The Research On 3D Face Reconstruction Based On Facial Photos

3D face reconstruction has great application value in the fields of identification and certification,film and television advertisements,game animation,medical cosmetic and virtual reality,etc.It has gained great attention from many domestic and foreign scholars.The traditional 3D face reconstruction method mainly extracts facial feature information based on multiple face images,and then completes the modification and optimization of the general model through an approximation algorithm.In recent years,a 3D face reconstruction technique based on a single face photo has emerged.Widely used in various fields,it is a new research direction of 3D face reconstruction.In this paper,based on the method of model modification,the realistic 3D face model of single face photo reconstruction is studied.The main work includes:(1)Face feature point positioning.Based on ORL face database,the active face model method is used to train the global facial shape model and local texture model in the target image.In the search phase of the model,the eyes and mouth centers in the photo are set as the initial position of the global shape model.The local texture model is used to complete the precise positioning of the facial feature points,and the plane information(x,y axis coordinates)of the feature points is stored at the same time to prepare for the adjustment phase of the three-dimensional grid model.(2)The depth of facial feature points is estimated based on the neural network method.For the problem that the initial weights and thresholds randomly generated in each layer of the traditional BP neural network lead to defects in the network training process,the L-M method is used to optimize the weights and thresholds in the neural network so that the depth of the facial feature points can be estimated.(3)The depth value of the facial feature points is estimated based on the nearest neighbor weighting method.Since the nearest neighbor method uses a certain optimal solution in the sample set to estimate the error of the feature information in the feature point,the K-Nearest Neighbor idea in the local linear embedding algorithm is used to estimate the depth of the facial feature point.Simultaneously in the Matlab platform to complete the simulation experiment,the experimental results of the above two methods are analyzed and compared,and the high-precision method is used to extractthe depth data(z-axis coordinates)of the feature points in the three-dimensional face reconstruction.(4)Proportional adjustment of average 3D face model.Firstly,the structure of the Candide-3 model is analyzed in detail.Then the average 3D mesh model of Candide-3 can be adjusted to a specific face mesh model based on the three-dimensional coordinate data of the facial features.Finally,texture mapping is performed to generate a strong sense of reality.A smooth 3D face model.Through the in-depth research of the above aspects,the realistic 3D face reconstruction based on a single face photo is finally achieved.The experimental results verify the accuracy and feasibility of the proposed method.