Research And Implementation Of 3D Face Reconstruction Algorithm Based On Deep Learning

The demand for 3D face models in the fields of face recognition,AR/VR technology,game animation,medical beauty,etc.is increasing.The purpose of the 3D face reconstruction task is to recover a 3D face model similar to the real face from the 2D face image.It has realistic effects as much as possible to meet the actual needs and is widely used.3D face reconstruction has achieved a number of research results,but there are still many problems that need to be resolved,including the unconstrained face reconstruction in natural scenes.Among them,the common complex conditions such as large poses,exaggerated expressions,occlusion,and insufficient lighting will make the face have invisible areas.The purpose of this paper is to propose a new method to solve this problem and effectively improve the reconstruction effect of the unconstrained face.Considering the superior performance of convolutional neural networks for image processing in deep learning,its excellent feature extraction capabilities,and non-linear expression capabilities are helpful to achieve 3D face reconstruction of 2D face images.Therefore,this paper proposes a new deep learning model to solve the problem of 3D reconstruction of unconstrained faces in natural scenes.The main work of this paper is as follows:1.Based on the classic 3D morphable model(3DMM)and the existing public face database,this paper uses a 3D morphable model that combines the shape and expression of the face.The purpose is to obtain the model parameters composed of shape coefficients,expression coefficients and pose coefficients through regression of the convolutional neural network and then to fit the 3D morphable model,so as to reconstruct the 3D face model represented by grid corresponding to the input 2D face image.2.In order to fully capture the apparent information and geometric structure information of the face,and be able to learn the face difference information between different individuals,this paper adopts the idea of face attribute decomposition and proposes an attribute decomposition learning network.After extracting the overall information of the face,the face attributes are decomposed into three parts:shape,expression,and pose through three branch networks,and each of them independently learns deeper corresponding attribute features.Through this network regression,more specific feature information of each attribute in face semantic information is obtained,and a more realistic 3D face model is reconstructed.3.Consider the overall optimization of the network to learn more detailed facial semantic information.On the basis of introducing the corresponding loss function of each face attribute to optimize the three branch networks,sparse face landmarks are introduced as constraints to optimize the edge features of the face,perfect the reconstruction effect of the geometric structure of important parts of the face,and then improve the accuracy of the overall reconstruction results of the 3D face model.4.Based on the attribute decomposition learning network 3D face reconstruction algorithm proposed in this paper,a convenient and fast 3D face reconstruction system is designed and implemented,which can realize the main functions such as 3D face reconstruction and face alignment.