Research On3D Face Visualization Based On Single Image

Recently,3D face visualization has been widely used in the fields of3D games,3Danimation and security. It is also the research focus in many fields, especially in thefield of computer graphics and pattern recognition. Among kinds of3D facevisualization methods, the single image based method is particularly important becauseof its simplicity and efficiency.This thesis aims to develop a3D face model corresponding to the face in the2Dphotos. The traditional statistical learning method first transforms the PCA statisticalmodel into a3D shape model for each face image. Then, it maps the image texture tothe model by using the orthogonal projection. Based on these researches, we havecompleted some works as follows:1. Unify the meshes of the3D face sample library.To obtain the statistical learning model, we must first build a3D face library,where each face sample has the same vertices and mesh topology. Many researchershave paid their attention to the vertex correspondence. However, the meshcorrespondence is rarely considered. In this thesis, we introduce a simple method tounify the meshes. The proposed method adjusts the triangular meshes of the samplesaccording to the transformation and connection relationship of the vertices.2. Solve the energy function in building shape model.To build a3D face model, we need to solve an energy function which is formulatedas a least squares problem. To avoid the over-fitting problem, we add a regularizationterm into to the energy function. In most cases, the optimized3D face shape model canmatch the face in the2D photo. However, the proposed method still has some defectscompared with the method which uses some prior knowledge as the constraint.3. Propose two methods of3D feature point depth estimation to improve theaccuracy of shape model.This thesis proposes two methods for feature point depth estimation, i.e.,weighted neighbors method and sparse representation method. Here, we use theestimated3D feature points to construct the3D face model. Experiment results show that the estimated feature point depth can improve the modeling accuracy and generatemore real and authentic3D face models.4. Improve the texture mapping based on segmentation template.In the texture mapping process, the constrained point based method usuallyproduces residual background. To overcome this problem, we proposed a segmentationbased method. A2D segmentation mask is first generated from the2D face image. Then,it is refined by erosion operation. The mask is used as a constraint to guide the texturemapping. Experiment results show that the proposed method can avoid the appearanceof the residual background effectively.