The Research Of Facial Morphable Model Based On3D Scanning

Morphable model is based on the3D face database and the statistics. It has strengthenson high realism of model and automation of3D face synthesis with a2D face image. Becauseof3D face synthesis widely used in many fields, such as information security, film andadvertisement, human computer interaction, computer aided instruction, morphable model hasbecome into a research focus in this area. This thesis just focuses on constructing morphablemodel. It mainly includes acquisition of prototypical3D face data, preprocessing andstandardizing of the data and structure designing of the database, and3D face synthesis. Wesummarize the relevant literatures sufficiently, propose our novel ideas on critical process ofconstructing morphable model, and show the experimental results. The main work includes:Acquisition and processing of prototypical3D face data. Firstly, choosingthree-dimensional optical scanner3DSS to obtain mass point cloud data of face sample, andpreprocessing the data about denoising, merging multi-view point data and data reduction.Because this method faces complex data post-processing, it is not suitable for the rapidestablishment of large scale face database. Secondly, preprocessing the3D face scanning spotacquired by Cyberware Laser Scanner through a texture amelioration method based on texturemapping and coordinate correcting. Making correspondence between vertices of prototypical3D faces is the key basis of constructing morphable model. After introduced formeralgorithms based on optical flow or TPS, a nonuniform resampling method for Pixel-wisecorrespondence between3D faces is proposed. The new method realizes automatic facessegmentation based on ASM algorithm, and then according every patch’s curvature todetermine the face overall resampling strategy. It also shows better results in reducing thenumber of mesh vertices and mesh triangles than former algorithm.3D Morphable model is constructed, and matching it with a face image for3D facialsynthesis. Morphable model is a kind of liner combination model. Adopting principalcomponent analysis to obtain the basis of the subspace then constructing the3D morphablemodel by the liner combination of the basis, and then evaluating the model through PSNR ofthe recovery sample. The specific person’s3D face is automatically reconstructed bymorphable model matched with his2D facial image. We use a Maxinum a Posteriori estimatorto get the goal function. Aimed the shortage of stochastic gradient descent method which isover-dependent on initial values, and has slowly convergent speed, we proposed the improved algorithm of Leven-Marquardt to optimize the goal function, deducing the matchingparameters to matching the eventually3d face model with the software development kit ofOpenGL.