3D Face Reconstruction From Single Facial Image Based On Convolutional Neural Network

Three-dimensional face reconstruction is widely used in medical,security,entertainment and other fields.However,there are two main kinds of problem when currently applied in real scenes: the three-dimensional simulation is unstable,resulting in a large difference in the shape of the face simulated by the same individual;The 3D simulation is generalized,which leads to the similar face shapes of different individuals.At present,the methods with good simulation effects all require expensive instruments and consume a lot of manpower time to obtain a highly accurate face reconstruction model,while the low cost experimental method is not ideal enough.In view of the shortcomings of the above research,combined with the actual application of the face picture is mostly a single face image of any gesture,put forward a method of 3D face reconstruction based on Convolutional Neural Network(CNN)from any single twodimensional face picture is proposed.The main contribution of this paper can be summarized as follows:(1)In the preprocessing of 3D face reconstruction,the feature points of the input 2D face picture need to be located to realize the mapping between the 2D picture and the 3D model.At the same time,for different face pictures of the same individual,the corresponding weights need to be determined according to the face poses in different pictures to complete the fusion of all the picture information of the individual.Aiming at the two problems,a large gesture face alignment method based on CNN is used,this method is based on the traditional convolutional neural network cascade structure,uses a novel network structure that introduces visual blocks,by introducing a visual layer in the visualization block,and integrates it into the CNN architecture to achieve end-toend training to reduce Training duration.In view of the particularity of the network model structure,it is proposed to use two loss functions in the training process of the model to enhance the fitting effect of the model.(2)A 3D face reconstruction method based on 3D Morphable Models(3DMM)and CNN which is for a single 2D face image at any angle is put forward.Combine the traditional 3DMM model theory and deep learning method to train a network model to achieve a single 2D face image output through the model corresponding to the 3D face shape.In order to enhance the robustness of the model,this paper proposes an improved multi-image deformation model matching method based on face pose estimation.Determine the weights of face images at different angles based on face poses to achieve the fusion of face feature information of the same individual at different angles,this fusion information is used as the expected output of the model,which makes the model obtained by the training reduce the difference between the face shapes reconstructed by different pictures of the same individual,and improves the stability and accuracy of model reconstruction.(3)Application of 3D Face Reconstruction Model based on CNN.Study the effect of 3D Face Reconstruction Model that applied to face recognition for the application scene of 3D face reconstruction model.Judge whether the face image belongs to the same individual based on the shape and texture information of the face image,so as to achieve the purpose of face recognition,and at the same time verify the robustness of the 3D face reconstruction system.