Research On 3D Face Reconstruction Based On Single Image

With the advancement of technology,3D face reconstruction research,as an important research content of computer graphics,computer vision and virtual reality technology,has been widely used in games,film,television and medical fields.Traditional 3D face reconstruction methods mostly rely on 3D scanners,multiple photos and binocular vision equipment to obtain facial feature information,and then use approximation optimization algorithms to complete face model reconstruction,modification and optimization,which is expensive and complicated.The reconstruction of 3D faces based on a single 2D image has been a new research direction in recent years.This paper focuses on the problems of inaccurate extraction of 2D feature points,slow extraction of 3D feature points,and low realism of 3D model deformation in the reconstruction of 3D faces based on 2D images.The main research contents include as follows:(1)To solve the problems of inaccurate calibration of 2d feature points and slow calibration of 3d feature points,this paper improves the 2d feature point extraction method based on deep learning and the 3d feature point extraction method based on key points from rough to fine.First of all,single face image is used as a driver,and the Adaboost algorithm is used to detect face parts,which reduces the training workload.Secondly,the convolutional neural network is used to train the model,so as to extract the two-dimensional face feature points.Finally,the 3d face feature points are roughly located by combining depth maps and supervised learning method,and the feature points of the face are screened by combining multiple local descriptors,so as to improve the extraction accuracy of feature points and complete the selection of 3d face feature points.Experimental results show that the method extracts two-dimensional feature points with high accuracy and improves the extraction efficiency of three-dimensional feature points.(2)To solve the problems of slow construction of the current general models and low sense of reality of 3d face deformation,an improved 3d face deformation algorithm based on thin plate spline is proposed in this paper.First,based on the average value of each sample feature point in the BJUT-3D face database of Beijing University of Technology,the general face model data is obtained,and the face samples in the database that are very close to the data are selected as the general face model.Secondly,using the two-dimensional feature points extracted from the input picture and the three-dimensional face feature points extracted from the general model to establish the correspondence between the thin plate splines and construct the personalized face model.Finally,the texture mapping of the personalized face model is carried out based on the feature point constraints.The experimental results show that the method is fast to construct the universal face model,accurate to deform the 3d model,and realistic to reconstruct the 3d face.(3)The 3d face reconstruction system based on single image is designed and implemented.This system is based on C ++ and uses QT to build a system interface.The system uses the C / S(client-server)architecture to update the input picture.The server matches the model according to the input picture to extract the feature points and deform the model,and displays the 3d face through OpenGL to complete the reconstruction of the 3d face.The experimental results show that the system can reconstruct 3d face automatically,quickly and realistically.