3D Face Reconstruction And Application Based On Single Image

3D face reconstruction is one of the active topics in computer graphics and vision.It has a wide range of applications in various fields including face recognition,face animation,video and game.Compared with the three-dimensional face reconstruction algorithm based on image sequence and sensor equipment,the face reconstruction based on single image does not need to rely on the face image sequence,and does not be limited to factors such as hardware equipment.Therefore,this thesis chooses the three-dimensional face reconstruction based on single image as the research topic.However,there are some problems in 3D face reconstruction based on single image:the geometric details of eyes,nose,mouth are not optimized enough;the local processing of face texture is not fine enough;the reconstructed face texture and the complexion given by the model are not consistent.To solve these problems,this thesis proposes a 3D face reconstruction method based on model segmentation and texture optimization.The main work is as follows:(1)The core tensor of complete face mesh model is constructed by the tensor decomposition method,and the core mesh model tensor of eyes,nose and mouth is constructed by the conditional random field segmentation method and tensor decomposition method.Then,we use the cascading method to detect the feature points of the input face image,and then we segment the input face image and database face images through feature points.At the same time,we calculate the structure similarity index of the input sub-block images and the sub-block database face images,and the corresponding sub-block mesh models of eyes,nose and mouth are obtained by the core tensor matching.Then,the sub-block mesh models and the complete face mesh model are fused by setting the weights.Next,Laplacian smoothing is applied to the face mesh model locally and globally,and finally the fine face mesh model is obtained.(2)Combining with the spectral clustering algorithm and hierarchical region tree algorithm,the face texture image is segmented and optimized,and the complexion of texture region is fused by the Poisson fusion algorithm.Then a 3D face model with realistic texture effect is reconstructed by the face texture mapping.During the face texture optimization,the complexion region of face image is detected by the complexion ellipse model,and then the complexion conversion is realized by using the edit propagation method based on local linear embedding algorithm,so that the complexion of the reconstructed face model is consistent with the complexion given by the human body model.(3)The proposed algorithm is compared with the reconstruction algorithm based on the deformation model and shadow change.The experimental results show that the 3D face model reconstructed by this algorithm has advantages in the local effect,fidelity and texture detail,and the running time efficiency of this algorithm is also reasonable.Finally,this thesis introduces the application of the reconstruction algorithm in the complexion conversion of human body model.The experimental results show that the complexion conversion effect of this algorithm is natural.Finally,the main research contents of this thesis are summarized,and the future work focus is pointed out.