Methods And Implementations Of 3D Face Reconstruction Based On Deep Learning From Profile

In recent years,3d face reconstruction has become a hot topic.It has excellent performance in computer vision,artificial intelligence and other fields.However,the monitors are generally located on the top of the face,and the overhead pictures will lack a lot of information about the face,which is not conducive to reconstruction.Therefore,there are relatively few studies on 3d face point cloud reconstruction using single overhead images.In this paper,the point cloud reconstruction of 3d human face is studied by using the 2d overlooking image of human face,and a Generative Adversarial Network is proposed for the reconstruction of 3d point cloud.The main contents are as follows :(1)A Wasserstein GAN is constructed based on the convolution of tree graph,which can generate the point cloud of 3d human face based on any noise input;(2)on the basis of(1),the conditional generation model is improved to take the overhead image as the condition and the sampled noise as the input for face reconstruction,and on the basis of the Wasserstein GAN native loss function,the Chamfer Distance is also used as the evaluation index;(3)in order to compare with(2),a conditional reconstruction model based on multi-angle facial image was designed;(4)collect and process the data needed for the experiment.The three models in this paper all use Wasserstein GAN as the main body and introduce the convolutional neural network based on tree as the generator.The convolutional neural network can obtain the ancestor information to enhance the feature representation ability and speed up the computation.Use the fully connected network as the discriminator.The paper starts from the acquisition of data,and gradually completes the overall work,including data processing,training network,experimental results comparison and so on.Better reconstruction results were obtained and the network speed was improved compared with other face reconstruction work.However,the model in(3)did not make significant improvement compared with(2)on the premise of losing the running time.Therefore,the model based on profile reconstruction can achieve the optimal effect under the balance.