Research On Facial Expression Synthesis Algorithm Based On Mesh Model

Facial expression is an important way to convey emotional information in daily interpersonal communications,and is a special means in addition to language communication.With the popularization of various smart devices and the rapid development of computer networks,a large volume of image and video data have emerged in our daily life,intelligent interpretation of these data plays a paramount role in computer vision,computer graphics,and the related applications,wherein facial expression synthesis become one of the hottest research topics.The diversity of facial contours and the complexity of the surface texture of the face make an expression vary from person to person,time to time,and conditions to conditions.The changes in facial expressions include not only the motion deformation of the overall facial features,but also the subtle changes of the local textures,which are often important visual cues for judging the expression categories,but synthesis of expression is a very difficult task in expression analysis.Therefore,it is a great challenge to maintain local details while controlling the geometric deformation of the face,and generating a new and realistic facial expression.Facial expression synthesis has potential applications in expression simulation,recognition and animation.Although much of the work mentioned in the current literatures is based on 3D models,the related algorithms usually require face matching and trimming,are thus they are time consuming.Based on the research of existing expression synthesis methods,this dissertation proposes a facial expression synthesis algorithm based on mesh model deformation in 2D space,and experiments are carried out on generally adopted database including the CK+ database and Multi-PIE face database.Firstly,this dissertation studies a moving least squares algorithm based on feature point control quadrilateral mesh deformation and its improved version called moving regularized least squares algorithm.Both of these algorithms establish a linear or nonlinear regression equation for the key feature points of an expression face image,control the movement of the vertices of the quadrilateral mesh,and then interpolate the mesh to obtain the values of the remaining pixels to achieve an expected image deformation.Although both algorithms can achieve a smooth and natural image deformation effect,they are only suitable for the deformation of small expressions,which may cause overlap for large expression deformation.Therefore,another algorithm based on Delaunay triangular mesh deformation using B-spline curve interpolation is proposed.Experimental results show that the proposed method can maintain the local area integrity,produces natural smooth image deformation,and adapts to both small expressions and large expressions.Secondly,based on the Delaunay triangular mesh deformation,a new model called Expression Ratio Image based on optical flow improvement is proposed.The method maps the details of the source expression to the corresponding position of the target neutral face image by fitting the difference in illumination between the source neutral face image and the expression face image.Experiments with expression synthesis involving large geometric deformation and illumination differences show that the algorithm can accurately maintain geometric deformation,and the synthesized expression is visually realistic.The expression synthesis method not only retains the facial features of the target person,but also contains the expression details of the source character,and produces natural,personalized and realistic expression.Finally,based on the basic expression generated by Delaunay triangular mesh deformation and ERI detail mapping,this dissertation uses an interpolation method to realize the transitional expression of existing facial expressions and generate new facial expressions.