Research On Monocular 3d Face Reconstruction Algorithm Based On Deep Learning

In recent years,3D face reconstruction has been a hot research topic in the field of computer vision.Therefore,3D face models have also become more and more widely used in security,medical,film and television fields.Accurate 3D face reconstruction technology relies on 3D scanning equipment,or the use of binocular cameras and other parallax relations to reconstruct the 3D face structure.Both the scanning device and the binocular camera have the disadvantages of high cost and slow reconstruction speed,which makes it very difficult in practical life applications.The reconstruction cost of three faces of a single image based on visual algorithm is low and the reconstruction speed is relatively fast.Therefore,three-dimensional face reconstruction of a single image has become a research focus in the field.This paper studies and implements two single-image rough three-dimensional face reconstruction algorithms based on different application scenarios with different emphasis on the speed and quality of reconstructing the face model,and proposes a single image dense three-dimensional face reconstruction algorithm.Rough 3D face reconstruction is suitable for applications such as assisted face recognition,expression migration,and face exchange.This paper studies and implements the following two reconstruction algorithms:(1)Single image 3D face reconstruction based on a parameter model.Based on the parameter model 3DMM coefficients,a convolutional neural network is used to predict the UV position map containing position information from the input single face map,and the UV position map is used to complete the two-dimensional to three-dimensional conversion to establish a three-dimensional face model.(2)Three-dimensional face reconstruction of a single image based on stereo pixels.The down-sampling of the dense point cloud of the 3D face model into the binary stereo pixel space,with 68 key points of the face as the regular item,the stereo pixel information is directly predicted from the input face image through the convolutional neural network,thereby establishing the 3D human Face stereo pixel model.Compared with other algorithms in the field,the two algorithms show a certain accuracy on the world common face data set,and show good robustness when reconstructing large-scale face images.Dense 3D face reconstruction is suitable for applications that require high quality face models such as plastic surgery and 3D movies.Inspired by the image style conversion task,this paper proposes a dense three-dimensional face reconstruction algorithm based on a generative adversarial network,that is,through the generative adversarial network,the corresponding depth map is predicted from the input face map,and the pixel-to-pixel prediction regression is completed.At the same time,the3 DMM coefficients are returned through the convolutional neural network to establish the deep 3DMM.After that,the depth 3DMM is combined with the depth information in the depth map to obtain a dense three-dimensional face model with details such as facial wrinkles.At the same time,the algorithm will also use the training data set for data enhancement,and finally compare and analyze with the more advanced reconstruction algorithm in the field.The results show that the algorithm is slightly lower in accuracy than the algorithm with the highest reconstruction accuracy in the field,but it is relatively robust when reconstructing large-scale pose face images.After completing the algorithm research work,this paper integrates the above algorithm model into a software system on the Windows side,that is,a monocular 3D face reconstruction system.This paper gives the design framework of the system,introduces the development and operation process of dense 3D face reconstruction,3D face impact of stereo pixels and 3D face reconstruction of video streaming,and realizes the 3D face image of any clearly visible face.reconstruction.