3D Face Reconstruction And Recognation Using Partial Differential Equation

Face recognition is an important research in security of biologicalinformation, but also in computer vision systems. due to the complexity of humanfaces and restrictions of the current technologies, the recognition of3D face hasmore challenge than2D face, at the AFMe time, there will be more features incomplex environment.The Partial Differential Equation (PDE) is an advanced geometric surfacerepresentation methods，especially it has seamless merging of surface patches andsmall number of parameters, based on these two advanced points, we first workon the method of3D face reconstruction from point cloud, and then investigatethe potential of pose, expression and occlusion. The main research contents andcontributions of this dissertation are as follows:We proposed a hybrid method to build the canonical depth map of3D face,at the AFMe time we have designed a handle method of3D point cloud. As theexisting methods to build the self-based depth map representations of3D facescans suffer either the susceptibility of anchor point detection or the robustness ofglobal feature extraction, we propose a hybrid method to construct the canonicaldepth map representations of3D face scans. Based on the canonicalrepresentation of3D face scans, we extended the depth map representation bywhich the3D face scans are normalized to be suitable for PDE geometric surfacemodeling.As the existed methods of PDE-based geometric modeling lack themechanism of geometric model selection, these methods tend to suffer theproblem of instability, redundancy and inefficiency. To address these deficiencies,and improve the accuracy of3D face reconstruction, we proposed two kinds ofmethods to sovle this problem. One is based on the Fourier series, we call thisFourier series fitting; and the other one is called the least-square approximation. Moreover, a method of PDE curves reconstruction based on least error isproposed. Furmore, we propose a adaptive reconstruction of3D face. Theexperiment results show that these methods can reduce the redundancy and moreimportant, they improve the accuracy of reconstruction and the model stability.The deformation and merit are the most important element for3D facerecognition. In the circumstance of using another way to extend3D face, a FacialDeformation Model based on Fourier Series (FDMFS) method for modeling isproposed. Based on these two situations, we proposed a Expression DeformationModel based on Fourier Series (EDMFS) and a Occlusion Deformation Modelbased on Fourier Series (ODMFS) and they are uniformly called FacialDeformation Model based on Fourier Series (FDMFS).Simulation results showthe FDMFS model has good description ability in many applications.