Study On Regularized Moving Least Square Method For Face Image Deformation

Face image deformation is an image processing technique that generates deformation mapping functions to produce smooth and natural deformation effects.This technology is widely used in many fields,such as,virtual reality,film animation production,image fusion and medical image processing.Firstly,the development of face image deformation technology is introduced.The mainstream face image deformation algorithms are classified and compared.In this thesis,the face image deformation technique based on Control Point's Moving Least Squares(MLS)is studied and improved.And the main flow of the face image deformation algorithm based on MLS is discussed.Then,the mathematical model of MLS algorithm is studied and three kinds of mapping transformations of MLS algorithm(affine transformation,similar transformation and rigid transformation)are derived.In order to improve the deformation quality of the above traditional MLS algorithm,this thesis proposes Regularized Moving Least Squares(RMLS)algorithm.In order to quantify and compare the deformation quality of various algorithms,this thesis proposes a measurement method,which uses Bezier curve to fit the face contour curve.The performance of different algorithms is compared with the error of the actual contour curve and the target contour curve.The simulation proves that the RMLS algorithm can obtain lower contour error than the MLS algorithm.Finally,the optimization scheme of face image deformation in video stream is studied.This thesis proposes a triangulation-based RMLS control point deformation algorithm,which simplifies the grid of rectangular splitting in RMLS algorithm and reduces the computational complexity.Since MLS algorithm and The RMLS algorithm does not consider the influence of face deformation on the background,this thesis adds local constraints to optimize the deformation quality.For the problem of face deformation jitter in the video stream scene,the stability function is used to stabilize the face feature points.And the contour error is measured to prove its validity.