Research On Covered 3D Face Reconstruction Algorithm Based On CNN

With the continuous development of science and technology and the continuous improvement of people's living standards,the 3D face model has more and more demands in communication,medical care,games,and animation.Especially as VR technology matures,creating a virtual environment that makes people feel like there is no longer a dream.And this requires the support of 3D reconstruction technology.In recent years,with the accelerating computing speed of hardware devices,deep learning has attracted more and more attention,and it has become one of the hottest areas of research in the academic world.And because convolutional calculations are suitable for image processing,Convolutional Neural Networks(CNN)continues to make progress in image-related research.This paper uses convolutional neural network to extract the high-dimensional features of face images,and studies the occlusion and three-dimensional reconstruction of occlusion face images.First of all,it studies the occlusion algorithm that blocks the face image,introduces and analyzes the advantages and disadvantages of the traditional face restoration algorithm.Most of the traditional algorithms use artificially calibrated feature points to recover from the symmetry and correlation of face images.However,artificially calibrating feature points is influenced by experience and there are limitations.This paper proposes a face image de-occlusion algorithm based on an auto encoder(Auto Encoder): Through the training of a large amount of data,it searches for high-dimensional features that are significant in face images;based on these features,a decoding network composed of transposed convolutions is used to generate images.;and calculate the loss between the generated image and the occlusion area of the real face image to perform inverse adjustment;finally obtain a network that can achieve occlusion effect for the occluded face image.Secondly,the image-based 3D face reconstruction algorithm is studied,and the advantages and disadvantages of the traditional reconstruction algorithm are introduced and analyzed.Most of the traditional algorithms use artificially-calibratedfeature points to achieve continuous reconstruction using iterative calculations.Not only are people limited,but calculations are time consuming.This paper proposes an occluded 3D face reconstruction algorithm based on E2FARMod: using convolutional neural network to extract features of face images,replacing the traditional manual calibration of feature points;through continuous training to determine between the face image and the 3D face model The intrinsic connection,in combination with the idea of the 3D Morphable Model(3DMM),transforms this relationship into corresponding shape features and reconstructs it;finally,a network is obtained that can achieve better three-dimensional reconstruction of occluded face images.Finally,experiments are performed on different datasets to show the effects of AutoEncoder-based face image de-occlusion algorithm and E2FARMod-based occlusion 3D face reconstruction algorithm,and compared with other 3D face reconstruction algorithms.The experimental results show that the proposed algorithm has a good effect on the occlusion and three-dimensional reconstruction of occluded face images.