feature extraction,and face recognition and classification algorithm.Among these,facial feature point location is to precisely locate the facial features and the facial contour,and provides a large number of accurate geometric information for the whole automated identification program,which is the data base of the whole program.Active Appearance Model is a classical algorithm in the research of facial feature points localization.It can locate the feature points of human faces to obtain the shape and texture information,and extract their high-level semantic meanings.On the basis of the classical Active Appearance Model,in order to improve the matching precision and the robustness in uncontrollable environment,this paper puts forward a new model fitting strategy,its main innovations are as follows:First,we propose a Cascaded Regression based Active Appearance Model.Through a cascade regression structure,several weak regressions are connected in series to form a strong regression,which will improve model matching accuracy.Secondly,Structured Features Fusion for Active Appearance Model is proposed.According to the characteristics of local descriptors,we combine the shape-indexed local features and global texture information to enhance the performance of Active Appearance Model fitting algorithms in terms of generalization capacity and accuracy.Lastly,we propose an optimization algorithm based on cascaded regression model.We uses an Non-Parametric Shape Model Adopted in CR to optimize the initial face shape.Then,the Active Appearance Model is used to accurately locate the feature points of the target face.Experimental results obtain on the XM2 VTS and BioID face datasets show that,compared with the traditional Active Appearance Model fitting algorithm,the improved algorithm proposed in this paper has better robustness and higher matching accuracy,and the improved Active Appearance Model is effective.