Detailed 3D Face Reconstruction Based On Neural Network

With the development of computer vision geometry,more and more researchers are pursuing the depth information contained in human faces.Therefore,the focus of research has been shifted from 2D plane face images to 3D vector space models.The face model in the 3D vector space is composed of a point cloud and a triangular grid connected to each other.In a constrained environment,the face will be affected by illumination,self-occlusion,etc.,the reconstructed model is noisy,and the prediction error of some vertex coordinates of the reconstructed model is relatively large.Therefore,in the method in this paper,the Basel face model(BFM)based on the three-dimensional deformation model(3DMM)is used to fit the average face.In view of the multi-information characteristics of human faces,the parameterized 3DMM method only needs to input multiple sets of simple principal component analysis parameters to obtain a smooth 3D model.Face alignment is one of the basic algorithms for face research.The method in this paper can perform face calibration on a large pose face image in a constrained environment,and detect its facial feature points after calibration,and then calculate the required parameters of 3DMM.Feature point detection may be inaccurate in some situations,so this paper uses a neural network to perform secondary processing on the calculated parameters to output a more complete model.This paper studies the pose and expression face reconstruction in unconstrained environment.The main methods proposed are face alignment and the detailed synthesis of reconstruction model.This paper mainly deals with 3D face reconstruction by neural network integrated by multi seed network.The network integrates the steps of face reconstruction,and the data obtained will be shared in the whole network,and the efficiency will be improved when the accuracy and robustness are guaranteed.The detail synthesis can solve the information loss caused by PCA algorithm.The detail synthesis network based on condition generated adversary network(CGAN)is responsible for processing details,and can handle the feature map better.The detail synthesis network outputs the displacement map,which is based on the mapping technology,which contains most of the face details.In the experimental part,the integrated network method proposed in this paper is compared with other advanced methods.In the evaluation experiment,first preprocess the reconstructed model and compare the real model to solve some problems that may affect the numerical calculation in the evaluation experiment.Then evaluate the integrated network proposed in this paper through the given evaluation criteria.In the model comparison calculation,the accuracy of the detected feature points and the reconstruction model in the constrained environment,the reconstruction model in the constrained environment,and the model robustness evaluation error of the four experiments are respectively calculated.The data results show that the method proposed in this paper has a higher degree of fit for facial images in every scenarios.