3D Reconstruction Of Human Face Based On Morphable Model

With the rapid development of computing technology,3D face model has been widely used in all walks of life.How to reconstruct realistic 3D face model has become the most challenging problem in the field of computer vision.The biggest advantage of 3D model is that it can reflect more internal information,which is not available in 2D pictures.Therefore,more and more people are aware of its potential application value and wide application prospect.Therefore,3D reconstruction of human face has become one of the most active research fields in the world.Now many researchers have used 3D reconstruction of moiphable model to study human face,is aimed at a single image reconstruction,as a result of the position of the single photo has less information,there is no guarantee that from all angles to rebuild the comparison ideal result,the traditional method based on deformation model,usually for prior model of shape and texture parameters are modified to obtain multiple views of the same model for alignment and modeling of the model.Such methods are ultimately about learning the parameters of a morphable model.The method in this paper is to use convolutional neural network to study the spatial information of 3D model,which is essentially different from the traditional parameter learning.Moreover,it can solve the tasks of face alignment and face 3D reconstruction at the same time.The 3D reconstruction method proposed in this paper mainly includes the following aspects:(1)In the process of 3D face reconstruction,face feature points should be detected at first and facial feature points is one of the important according to the sparse face alignment,is also a prerequisite for the face must be completed before 3D reconstruction work,based on the cascade regression method to complete the feature point detection,and use the UV position maps for human face three-dimensional data input,the method in the UV space to record the three-dimensional space of full face,and offers a closely corresponding to each point in the UV space semantic.The problem of solving parameters of morphable model is transformed into the problem of learning image features.(2)The structure of the convolutional neural network with weighted loss is trained,and the higher-order characteristic information in the UV location map is learned by using the convolutional neural network.Different weight masks are set during the training,which assign different weights to each point on the location map and calculate the weighted loss.In this way,more emphasis is placed on the key recognition area.Through deconvolution,the UV location map with the same size as the input is restored,and then the real 3D face model is obtained through meshing and texture mapping.Face alignment and 3D face reconstruction are solved in an end-to-end way.Compared with other reconstruction methods through experiments,qualitative and quantitative analysis proves that the face reconstruction method presented in this paper has better effect and stronger robustness in face alignment and 3D model of face,and its computing speed is also improved compared with the previous methods.(3)In order to verify the proposed method in practice,a 3D face reconstruction visualization system is designed and made by combining the method in this paper with relevant development software.Users can quickly realize the key steps of face 3D reconstruction through simple operations,and hide the intermediate links that are not important to the experience-seekers,so as to show the results of the theoretical method succinctly and intuitively.