Face Animation Research

As a hot research topic in computer vision and computer graphics,facial animation technology has broad application prospects in many industries such as animation and games.As the most common facial animation method,the Blendshape Face Model,can parameterize the deformation of the mesh model easily and efficiently.However,this method has long faced two complex problems:one is how to produce a set of complete and realistic expression base models,and the other is how to analyze the video signal in real time to obtain Blendshape weihts.In response to the above two issues,the paper has carried out the following work:1.Aiming at the high realism requirement of real face expression model,the automatic expression base acquisition process based on face reconstruction is realized.The specific implementation method is as follows:firstly,the muti-view 3D reconstruction technique is used to obtain face model at sub-millimeter-precision level,and then the non-rigid iterative nearest point algorithm is used to register the model to obtain a uniform topology,and finally the deformation migration algorithm is used generate a set of complete and realistic expression base models.2.Aiming at the high flexibility of stylized characters on facial expression models,an efficient expression base acquisition process based on rigging technology is proposed.The specific implementation method is to first create a scientific and reasonable facial bone hierarchical structure,and then correctly assign the skin weights of the vertices at different joints,and finally create a user control interface that supports parameter input.By adjusting the parameters of each skeletal deformer,the mesh model is flexibly and parametrically deformed,and a set of expression-based models that are complete and conform to the characteristics of stylized characters are obtained.3.In order to extract expression parameters from video images in real time,a learning-based Blendshape weights solving method is used.The traditional expression parameter calculation method based on optimization algorithm needs to perform three-dimensional modeling on the performer.Firstly,the expression parameter of the specific user is calculated by optimizing the error function between the position of the 3D feature point projection of the performer and the position of the feature point detected by the input image.Then,through facial expression cloning technology to achieve facial animation.Unlike the optimization algorithm-based approach,the learning-based method does not treat the performer as a specific user,and does not require three-dimensional modeling and additional pre-processing.Firstly,the position of key points in the face image are obtained based on face detection and face alignment algorithm.Then a trained model are used to perform expression classification and intensity regression on key points to obtain action unit parameters.Finally,the method creates a mapping relationship between action units and the base models to analyze the video signal in real time and obtain expression parameters.