3D Facial Expression Transfer Based On Personalized Detailed Features-preserving

With the rapid development of 3D computer graphics in recent years,3D facial expression transfer has attracted extensive attention of many scholars.At present,the expression transfer method has been widely used in many fields such as computer animation,special effects,interactive virtual simulation,teleconference.3D facial expression transfer technology effectively avoids the tedious work of the animators for making expression animation sequences for new models and improves the reuse rate of existing animations and the synthesis efficiency of new animations.Therefore,the expression transfer method can provide the new ways and methods for the generation of realistic expression animations.A good 3D facial expression transfer technology must meet three main indicators:1)the process of facial expression transfer does not require manual intervention;2)the facial expression of the target model is natural and has the rich detailed information;3)the expression transfer is real-time.These indicators directly affect the application of the technologies in relevant practical projects and engineering.Most of the existing facial expression transfer methods and processes have some defects in the above three indicators.Therefore,we propose a 3D facial expression transfer method with personalized detailed features-preserving,in order to achieve efficient,robust,high-realistic and automatic expression transfer.The specific work and innovations of this paper are as follows:1.We propose a 3D facial expression transfer technique based on an improved Co-learning method.The dimensionality reduction by unsupervised regression method used in Co-learning method has high time complexity.In order to improve the learning efficiency of expression subspace,we introduce the parametric dimensionality reduction by unsupervised regression method to solve the mapping function.In order to further shorten Co-training time before expression synthesis,we improve the projection process of the Co-learning method,and replace the gradient descent method with the Gauss Newton method,which converges faster.Compared with the traditional expression transfer methods,the proposed expression transfer method not only reduces manual intervention in the transfer process,but also realizes expression transfer in real-time.2.We propose a new 3D expression transfer algorithm based on detailed feature extraction and multi-scale reconstruction,which can well reconstruct the personalized detailed features of the facial expressions.1)Detailed feature extraction.We use the Laplacian smooth algorithm to extract the detailed features such as wrinkles and folds of the model,which can decompose the facial expression animation into the large-scale global deformation and the fine-scale detailed motions.2)Multi-scale reconstruction.The large-scale global deformation represents the global deformation caused by the facial muscle movement,and the fine-scale details movements represent the changes in the fine expression details such as wrinkles and folds caused by the local skin deformation.We use the facial expression transfer method based on the improved parametric dimensionality reduction by unsupervised regression method to transfer the large-scale global deformation,so as to transfer the facial expression of the source model to the target model,and propose a new geometric vector mapping strategy to reconstruct the fine-scale detailed motion of the target model.In the fine-scale detailed motion reconstruction stage,we present an objective function based on normal constraints and displacement constraints to solve the geometric vectors for reconstructing details.By reconstructing the detailed motions of the target model after the large-scale global deformation transfer,it can effectively restore the personalized detailed features of the target model and make the generated expressions more real and natural.This paper proposes a 3D facial expression transfer method based on the improved Co-learning method and a 3D facial expression transfer method based on detailed feature extraction and multi-scale reconstruction.The methods reduce manual intervention in the process of expression transfer,improve the speed of expression transfer,and reconstruct the personalized details of the target model with high precision,making the generated expressions real and natural.Exploring the abstract representation of the expression,and further improving the processing ability of our method in the extreme expressional deformation cases is our future research work.