Research On Refined 3D Face Reconstruction Algorithm Based On Single View

As a symbolic feature of human beings,human face contains extremely abundant personal information.It has a wide range of applications in medical beauty,film entertainment and information security.With the continuous development of artificial intelligence and deep learning in recent years,three-dimensional face reconstruction technology has become an important research topic in the field of computer vision and image processing.The 3D structure of human face is not limited to a single face pose and lighting environment.It can better describe the feature information and is more operability.Although 3D face reconstruction technology has been greatly developed in recent years,there are still many limitations that constrain the development of this field.On the one hand,although using complex hardware devices can complete highprecision and high-reality 3D face reconstruction,the usage scenarios of hardware devices are limited,the cost is high,and the processing is complicated and inflexible,which makes it impossible to be widely used in practical applications.On the other hand,image-based 3D face reconstruction technology mostly requires high quality and quantity of input images,and the effect of 3D simulation is unstable.In view of the shortcomings of the existing methods,this paper will study the refined 3D face reconstruction from three steps: single view image conversion,geometric space reconstruction and facial detail reconstruction.The main research work is as follows:(1)A Multi-Loss Image Conversion Method based on GAN(MLICM-GAN)is proposed.This method utilizes the image generation capability of GAN to convert a 2D face photo into a face depth map and correspondence map.In order to capture the face information better,an improved U-Net framework and a joint loss function are used to train the generator network,so that the converted image has better depth information and edge information.The experimental results show that the method can effectively transform the input 2D face image into depth map and correspondence map.(2)In order to transform face images from 2D space to 3D space,an improved 3D face reconstruction algorithm based on surface-based embedded deformation is adopted(MDF-IR).The method first converts the face depth map into a triangular mesh,and then adopts an improved non-rigid registration algorithm for the template face based on the mesh.In the process of iterative deformation,the mixed data fitting term and the improved regularization term are combined to form an energy function,thereby obtaining a 3D face with smoothing facial structure including facial expressions and facial features such as pleats.The experimental results show that the method basically restores facial features and facial expressions,and the texture mapping is complete.(3)In order to obtain a more refined 3D face model,a refined 3d face reconstruction algorithm based on low-pass filtering(MNVAC-DF)was designed on the basis of research work(2).The method is based on the low-pass filtering of the image,and the vertices of the model are displaced along the normal vector,and then the high-frequency components carrying the facial detail texture are superimposed on the smooth human face,and the facial details such as wrinkles are restored on a more detailed level.The experimental results show that the algorithm can effectively restore high-frequency facial texture.Compared with similar methods,the reconstructed 3D face model of the algorithm is closer to the input face,which not only restores the angle of the image,but also restores the texture details of the face more finely.