Research On 3D Face Reconstruction Methods Based On Single View Of Face

In recent years,with the rapid development of artificial intelligence technology,the research of 3D face reconstruction based on a single view of face has brought great convenience to people's daily life and entertainment activities.At the same time,it has always been a great challenge in the research field of computer vision to reconstruct 3D face simply,quickly and efficiently.To this end,this thesis has completed the following research work:Firstly,this thesis reconstructs a 3D face based on the face front view.After selecting the front image of the face,this thesis utilizes the active shape model algorithm to perform the face alignment to obtain the landmarks.Subsequently,the model generated by Facegen was modified as the reference 3D face model,and a mapping relationship between the landmarks and the model vertexes was established.According to the mapping relationship,the radial basis function interpolation algorithm is used for interpolation,so that the reference 3D face model can be transformed into a personalized 3D face model.In order to reconstruct a realistic 3D face,this thesis then uses Open GL texture mapping technology to realize the texture mapping of the 3D face model.Secondly,this thesis designs and develops a 3D face reconstruction system based on the front view of the face.The system integrates and perfects all the face reconstruction processes described above,and reasonably imagines the 3D face reconstruction scene based on the front view of the face,so that users can build,adjust and output the 3D face model according to the system function and interaction logic,thereby achieving the one-stop 3D face reconstruction.Finally,this thesis represents a 3D face model as a position map,and combines neural network to implement the 3D face reconstruction based on complex pose face images.The neural network first down-samples the face image for extracting the feature maps,and then uses the feature maps for up-sampling to generate the position map.In the neural network,the down-sampling module is formed by stacking residual blocks,and the up-sampling module is composed of shortcut-upsampling blocks proposed in this thesis.In addition,this thesis also proposes a dynamic weight loss function to promote further convergence of the neural network.The experimental results show that the neural network with the shortcut-upsampling blocks can quickly achieve excellent 3D face reconstruction when the neural network model size is only 121M.