3D Face Alignment And Reconstruction Based On Convolutional Neural Network

With the advent of the digital age,human facial feature information is widely used in security monitoring,medical imaging,entertainment and audio-visual industries.Compared with two-dimensional faces,three-dimensional faces contain richer information and are not affected by facial poses,which is suitable for different scenes.Face alignment and face reconstruction in the field of computer vision are two highly related problems,which are often used in the same related face task.The traditional face alignment model is not ideal when dealing with large poses face,face reconstruction model also has the problem of high model redundancy and low accuracy.With the development of deep learning,the performance of these two algorithm models has been continuously improved,and two tasks can be achieved in the same algorithm model.This paper is based on the convolutional neural network and combines the relevant knowledge of 3D face reconstruction and alignment.Two improvements to the existing model are proposed.(1)A neural network composed of up-sampling network and down-sampling network is proposed,a subsampling network composed of dense convolution module and transition layer is designed,and then the two kinds of transpose convolution are combined as the upsampling network to return the 3D face information from the extracted feature images.The network effectively extracts the features of 2D face images and reduces the model parameters.(2)A weighted loss function is proposed,in this paper,image structure similarity is used to measure the difference between the predicted 3D face information and the real 3D face information,and the weight matrix is used to focus the training on the key parts of the face.Two improvements were combined to form the algorithm in this paper.T the experimental results on the AFLW2000-3D data set show that the model has improved the performance of face alignment and 3D face reconstruction in each pose.